Update version number in README.md to 2.1.6

Pre-load NVIDIA shared libraries on Linux

.gitignore

@@ -27,3 +27,5 @@ faceswap/
 switch_states.json
 /models
 install.bat
+/.claude
+*.bat

README.md

@@ -1,4 +1,4 @@
-<h1 align="center">Deep-Live-Cam 2.0.5c</h1>
+<h1 align="center">Deep-Live-Cam 2.1.6</h1>
 
 <p align="center">
   Real-time face swap and video deepfake with a single click and only a single image.
@@ -30,11 +30,11 @@ By using this software, you agree to these terms and commit to using it in a man
 
 Users are expected to use this software responsibly and legally. If using a real person's face, obtain their consent and clearly label any output as a deepfake when sharing online. We are not responsible for end-user actions.
 
-## Exclusive v2.6d Quick Start - Pre-built (Windows/Mac Silicon)
+## Exclusive v2.7 beta Quick Start - Pre-built (Windows/Mac Silicon/CPU)
 
   <a href="https://deeplivecam.net/index.php/quickstart"> <img src="media/Download.png" width="285" height="77" />
 
-##### This is the fastest build you can get if you have a discrete NVIDIA or AMD GPU or Mac Silicon, And you'll receive special priority support.
+##### This is the fastest build you can get if you have a discrete NVIDIA or AMD GPU, CPU or Mac Silicon, And you'll receive special priority support. 2.7 beta is the best you can have with 30+ extra features than the open source version.
  
 ###### These Pre-builts are perfect for non-technical users or those who don't have time to, or can't manually install all the requirements. Just a heads-up: this is an open-source project, so you can also install it manually. 
 
@@ -142,7 +142,7 @@ pip install -r requirements.txt
 ```
 For Linux:
 ```bash
-# Ensure you use the installed Python 3.10
+# Ensure you use the installed Python 3.11
 python3 -m venv venv
 source venv/bin/activate
 pip install -r requirements.txt
@@ -157,7 +157,7 @@ Apple Silicon (M1/M2/M3) requires specific setup:
 brew install python@3.11
 
 # Install tkinter package (required for the GUI)
-brew install python-tk@3.10
+brew install python-tk@3.11
 
 # Create and activate virtual environment with Python 3.11
 python3.11 -m venv venv
@@ -214,7 +214,7 @@ python run.py --execution-provider cuda
 
 Apple Silicon (M1/M2/M3) specific installation:
 
-1. Make sure you've completed the macOS setup above using Python 3.10.
+1. Make sure you've completed the macOS setup above using Python 3.11.
 2. Install dependencies:
 
 ```bash
@@ -222,25 +222,25 @@ pip uninstall onnxruntime onnxruntime-silicon
 pip install onnxruntime-silicon==1.13.1
 ```
 
-3. Usage (important: specify Python 3.10):
+3. Usage:
 
 ```bash
-python3.10 run.py --execution-provider coreml
+python3.11 run.py --execution-provider coreml
 ```
 
 **Important Notes for macOS:**
-- You **must** use Python 3.10, not newer versions like 3.11 or 3.13
-- Always run with `python3.10` command not just `python` if you have multiple Python versions installed
-- If you get error about `_tkinter` missing, reinstall the tkinter package: `brew reinstall python-tk@3.10`
+- You **must** use Python 3.11, not newer versions like 3.13
+- Always run with `python3.11` command not just `python` if you have multiple Python versions installed
+- If you get error about `_tkinter` missing, reinstall the tkinter package: `brew reinstall python-tk@3.11`
 - If you get model loading errors, check that your models are in the correct folder
 - If you encounter conflicts with other Python versions, consider uninstalling them:
   ```bash
   # List all installed Python versions
   brew list | grep python
-  
+
   # Uninstall conflicting versions if needed
-  brew uninstall --ignore-dependencies python@3.11 python@3.13
-  
+  brew uninstall --ignore-dependencies python@3.13
+
   # Keep only Python 3.11
   brew cleanup
   ```
@@ -309,6 +309,9 @@ python run.py --execution-provider openvino
 -   Use a screen capture tool like OBS to stream.
 -   To change the face, select a new source image.
 
+## Download all models in this huggingface link
+- [**Download models here**](https://huggingface.co/hacksider/deep-live-cam/tree/main)
+
 ## Command Line Arguments (Unmaintained)
 
 ```
@@ -362,6 +365,7 @@ Looking for a CLI mode? Using the -s/--source argument will make the run program
 -   [kier007](https://github.com/kier007): for improving the user experience
 -   [qitianai](https://github.com/qitianai): for multi-lingual support
 -   [laurigates](https://github.com/laurigates): Decoupling stuffs to make everything faster!
+-   [maxwbuckley](https://github.com/maxwbuckley): For making the effort to optimize this for mac!
 -   and [all developers](https://github.com/hacksider/Deep-Live-Cam/graphs/contributors) behind libraries used in this project.
 -   Footnote: Please be informed that the base author of the code is [s0md3v](https://github.com/s0md3v/roop)
 -   All the wonderful users who helped make this project go viral by starring the repo ❤️

benchmark_pipeline.py

@@ -0,0 +1,178 @@
+"""Standalone pipeline benchmark — no UI required.
+
+Captures 200 frames from the webcam and runs the full face swap pipeline,
+printing per-stage timing and effective FPS.
+"""
+import os, sys, time, cv2, numpy as np, queue, threading
+
+# PATH fix for cuDNN (Windows only)
+if sys.platform == "win32":
+    _sp = os.path.join(sys.prefix, "Lib", "site-packages")
+    _torch_lib = os.path.join(_sp, "torch", "lib")
+    if os.path.isdir(_torch_lib):
+        os.environ["PATH"] = _torch_lib + os.pathsep + os.environ["PATH"]
+
+import insightface
+from insightface.app import FaceAnalysis
+from insightface.utils import face_align
+from modules.processors.frame.face_swapper import _fast_paste_back
+from modules import platform_info
+
+platform_info.print_banner()
+
+# Pick providers based on what's actually available on this machine.
+if platform_info.HAS_CUDA_PROVIDER:
+    _providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+elif platform_info.HAS_COREML_PROVIDER:
+    _providers = ["CoreMLExecutionProvider", "CPUExecutionProvider"]
+else:
+    _providers = ["CPUExecutionProvider"]
+
+# --- Init models (same as the app) ---
+print(f"Loading models with providers={_providers}...")
+fa = FaceAnalysis(
+    name="buffalo_l",
+    providers=_providers,
+    allowed_modules=["detection", "recognition", "landmark_2d_106"],
+)
+fa.prepare(ctx_id=0, det_size=(640, 640))
+swap_model = insightface.model_zoo.get_model(
+    "models/inswapper_128.onnx",
+    providers=_providers,
+)
+face_size = swap_model.input_size[0]
+aimg_dummy = np.empty((face_size, face_size, 3), dtype=np.uint8)
+
+# --- Camera setup ---
+# Windows: DirectShow explicit for MJPEG 1080p60 support.
+# macOS/Linux: default backend (AVFoundation / V4L2).
+print("Opening camera at 1080p60 MJPEG...")
+if sys.platform == "win32":
+    cap = cv2.VideoCapture(0, cv2.CAP_DSHOW)
+else:
+    cap = cv2.VideoCapture(0)
+cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc(*"MJPG"))
+cap.set(cv2.CAP_PROP_FRAME_WIDTH, 1920)
+cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 1080)
+cap.set(cv2.CAP_PROP_FPS, 60)
+time.sleep(0.5)
+
+# Warmup + get source face
+for _ in range(15):
+    cap.read()
+ret, src_frame = cap.read()
+faces = fa.get(src_frame)
+if not faces:
+    print("ERROR: No face detected in warmup frame")
+    cap.release()
+    sys.exit(1)
+source_face = faces[0]
+print(f"Source face acquired. Frame: {src_frame.shape}")
+
+# --- Capture thread (same as app) ---
+capture_queue = queue.Queue(maxsize=2)
+stop_event = threading.Event()
+
+def capture_thread():
+    while not stop_event.is_set():
+        ret, frame = cap.read()
+        if not ret:
+            break
+        try:
+            capture_queue.put_nowait(frame)
+        except queue.Full:
+            try: capture_queue.get_nowait()
+            except queue.Empty: pass
+            try: capture_queue.put_nowait(frame)
+            except queue.Full: pass
+
+cap_t = threading.Thread(target=capture_thread, daemon=True)
+cap_t.start()
+
+# --- Warmup processing ---
+print("Warming up pipeline...")
+for _ in range(20):
+    try:
+        frame = capture_queue.get(timeout=0.1)
+    except queue.Empty:
+        continue
+    f = frame.copy()
+    det_faces = fa.get(f)
+    if det_faces:
+        tgt = min(det_faces, key=lambda x: x.bbox[0])
+        bgr_fake, M = swap_model.get(f, tgt, source_face, paste_back=False)
+        _fast_paste_back(f, bgr_fake, aimg_dummy, M)
+
+# --- Benchmark ---
+N = 200
+print(f"\nBenchmarking {N} frames...")
+
+t_queue, t_det, t_onnx, t_paste, t_copy, t_cvt, t_total = [], [], [], [], [], [], []
+det_count = 0
+cached_face = None
+
+for i in range(N):
+    tt = time.perf_counter()
+
+    t0 = time.perf_counter()
+    try:
+        frame = capture_queue.get(timeout=0.1)
+    except queue.Empty:
+        continue
+    t_queue.append((time.perf_counter() - t0) * 1000)
+
+    # Detection every 3rd frame — det-only (no landmark/recognition)
+    det_count += 1
+    if det_count % 3 == 0:
+        t0 = time.perf_counter()
+        from insightface.app.common import Face as _Face
+        bboxes, kpss = fa.det_model.detect(frame, max_num=0, metric='default')
+        if bboxes.shape[0] > 0:
+            idx = int(bboxes[:, 0].argmin())
+            cached_face = _Face(bbox=bboxes[idx, :4], kps=kpss[idx], det_score=bboxes[idx, 4])
+        t_det.append((time.perf_counter() - t0) * 1000)
+
+    if cached_face is not None:
+        # No frame.copy() — _fast_paste_back writes in-place, we own the frame
+        t0 = time.perf_counter()
+        bgr_fake, M = swap_model.get(frame, cached_face, source_face, paste_back=False)
+        t_onnx.append((time.perf_counter() - t0) * 1000)
+
+        t0 = time.perf_counter()
+        result = _fast_paste_back(frame, bgr_fake, aimg_dummy, M)
+        t_paste.append((time.perf_counter() - t0) * 1000)
+
+        # Display prep — resize then flip (no cvtColor needed)
+        t0 = time.perf_counter()
+        small = cv2.resize(result, (640, 360))
+        _ = small[:, :, ::-1]  # BGR→RGB zero-copy
+        t_cvt.append((time.perf_counter() - t0) * 1000)
+
+    t_total.append((time.perf_counter() - tt) * 1000)
+
+stop_event.set()
+cap.release()
+
+# --- Results ---
+def s(name, arr):
+    if not arr:
+        return
+    avg = sum(arr) / len(arr)
+    print(f"  {name:25s}: avg={avg:6.1f}ms  min={min(arr):5.1f}ms  max={max(arr):6.1f}ms  n={len(arr)}")
+
+print(f"\n{'='*55}")
+print(f"  1080p Pipeline Benchmark ({len(t_total)} frames)")
+print(f"{'='*55}")
+s("queue.get (wait for cam)", t_queue)
+s("detection (fa.get)", t_det)
+s("frame.copy()", t_copy)
+s("ONNX swap", t_onnx)
+s("_fast_paste_back", t_paste)
+s("cvtColor BGR->RGB", t_cvt)
+s("TOTAL per frame", t_total)
+
+avg_total = sum(t_total) / len(t_total)
+avg_queue = sum(t_queue) / len(t_queue)
+print(f"\n  Effective FPS:          {1000/avg_total:.1f}")
+print(f"  FPS (excl. cam wait):   {1000/(avg_total - avg_queue):.1f}")
+print(f"{'='*55}")

modules/core.py

@@ -2,7 +2,7 @@ import os
 import sys
 # single thread doubles cuda performance - needs to be set before torch import
 if any(arg.startswith('--execution-provider') for arg in sys.argv):
-    os.environ['OMP_NUM_THREADS'] = '1'
+    os.environ['OMP_NUM_THREADS'] = '6'
 # reduce tensorflow log level
 os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
 import warnings
@@ -17,12 +17,16 @@ try:
 except ImportError:
     HAS_TORCH = False
 import onnxruntime
-import tensorflow
+try:
+    import tensorflow
+    HAS_TENSORFLOW = True
+except ImportError:
+    HAS_TENSORFLOW = False
 
 import modules.globals
 import modules.metadata
 import modules.ui as ui
-from modules.processors.frame.core import get_frame_processors_modules
+from modules.processors.frame.core import get_frame_processors_modules, process_video_in_memory
 from modules.utilities import has_image_extension, is_image, is_video, detect_fps, create_video, extract_frames, get_temp_frame_paths, restore_audio, create_temp, move_temp, clean_temp, normalize_output_path
 
 if HAS_TORCH and 'ROCMExecutionProvider' in modules.globals.execution_providers:
@@ -53,7 +57,7 @@ def parse_args() -> None:
     program.add_argument('--live-mirror', help='The live camera display as you see it in the front-facing camera frame', dest='live_mirror', action='store_true', default=False)
     program.add_argument('--live-resizable', help='The live camera frame is resizable', dest='live_resizable', action='store_true', default=False)
     program.add_argument('--max-memory', help='maximum amount of RAM in GB', dest='max_memory', type=int, default=suggest_max_memory())
-    program.add_argument('--execution-provider', help='execution provider', dest='execution_provider', default=['cpu'], choices=suggest_execution_providers(), nargs='+')
+    program.add_argument('--execution-provider', help='execution provider', dest='execution_provider', default=[suggest_default_execution_provider()], choices=suggest_execution_providers(), nargs='+')
     program.add_argument('--execution-threads', help='number of execution threads', dest='execution_threads', type=int, default=suggest_execution_threads())
     program.add_argument('-v', '--version', action='version', version=f'{modules.metadata.name} {modules.metadata.version}')
 
@@ -127,6 +131,15 @@ def suggest_max_memory() -> int:
     return 16
 
 
+def suggest_default_execution_provider() -> str:
+    """Pick the best available provider: cuda > rocm > coreml > dml > cpu."""
+    available = encode_execution_providers(onnxruntime.get_available_providers())
+    for pref in ('cuda', 'rocm', 'coreml', 'dml'):
+        if pref in available:
+            return pref
+    return 'cpu'
+
+
 def suggest_execution_providers() -> List[str]:
     return encode_execution_providers(onnxruntime.get_available_providers())
 
@@ -143,8 +156,7 @@ def suggest_execution_threads() -> int:
     if 'ROCMExecutionProvider' in modules.globals.execution_providers:
         return 1
     if 'CUDAExecutionProvider' in modules.globals.execution_providers:
-        # For CUDA, use more threads for parallel frame processing
-        return min(cpu_count, 16)
+        return 2
     
     # For CPU execution, use most cores but leave some for system
     return max(4, min(cpu_count - 2, 16))
@@ -152,9 +164,10 @@ def suggest_execution_threads() -> int:
 
 def limit_resources() -> None:
     # prevent tensorflow memory leak
-    gpus = tensorflow.config.experimental.list_physical_devices('GPU')
-    for gpu in gpus:
-        tensorflow.config.experimental.set_memory_growth(gpu, True)
+    if HAS_TENSORFLOW:
+        gpus = tensorflow.config.experimental.list_physical_devices('GPU')
+        for gpu in gpus:
+            tensorflow.config.experimental.set_memory_growth(gpu, True)
     # limit memory usage
     if modules.globals.max_memory:
         memory = modules.globals.max_memory * 1024 ** 3
@@ -223,40 +236,70 @@ def start() -> None:
     if modules.globals.nsfw_filter and ui.check_and_ignore_nsfw(modules.globals.target_path, destroy):
         return
 
-    extraction_start = time.time()
-    if not modules.globals.map_faces:
-        update_status('Creating temp resources...')
-        create_temp(modules.globals.target_path)
-        update_status('Extracting frames...')
-        extract_frames(modules.globals.target_path)
-    extraction_time = time.time() - extraction_start
-    update_status(f'Frame extraction completed in {extraction_time:.2f}s')
-
-    temp_frame_paths = get_temp_frame_paths(modules.globals.target_path)
-    total_frames = len(temp_frame_paths)
-    update_status(f'Processing {total_frames} frames with {modules.globals.execution_threads} threads...')
-    
-    processing_start = time.time()
-    for frame_processor in get_frame_processors_modules(modules.globals.frame_processors):
-        update_status('Progressing...', frame_processor.NAME)
-        frame_processor.process_video(modules.globals.source_path, temp_frame_paths)
-        release_resources()
-    processing_time = time.time() - processing_start
-    fps_processing = total_frames / processing_time if processing_time > 0 else 0
-    update_status(f'Frame processing completed in {processing_time:.2f}s ({fps_processing:.2f} fps)')
-    
-    # handles fps
-    encoding_start = time.time()
+    # Detect FPS early (needed by both pipelines)
     if modules.globals.keep_fps:
         update_status('Detecting fps...')
         fps = detect_fps(modules.globals.target_path)
-        update_status(f'Creating video with {fps} fps...')
-        create_video(modules.globals.target_path, fps)
     else:
-        update_status('Creating video with 30.0 fps...')
-        create_video(modules.globals.target_path)
-    encoding_time = time.time() - encoding_start
-    update_status(f'Video encoding completed in {encoding_time:.2f}s')
+        fps = 30.0
+
+    video_created = False
+
+    # --- In-memory pipeline (non-map_faces only) ---
+    # Reads frames from FFmpeg pipe, processes in memory, encodes directly.
+    # Eliminates all per-frame PNG disk I/O for a major speed-up.
+    if not modules.globals.map_faces:
+        update_status(f'Processing video in-memory at {fps} fps...')
+        create_temp(modules.globals.target_path)
+
+        processing_start = time.time()
+        video_created = process_video_in_memory(
+            modules.globals.source_path,
+            modules.globals.target_path,
+            fps,
+        )
+        processing_time = time.time() - processing_start
+        release_resources()
+
+        if video_created:
+            update_status(f'In-memory processing + encoding completed in {processing_time:.2f}s')
+
+    # --- Disk-based fallback (required for map_faces, or if pipe failed) ---
+    if not video_created:
+        if not modules.globals.map_faces:
+            update_status('Falling back to disk-based processing...')
+
+        extraction_start = time.time()
+        if not modules.globals.map_faces:
+            create_temp(modules.globals.target_path)
+            update_status('Extracting frames...')
+            extract_frames(modules.globals.target_path)
+        extraction_time = time.time() - extraction_start
+
+        temp_frame_paths = get_temp_frame_paths(modules.globals.target_path)
+        total_frames = len(temp_frame_paths)
+        update_status(f'Processing {total_frames} frames with {modules.globals.execution_threads} threads...')
+
+        processing_start = time.time()
+        for frame_processor in get_frame_processors_modules(modules.globals.frame_processors):
+            update_status('Progressing...', frame_processor.NAME)
+            frame_processor.process_video(modules.globals.source_path, temp_frame_paths)
+            release_resources()
+        processing_time = time.time() - processing_start
+        fps_processing = total_frames / processing_time if processing_time > 0 else 0
+        update_status(f'Frame processing completed in {processing_time:.2f}s ({fps_processing:.2f} fps)')
+
+        encoding_start = time.time()
+        update_status(f'Creating video with {fps} fps...')
+        video_created = create_video(modules.globals.target_path, fps)
+        encoding_time = time.time() - encoding_start
+        if video_created:
+            update_status(f'Video encoding completed in {encoding_time:.2f}s')
+
+    if not video_created:
+        update_status('Video encoding failed. No temporary output video was created.')
+        clean_temp(modules.globals.target_path)
+        return
     
     # handle audio
     if modules.globals.keep_audio:
@@ -272,8 +315,8 @@ def start() -> None:
     clean_temp(modules.globals.target_path)
     
     total_time = time.time() - start_time
-    if is_video(modules.globals.target_path):
-        update_status(f'Processing to video succeed! Total time: {total_time:.2f}s')
+    if is_video(modules.globals.target_path) and modules.globals.output_path and os.path.isfile(modules.globals.output_path):
+        update_status(f'Video processing succeeded! Total time: {total_time:.2f}s')
     else:
         update_status('Processing to video failed!')
 
@@ -291,9 +334,12 @@ def run() -> None:
     for frame_processor in get_frame_processors_modules(modules.globals.frame_processors):
         if not frame_processor.pre_check():
             return
+    # Pre-load face analyser in main thread before GUI starts
+    #from modules.face_analyser import get_face_analyser
+    #get_face_analyser()
     limit_resources()
     if modules.globals.headless:
         start()
     else:
         window = ui.init(start, destroy, modules.globals.lang)
-        window.mainloop()
+        window.mainloop()

modules/face_analyser.py

@@ -16,6 +16,8 @@ from pathlib import Path
 FACE_ANALYSER = None
 FACE_ANALYSER_LOCK = threading.Lock()
 
+DET_SIZE = (640, 640)
+
 
 def get_face_analyser() -> Any:
     """Get face analyser with thread-safe initialization."""
@@ -25,29 +27,196 @@ def get_face_analyser() -> Any:
         with FACE_ANALYSER_LOCK:
             # Double-check after acquiring lock
             if FACE_ANALYSER is None:
+                from modules.processors.frame._onnx_enhancer import (
+                    build_provider_config,
+                )
+                providers = build_provider_config()
                 FACE_ANALYSER = insightface.app.FaceAnalysis(
                     name='buffalo_l',
-                    providers=modules.globals.execution_providers,
-                    allowed_modules=['detection', 'recognition']
+                    providers=providers,
+                    allowed_modules=['detection', 'recognition', 'landmark_2d_106']
                 )
-                FACE_ANALYSER.prepare(ctx_id=0, det_size=(320, 320))
+                FACE_ANALYSER.prepare(ctx_id=0, det_size=DET_SIZE)
+                _optimize_det_model(FACE_ANALYSER, providers)
     return FACE_ANALYSER
 
 
-def get_one_face(frame: Frame) -> Any:
-    face = get_face_analyser().get(frame)
+def _optimize_det_model(fa: Any, providers) -> None:
+    """Replace the detection model's ONNX session with a CoreML-optimized one.
+
+    Folds dynamic Shape→Gather chains into constants (the input size is
+    fixed at det_size), eliminating CPU↔ANE partition boundaries in the
+    RetinaFace FPN upsampling path.  21ms → 4ms on M3 Max.
+    """
+    from modules.onnx_optimize import optimize_for_coreml, IS_APPLE_SILICON
+    if not IS_APPLE_SILICON:
+        return
+
+    det_model = fa.det_model
+    model_path = getattr(det_model, 'model_file', None)
+    if model_path is None or not os.path.exists(model_path):
+        return
+
+    input_shape = (1, 3, DET_SIZE[1], DET_SIZE[0])
+    optimized_path = optimize_for_coreml(model_path, input_shape=input_shape)
+    if optimized_path == model_path:
+        return
+
+    import onnxruntime
+    session_options = onnxruntime.SessionOptions()
+    session_options.graph_optimization_level = (
+        onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+    )
+
+    # Route detection to GPU shader cores (CPUAndGPU) instead of ANE.
+    # This lets detection run concurrently with the swap model on the
+    # ANE, overlapping the two inference calls.  Detection is fast
+    # enough on GPU (~4ms) and this frees ANE for the heavier swap.
+    det_providers = []
+    for p in providers:
+        name = p[0] if isinstance(p, tuple) else p
+        if name == "CoreMLExecutionProvider":
+            det_providers.append((
+                "CoreMLExecutionProvider",
+                {"ModelFormat": "MLProgram", "MLComputeUnits": "CPUAndGPU"},
+            ))
+        else:
+            det_providers.append(p)
+
+    det_model.session = onnxruntime.InferenceSession(
+        optimized_path, sess_options=session_options, providers=det_providers,
+    )
+
+
+def _needs_landmark() -> bool:
+    """Check whether any active feature requires 106-point landmarks.
+
+    Landmarks are needed by face enhancers and mouth masking, but not
+    by the face swapper alone.
+    """
+    if getattr(modules.globals, "mouth_mask", False):
+        return True
+    processors = getattr(modules.globals, "frame_processors", [])
+    return any(p in processors for p in
+               ("face_enhancer", "face_enhancer_gpen256", "face_enhancer_gpen512"))
+
+
+def _is_dml() -> bool:
+    return any("DmlExecutionProvider" in p for p in modules.globals.execution_providers)
+
+
+def _analyse_faces(frame: Frame) -> list:
+    """Run face detection, then recognition (and optionally landmark).
+
+    Replaces InsightFace's ``FaceAnalysis.get()`` to skip the
+    landmark_2d_106 model when only face_swapper is active (saves ~1ms
+    per face and avoids an unnecessary ONNX session call).
+    """
+    fa = get_face_analyser()
+
+    bboxes, kpss = fa.det_model.detect(frame, max_num=0, metric="default")
+    if bboxes.shape[0] == 0:
+        return []
+
+    need_landmark = _needs_landmark()
+    rec_model = fa.models.get("recognition")
+    lmk_model = fa.models.get("landmark_2d_106") if need_landmark else None
+
+    from insightface.app.common import Face
+
+    faces = []
+    for i in range(bboxes.shape[0]):
+        face = Face(bbox=bboxes[i, 0:4],
+                    kps=kpss[i] if kpss is not None else None,
+                    det_score=bboxes[i, 4])
+        if rec_model is not None:
+            rec_model.get(frame, face)
+        if lmk_model is not None:
+            lmk_model.get(frame, face)
+        faces.append(face)
+
+    return faces
+
+
+def get_one_face(frame: Frame, faces: Any = None) -> Any:
+    if faces is None:
+        if _is_dml():
+            with modules.globals.dml_lock:
+                faces = _analyse_faces(frame)
+        else:
+            faces = _analyse_faces(frame)
     try:
-        return min(face, key=lambda x: x.bbox[0])
+        return min(faces, key=lambda x: x.bbox[0])
     except ValueError:
         return None
 
 
 def get_many_faces(frame: Frame) -> Any:
     try:
-        return get_face_analyser().get(frame)
+        if _is_dml():
+            with modules.globals.dml_lock:
+                return _analyse_faces(frame)
+        else:
+            return _analyse_faces(frame)
     except IndexError:
         return None
 
+def detect_one_face_fast(frame: Frame) -> Any:
+    """Detection-only — skips landmark and recognition models.
+
+    Returns a Face with bbox, kps, det_score (enough for face swap).
+    ~10ms vs ~16ms for full get_one_face() at 1080p.
+    """
+    from insightface.app.common import Face
+    fa = get_face_analyser()
+    bboxes, kpss = fa.det_model.detect(frame, max_num=0, metric='default')
+    if bboxes.shape[0] == 0:
+        return None
+    idx = int(bboxes[:, 0].argmin())
+    return Face(bbox=bboxes[idx, :4], kps=kpss[idx], det_score=bboxes[idx, 4])
+
+
+def detect_many_faces_fast(frame: Frame) -> Any:
+    """Detection-only multi-face — skips landmark and recognition."""
+    from insightface.app.common import Face
+    fa = get_face_analyser()
+    bboxes, kpss = fa.det_model.detect(frame, max_num=0, metric='default')
+    if bboxes.shape[0] == 0:
+        return None
+    return [Face(bbox=bboxes[i, :4], kps=kpss[i], det_score=bboxes[i, 4])
+            for i in range(bboxes.shape[0])]
+
+
+def ensure_landmarks(frame: Frame, faces: Any) -> None:
+    """Run the 2d106 landmark model in-place on faces that lack it.
+
+    The fast webcam path (detect_one_face_fast / detect_many_faces_fast)
+    produces detection-only Face objects with no ``landmark_2d_106``.
+    Mouth masking needs those landmarks, so add them on demand only when
+    the feature is active — keeping the fast path fast otherwise.
+    """
+    if faces is None:
+        return
+    if not isinstance(faces, (list, tuple)):
+        faces = [faces]
+
+    fa = get_face_analyser()
+    lmk_model = fa.models.get("landmark_2d_106")
+    if lmk_model is None:
+        return
+
+    for face in faces:
+        if face is None:
+            continue
+        # insightface Face is a dict; missing keys raise AttributeError,
+        # so getattr(..., None) is the safe presence check.
+        if getattr(face, "landmark_2d_106", None) is None:
+            try:
+                lmk_model.get(frame, face)
+            except Exception as e:  # pragma: no cover - never break the swap
+                print(f"Error computing 2d106 landmarks: {e}")
+
+
 def has_valid_map() -> bool:
     for map in modules.globals.source_target_map:
         if "source" in map and "target" in map:
@@ -196,4 +365,4 @@ def dump_faces(centroids: Any, frame_face_embeddings: list):
 
                     if temp_frame[int(y_min):int(y_max), int(x_min):int(x_max)].size > 0:
                         cv2.imwrite(temp_directory_path + f"/{i}/{frame['frame']}_{j}.png", temp_frame[int(y_min):int(y_max), int(x_min):int(x_max)])
-                j += 1
+                j += 1

modules/globals.py

@@ -63,6 +63,7 @@ show_mouth_mask_box: bool = False  # Visualize the mouth mask area (for debuggin
 mask_feather_ratio: int = 12       # Denominator for feathering calculation (higher = smaller feather)
 mask_down_size: float = 0.1        # Expansion factor for lower lip mask (relative)
 mask_size: float = 1.0             # Expansion factor for upper lip mask (relative)
+mouth_mask_size: float = 0.0       # Mouth mask size (0-100; 0=off, 100=mouth to chin)
 
 # --- START: Added for Frame Interpolation ---
 enable_interpolation: bool = True # Toggle temporal smoothing
@@ -70,3 +71,6 @@ interpolation_weight: float = 0  # Blend weight for current frame (0.0-1.0). Low
 # --- END: Added for Frame Interpolation ---
 
 # --- END OF FILE globals.py ---
+
+import threading
+dml_lock = threading.Lock()

modules/gpu_processing.py

@@ -18,6 +18,7 @@ Usage
 
 from __future__ import annotations
 
+import os
 import cv2
 import numpy as np
 from typing import Tuple, Optional
@@ -27,27 +28,25 @@ from typing import Tuple, Optional
 # ---------------------------------------------------------------------------
 CUDA_AVAILABLE: bool = False
 
-try:
-    # cv2.cuda.GpuMat is only present when OpenCV is compiled with CUDA
-    _test_mat = cv2.cuda.GpuMat()
-    # Verify we have the required filter / image-processing functions
-    _has_gauss = hasattr(cv2.cuda, "createGaussianFilter")
-    _has_resize = hasattr(cv2.cuda, "resize")
-    _has_cvt = hasattr(cv2.cuda, "cvtColor")
-    if _has_gauss and _has_resize and _has_cvt:
-        CUDA_AVAILABLE = True
-        print("[gpu_processing] OpenCV CUDA support detected – GPU-accelerated processing enabled.")
-    else:
-        missing = []
-        if not _has_gauss:
-            missing.append("createGaussianFilter")
-        if not _has_resize:
-            missing.append("resize")
-        if not _has_cvt:
-            missing.append("cvtColor")
-        print(f"[gpu_processing] cv2.cuda.GpuMat exists but missing: {', '.join(missing)} – falling back to CPU.")
-except Exception:
-    print("[gpu_processing] OpenCV CUDA not available – using CPU fallback for all operations.")
+# OpenCV CUDA per-operation acceleration is DISABLED by default.
+# Each gpu_* call uploads to GPU, processes, then downloads back to CPU.
+# At webcam resolution (~960x540) this upload/download overhead far exceeds
+# the time saved on the actual operation, making it slower than pure CPU.
+# The heavy lifting (face detection, swap, enhancement) runs on GPU via
+# ONNX Runtime's CUDAExecutionProvider, which is where GPU matters.
+#
+# To force-enable, set OPENCV_CUDA_PROCESSING=1 in your environment.
+if os.environ.get("OPENCV_CUDA_PROCESSING") == "1":
+    try:
+        _test_mat = cv2.cuda.GpuMat()
+        _has_gauss = hasattr(cv2.cuda, "createGaussianFilter")
+        _has_resize = hasattr(cv2.cuda, "resize")
+        _has_cvt = hasattr(cv2.cuda, "cvtColor")
+        if _has_gauss and _has_resize and _has_cvt:
+            CUDA_AVAILABLE = True
+            print("[gpu_processing] OpenCV CUDA processing enabled via OPENCV_CUDA_PROCESSING=1.")
+    except Exception:
+        pass
 
 
 # ---------------------------------------------------------------------------

modules/metadata.py

@@ -1,3 +1,3 @@
 name = 'Deep-Live-Cam'
-version = '2.0.3c'
+version = '2.1.5'
 edition = 'GitHub Edition'

modules/onnx_optimize.py

@@ -0,0 +1,550 @@
+"""ONNX model optimizations for CoreML execution on Apple Silicon.
+
+Each pass eliminates a different CPU↔ANE round-trip that ORT's CoreML EP
+would otherwise introduce:
+
+1. **Shape/Gather constant folding** — Dynamic ``Shape`` → ``Gather`` chains
+   (e.g. for FPN upsample target sizes in RetinaFace) force ops onto CPU even
+   when the input dimensions are known at load time.  We run ONNX shape
+   inference with the known input size and replace these chains with constants.
+   Float32-noise-level differences only (max ~6e-6).
+
+2. **Pad(reflect) decomposition** — CoreML doesn't support ``Pad(mode=reflect)``.
+   Models using reflect padding (e.g. inswapper_128) get split into many CoreML
+   subgraphs with CPU fallbacks between each.  We rewrite each ``Pad(reflect)``
+   as equivalent ``Slice`` + ``Concat`` ops that CoreML handles natively.
+   Bit-for-bit identical output. (Fixed upstream in microsoft/onnxruntime#28073.)
+
+3. **Split → Slice decomposition** — CoreML's EP doesn't support the ONNX
+   ``Split`` op, causing partition boundaries in models with channel-wise
+   splits (e.g. GFPGAN's SFT modulation). Each 2-way Split becomes two Slices.
+
+4. **Scalar Gather widening** — ORT's CoreML EP rejects ``Gather`` nodes with
+   rank-0 (scalar) indices. StyleGAN-derived models (GFPGAN) slice per-layer
+   style codes using exactly this pattern. We widen each scalar index to
+   ``[1]`` and squeeze the added axis on the Gather output.
+   (Filed upstream as microsoft/onnxruntime#28180.)
+
+All passes are cached on disk with a ``_coreml`` suffix so the rewrite cost
+is paid only once per model.
+"""
+
+import os
+import platform
+
+import numpy as np
+
+IS_APPLE_SILICON = platform.system() == "Darwin" and platform.machine() == "arm64"
+
+
+def optimize_for_coreml(model_path: str, input_shape: tuple = None) -> str:
+    """Return path to a CoreML-optimized ONNX model.
+
+    Applies all applicable optimizations and caches the result next to
+    the original model (with ``_coreml`` suffix).
+
+    Args:
+        model_path: Path to the original ONNX model.
+        input_shape: Optional fixed input shape (e.g. ``(1, 3, 640, 640)``).
+            When provided, enables Shape/Gather constant folding.
+
+    Returns the optimized path, or the original path if no optimizations
+    apply or we're not on Apple Silicon.
+    """
+    if not IS_APPLE_SILICON:
+        return model_path
+
+    base, ext = os.path.splitext(model_path)
+    optimized_path = f"{base}_coreml{ext}"
+    if os.path.exists(optimized_path):
+        if os.path.getmtime(optimized_path) >= os.path.getmtime(model_path):
+            return optimized_path
+
+    import onnx
+    from onnx import numpy_helper
+
+    model = onnx.load(model_path)
+    changed = False
+
+    if _fold_shape_gather(model, input_shape):
+        changed = True
+
+    # TODO(ort>=1.26): drop this pass. Fixed upstream by microsoft/onnxruntime#28073.
+    if _decompose_reflect_pad(model):
+        changed = True
+
+    if _decompose_split(model):
+        changed = True
+
+    # TODO: drop this pass once microsoft/onnxruntime#28180 ships. The CoreML
+    # Gather op builder rejects rank-0 (scalar) indices; we widen them to [1]
+    # + Squeeze so StyleGAN-family models (GFPGAN) stay on ANE.
+    if _rewrite_scalar_gather(model):
+        changed = True
+
+    if not changed:
+        return model_path
+
+    # Preserve insightface's emap convention: the INSwapper class reads
+    # graph.initializer[-1] as the embedding map.  If the original model
+    # had a (512, 512) matrix as its last initializer, keep it last.
+    _preserve_emap_position(model, numpy_helper)
+
+    onnx.save(model, optimized_path)
+    return optimized_path
+
+
+# ---------------------------------------------------------------------------
+# Pass 1: Fold Shape → Gather chains into constants
+# ---------------------------------------------------------------------------
+
+def _fold_shape_gather(model, input_shape) -> bool:
+    """Replace dynamic Shape→Gather chains with constants when input size is known.
+
+    Only removes a Shape node when ALL of its consumers are Gather nodes
+    that are also being folded.  This prevents breaking graphs where
+    a Shape output feeds into other ops as well.
+    """
+    if input_shape is None:
+        return False
+
+    from onnx import numpy_helper, shape_inference
+
+    graph = model.graph
+
+    # Set fixed input dimensions for shape inference
+    inp = graph.input[0]
+    dims = inp.type.tensor_type.shape.dim
+    for i, size in enumerate(input_shape):
+        if i < len(dims):
+            dims[i].dim_value = size
+
+    try:
+        model_inferred = shape_inference.infer_shapes(model)
+    except Exception:
+        return False
+
+    # Extract inferred shapes
+    value_shapes = {}
+    for vi in list(model_inferred.graph.value_info) + list(graph.input) + list(graph.output):
+        shape_dims = vi.type.tensor_type.shape.dim
+        shape = []
+        for d in shape_dims:
+            if d.dim_value > 0:
+                shape.append(d.dim_value)
+            else:
+                shape.append(None)
+        value_shapes[vi.name] = shape
+
+    inits = {init.name: numpy_helper.to_array(init) for init in graph.initializer}
+
+    # Build consumer map: output_name → list of consuming nodes
+    consumers = {}
+    for node in graph.node:
+        for i in node.input:
+            consumers.setdefault(i, []).append(node)
+
+    # Also check graph outputs — an output name consumed by the graph
+    # output list must not be removed
+    graph_output_names = {o.name for o in graph.output}
+
+    # Find Shape nodes with fully-known output
+    shape_constants = {}
+    for node in graph.node:
+        if node.op_type == "Shape":
+            inp_shape = value_shapes.get(node.input[0])
+            if inp_shape and all(isinstance(d, int) for d in inp_shape):
+                shape_constants[node.output[0]] = np.array(inp_shape, dtype=np.int64)
+
+    if not shape_constants:
+        return False
+
+    # Find Gather nodes consuming Shape constants
+    gather_constants = {}
+    for node in graph.node:
+        if node.op_type == "Gather" and node.input[0] in shape_constants:
+            idx_name = node.input[1]
+            if idx_name in inits:
+                idx = int(inits[idx_name])
+                val = int(shape_constants[node.input[0]][idx])
+                gather_constants[node.output[0]] = np.array(val, dtype=np.int64)
+
+    if not gather_constants:
+        return False
+
+    # Determine which Gather nodes to fold (always safe — we replace
+    # the output with a constant initializer)
+    gather_remove_ids = set()
+    for node in graph.node:
+        if node.op_type == "Gather" and node.output[0] in gather_constants:
+            gather_remove_ids.add(id(node))
+
+    # Determine which Shape nodes are safe to remove: only if ALL
+    # consumers of the Shape output are Gather nodes being folded,
+    # and the output isn't a graph output.
+    shape_remove_ids = set()
+    for node in graph.node:
+        if node.op_type == "Shape" and node.output[0] in shape_constants:
+            out_name = node.output[0]
+            if out_name in graph_output_names:
+                continue
+            node_consumers = consumers.get(out_name, [])
+            if all(id(c) in gather_remove_ids for c in node_consumers):
+                shape_remove_ids.add(id(node))
+
+    remove_ids = gather_remove_ids | shape_remove_ids
+
+    # Add Gather output constants as initializers
+    existing = {i.name for i in graph.initializer}
+    for name, val in gather_constants.items():
+        if name not in existing:
+            graph.initializer.append(numpy_helper.from_array(val, name=name))
+
+    new_nodes = [n for n in graph.node if id(n) not in remove_ids]
+    del graph.node[:]
+    graph.node.extend(new_nodes)
+    return True
+
+
+# ---------------------------------------------------------------------------
+# Pass 2: Decompose Pad(reflect) → Slice + Concat
+#
+# TEMPORARY: fixed upstream in microsoft/onnxruntime#28073 (merged 2026-04-20).
+# Once the ORT floor is >= 1.26.0, MLProgram handles Pad(mode=reflect) natively
+# via MIL tensor_operation.pad and this entire pass can be deleted.
+# ---------------------------------------------------------------------------
+
+def _decompose_reflect_pad(model) -> bool:
+    """Rewrite Pad(reflect) as Slice+Concat sequences CoreML can handle."""
+    from onnx import numpy_helper, helper
+
+    graph = model.graph
+    inits = {init.name: numpy_helper.to_array(init) for init in graph.initializer}
+
+    reflect_pads = []
+    for node in graph.node:
+        if node.op_type == "Pad":
+            mode = "constant"
+            for attr in node.attribute:
+                if attr.name == "mode":
+                    mode = attr.s.decode()
+            if mode == "reflect" and len(node.input) > 1 and node.input[1] in inits:
+                reflect_pads.append(node)
+
+    if not reflect_pads:
+        return False
+
+    existing_names = {i.name for i in graph.initializer}
+
+    def ensure_const(name, value):
+        if name not in existing_names:
+            graph.initializer.append(
+                numpy_helper.from_array(np.array(value, dtype=np.int64), name=name)
+            )
+            existing_names.add(name)
+
+    ensure_const("_rp_ax2", [2])
+    ensure_const("_rp_ax3", [3])
+
+    max_pad = 0
+    for node in reflect_pads:
+        pads = inits[node.input[1]].tolist()
+        max_pad = max(max_pad, int(pads[2]), int(pads[3]))
+
+    for v in range(1, max_pad + 2):
+        ensure_const(f"_rp_p{v}", [v])
+        ensure_const(f"_rp_n{v}", [-v])
+
+    _counter = [0]
+
+    def uid():
+        _counter[0] += 1
+        return _counter[0]
+
+    pad_ids = {id(n) for n in reflect_pads}
+    pad_init_names = set()
+
+    new_nodes = []
+    for node in graph.node:
+        if id(node) not in pad_ids:
+            new_nodes.append(node)
+            continue
+
+        pads = inits[node.input[1]].tolist()
+        h_pad, w_pad = int(pads[2]), int(pads[3])
+
+        for inp in node.input[1:]:
+            if inp in inits:
+                pad_init_names.add(inp)
+
+        current = node.input[0]
+
+        if h_pad > 0:
+            top = []
+            for i in range(h_pad, 0, -1):
+                name = f"_rp_t{uid()}"
+                new_nodes.append(helper.make_node(
+                    "Slice",
+                    inputs=[current, f"_rp_p{i}", f"_rp_p{i+1}", "_rp_ax2"],
+                    outputs=[name],
+                ))
+                top.append(name)
+
+            bot = []
+            for i in range(1, h_pad + 1):
+                name = f"_rp_b{uid()}"
+                new_nodes.append(helper.make_node(
+                    "Slice",
+                    inputs=[current, f"_rp_n{i+1}", f"_rp_n{i}", "_rp_ax2"],
+                    outputs=[name],
+                ))
+                bot.append(name)
+
+            h_out = f"_rp_h{uid()}"
+            new_nodes.append(helper.make_node(
+                "Concat", inputs=top + [current] + bot, outputs=[h_out], axis=2
+            ))
+            current = h_out
+
+        if w_pad > 0:
+            left = []
+            for i in range(w_pad, 0, -1):
+                name = f"_rp_l{uid()}"
+                new_nodes.append(helper.make_node(
+                    "Slice",
+                    inputs=[current, f"_rp_p{i}", f"_rp_p{i+1}", "_rp_ax3"],
+                    outputs=[name],
+                ))
+                left.append(name)
+
+            right = []
+            for i in range(1, w_pad + 1):
+                name = f"_rp_r{uid()}"
+                new_nodes.append(helper.make_node(
+                    "Slice",
+                    inputs=[current, f"_rp_n{i+1}", f"_rp_n{i}", "_rp_ax3"],
+                    outputs=[name],
+                ))
+                right.append(name)
+
+            new_nodes.append(helper.make_node(
+                "Concat",
+                inputs=left + [current] + right,
+                outputs=[node.output[0]],
+                axis=3,
+            ))
+        elif h_pad > 0:
+            new_nodes.append(helper.make_node(
+                "Identity", inputs=[current], outputs=[node.output[0]]
+            ))
+
+    # Remove old Pad initializers
+    clean_inits = [i for i in graph.initializer if i.name not in pad_init_names]
+    del graph.initializer[:]
+    graph.initializer.extend(clean_inits)
+
+    del graph.node[:]
+    graph.node.extend(new_nodes)
+    return True
+
+
+# ---------------------------------------------------------------------------
+# Pass 3: Decompose Split → Slice pairs
+# ---------------------------------------------------------------------------
+
+def _decompose_split(model) -> bool:
+    """Rewrite Split(axis=1) as Slice pairs that CoreML can handle.
+
+    CoreML's EP doesn't support the ONNX ``Split`` op, causing partition
+    boundaries in models that use channel-wise splits (e.g. GFPGAN's SFT
+    modulation layers).  Each Split with two outputs becomes two Slice ops.
+    """
+    from onnx import numpy_helper, helper
+
+    graph = model.graph
+
+    splits = []
+    for node in graph.node:
+        if node.op_type == "Split":
+            axis = 0
+            split_sizes = []
+            for attr in node.attribute:
+                if attr.name == "axis":
+                    axis = attr.i
+                if attr.name == "split":
+                    split_sizes = list(attr.ints)
+            if axis == 1 and len(split_sizes) == 2 and len(node.output) == 2:
+                splits.append((node, split_sizes))
+
+    if not splits:
+        return False
+
+    existing = {i.name for i in graph.initializer}
+
+    def ensure_const(name, value):
+        if name not in existing:
+            graph.initializer.append(
+                numpy_helper.from_array(np.array(value, dtype=np.int64), name=name)
+            )
+            existing.add(name)
+
+    ensure_const("_sp_ax1", [1])
+
+    # Collect all needed boundary constants
+    for _, (a, b) in splits:
+        ensure_const(f"_sp_s0", [0])
+        ensure_const(f"_sp_s{a}", [a])
+        ensure_const(f"_sp_s{a + b}", [a + b])
+
+    split_ids = {id(node) for node, _ in splits}
+    replacements = {}
+    for node, (a, b) in splits:
+        slice0 = helper.make_node(
+            "Slice",
+            inputs=[node.input[0], "_sp_s0", f"_sp_s{a}", "_sp_ax1"],
+            outputs=[node.output[0]],
+        )
+        slice1 = helper.make_node(
+            "Slice",
+            inputs=[node.input[0], f"_sp_s{a}", f"_sp_s{a + b}", "_sp_ax1"],
+            outputs=[node.output[1]],
+        )
+        replacements[id(node)] = [slice0, slice1]
+
+    new_nodes = []
+    for node in graph.node:
+        if id(node) in split_ids:
+            new_nodes.extend(replacements[id(node)])
+        else:
+            new_nodes.append(node)
+
+    del graph.node[:]
+    graph.node.extend(new_nodes)
+    return True
+
+
+# ---------------------------------------------------------------------------
+# Pass 4: Widen scalar Gather indices to [1] + Squeeze
+#
+# TEMPORARY: filed upstream as microsoft/onnxruntime#28180. ORT's CoreML EP
+# GatherOpBuilder::IsOpSupportedImpl rejects rank-0 (scalar) indices with
+# `Gather does not support scalar 'indices'`. The builder's own comment
+# describes the workaround (promote to [1], squeeze the added axis) but
+# doesn't apply it. We do the same thing at the ONNX level so StyleGAN-
+# family models (GFPGAN is the hot example — 16 per-layer style-code
+# slices) don't split the CoreML subgraph. Once the upstream fix ships
+# and the ORT floor is raised, delete this pass.
+# ---------------------------------------------------------------------------
+
+def _rewrite_scalar_gather(model) -> bool:
+    """Rewrite Gather(data, scalar_idx) as Gather(data, [scalar_idx]) + Squeeze.
+
+    Only touches Gather nodes whose index is a rank-0 int64 constant or
+    initializer; everything else passes through unchanged. The rewrite
+    is semantically identical — indices get an added leading axis, the
+    Squeeze removes it after the gather.
+    """
+    from onnx import numpy_helper, helper, TensorProto
+
+    graph = model.graph
+
+    # Opset 13 moved Squeeze's axes from attribute to input.
+    opset = next(
+        (o.version for o in model.opset_import if o.domain in ("", "ai.onnx")),
+        11,
+    )
+
+    const_values = {}
+    for n in graph.node:
+        if n.op_type == "Constant":
+            for a in n.attribute:
+                if a.name == "value":
+                    const_values[n.output[0]] = a.t
+    init_values = {i.name: i for i in graph.initializer}
+
+    def scalar_int64(name):
+        """Return int value if `name` resolves to a rank-0 int64 constant, else None."""
+        tensor = const_values.get(name) or init_values.get(name)
+        if tensor is None or tensor.data_type != TensorProto.INT64:
+            return None
+        arr = numpy_helper.to_array(tensor)
+        return int(arr) if arr.ndim == 0 else None
+
+    rewrote = 0
+    new_nodes = []
+    for n in graph.node:
+        if n.op_type == "Gather":
+            val = scalar_int64(n.input[1])
+            if val is not None:
+                axis = next((a.i for a in n.attribute if a.name == "axis"), 0)
+                idx_1d_name = f"{n.input[1]}_1d_{rewrote}"
+                idx_const = helper.make_node(
+                    "Constant",
+                    inputs=[],
+                    outputs=[idx_1d_name],
+                    value=helper.make_tensor(idx_1d_name, TensorProto.INT64, [1], [val]),
+                )
+                gather_out = f"{n.output[0]}_pre_squeeze_{rewrote}"
+                new_gather = helper.make_node(
+                    "Gather",
+                    inputs=[n.input[0], idx_1d_name],
+                    outputs=[gather_out],
+                    name=n.name,
+                    axis=axis,
+                )
+                if opset < 13:
+                    squeeze = helper.make_node(
+                        "Squeeze",
+                        inputs=[gather_out],
+                        outputs=[n.output[0]],
+                        name=(n.name or "gather") + "_squeeze",
+                        axes=[axis],
+                    )
+                    new_nodes.extend([idx_const, new_gather, squeeze])
+                else:
+                    axes_name = f"{idx_1d_name}_sq_axes"
+                    axes_const = helper.make_node(
+                        "Constant",
+                        inputs=[],
+                        outputs=[axes_name],
+                        value=helper.make_tensor(axes_name, TensorProto.INT64, [1], [axis]),
+                    )
+                    squeeze = helper.make_node(
+                        "Squeeze",
+                        inputs=[gather_out, axes_name],
+                        outputs=[n.output[0]],
+                        name=(n.name or "gather") + "_squeeze",
+                    )
+                    new_nodes.extend([idx_const, axes_const, new_gather, squeeze])
+                rewrote += 1
+                continue
+        new_nodes.append(n)
+
+    if rewrote == 0:
+        return False
+
+    del graph.node[:]
+    graph.node.extend(new_nodes)
+    return True
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def _preserve_emap_position(model, numpy_helper):
+    """Keep the insightface emap (512×512 matrix) as the last initializer."""
+    graph = model.graph
+    emap_init = None
+    for init in graph.initializer:
+        if not init.name.startswith("_rp_"):
+            arr = numpy_helper.to_array(init)
+            if len(arr.shape) == 2 and arr.shape[0] == 512 and arr.shape[1] == 512:
+                emap_init = init
+                break
+
+    if emap_init is not None:
+        inits = [i for i in graph.initializer if i.name != emap_init.name]
+        del graph.initializer[:]
+        graph.initializer.extend(inits)
+        graph.initializer.append(emap_init)

modules/platform_info.py

@@ -0,0 +1,80 @@
+"""Centralized platform + accelerator detection.
+
+Imported once at startup to expose typed flags the rest of the codebase
+can branch on without re-querying `platform`, `torch.cuda`, or
+`onnxruntime.get_available_providers()` repeatedly.
+
+The banner printed by :func:`print_banner` is the single user-facing
+report of which code path the app will take.
+"""
+from __future__ import annotations
+
+import platform as _platform
+import sys
+from typing import List, Tuple
+
+IS_WINDOWS: bool = _platform.system() == "Windows"
+IS_MACOS: bool = _platform.system() == "Darwin"
+IS_LINUX: bool = _platform.system() == "Linux"
+IS_APPLE_SILICON: bool = IS_MACOS and _platform.machine() == "arm64"
+
+
+def _detect_torch_cuda() -> bool:
+    try:
+        import torch  # noqa: WPS433 — local import, avoid hard dep at module load
+        return bool(torch.cuda.is_available())
+    except Exception:
+        return False
+
+
+def _detect_onnx_providers() -> List[str]:
+    try:
+        import onnxruntime
+        return list(onnxruntime.get_available_providers())
+    except Exception:
+        return []
+
+
+HAS_TORCH_CUDA: bool = _detect_torch_cuda()
+ONNX_PROVIDERS: List[str] = _detect_onnx_providers()
+HAS_CUDA_PROVIDER: bool = "CUDAExecutionProvider" in ONNX_PROVIDERS
+HAS_COREML_PROVIDER: bool = "CoreMLExecutionProvider" in ONNX_PROVIDERS
+HAS_DML_PROVIDER: bool = "DmlExecutionProvider" in ONNX_PROVIDERS
+
+
+def camera_backends() -> List[Tuple[int, int]]:
+    """Return an ordered list of ``(device_index, cv2_backend)`` attempts.
+
+    Windows prefers MSMF (60fps capable) with DirectShow as fallback.
+    macOS/Linux use the default backend (AVFoundation / V4L2).
+    """
+    import cv2
+    if IS_WINDOWS:
+        return [
+            (0, cv2.CAP_MSMF),
+            (0, cv2.CAP_DSHOW),
+            (0, cv2.CAP_ANY),
+        ]
+    return [(0, cv2.CAP_ANY)]
+
+
+def accelerator_label() -> str:
+    if HAS_CUDA_PROVIDER:
+        return "CUDA (NVIDIA)"
+    if IS_APPLE_SILICON and HAS_COREML_PROVIDER:
+        return "CoreML (Apple Neural Engine)"
+    if HAS_COREML_PROVIDER:
+        return "CoreML"
+    if HAS_DML_PROVIDER:
+        return "DirectML"
+    return "CPU"
+
+
+def print_banner() -> None:
+    """Print a one-line summary of the platform + accelerator selection."""
+    os_label = f"{_platform.system()} {_platform.machine()}"
+    print(
+        f"[platform] {os_label} | python {sys.version.split()[0]} | "
+        f"accelerator: {accelerator_label()} | providers: {ONNX_PROVIDERS}",
+        flush=True,
+    )

modules/processors/frame/_onnx_enhancer.py

@@ -21,10 +21,104 @@ IS_APPLE_SILICON = platform.system() == "Darwin" and platform.machine() == "arm6
 THREAD_SEMAPHORE = threading.Semaphore(min(max(1, (os.cpu_count() or 1)), 8))
 
 
+def build_provider_config(providers=None):
+    """Wrap raw provider name strings with optimised CUDA / CoreML options.
+
+    Providers that are already ``(name, options_dict)`` tuples are passed
+    through unchanged.  Non-CUDA providers are left as bare strings.
+    """
+    if providers is None:
+        providers = modules.globals.execution_providers
+
+    config = []
+    for p in providers:
+        if isinstance(p, tuple):
+            # Already configured – pass through
+            config.append(p)
+        elif p == "CUDAExecutionProvider":
+            # Use bare provider — ONNX Runtime's defaults are fastest on
+            # modern GPUs (Blackwell/sm_120).  Custom options like
+            # EXHAUSTIVE cudnn_conv_algo_search hurt performance on these
+            # architectures.
+            config.append(p)
+        elif p == "CoreMLExecutionProvider" and IS_APPLE_SILICON:
+            config.append((
+                "CoreMLExecutionProvider",
+                {
+                    "ModelFormat": "MLProgram",
+                    "MLComputeUnits": "ALL",
+                    "AllowLowPrecisionAccumulationOnGPU": 1,
+                },
+            ))
+        else:
+            config.append(p)
+    return config
+
+
+def run_inference(session: onnxruntime.InferenceSession,
+                  input_name: str,
+                  input_tensor: "np.ndarray") -> "np.ndarray":
+    """Run ONNX inference, using IO binding when a CUDA session is active.
+
+    IO binding avoids redundant host↔device copies by transferring the
+    input tensor directly to GPU memory and letting ONNX Runtime allocate
+    the output on the device.  Falls back to the standard ``session.run``
+    path for non-CUDA providers or if binding fails.
+    """
+    if "CUDAExecutionProvider" in session.get_providers():
+        try:
+            io_binding = session.io_binding()
+
+            # Input: numpy → GPU
+            ort_input = onnxruntime.OrtValue.ortvalue_from_numpy(
+                input_tensor, "cuda", 0,
+            )
+            io_binding.bind_ortvalue_input(input_name, ort_input)
+
+            # Output: allocate on GPU (avoids a CPU-side allocation)
+            output_name = session.get_outputs()[0].name
+            io_binding.bind_output(output_name, "cuda", 0)
+
+            session.run_with_iobinding(io_binding)
+
+            return io_binding.get_outputs()[0].numpy()
+        except Exception:
+            # Fall back to standard path (e.g. ORT version mismatch,
+            # unsupported op, or VRAM pressure)
+            pass
+
+    return session.run(None, {input_name: input_tensor})[0]
+
+
 def create_onnx_session(model_path: str) -> onnxruntime.InferenceSession:
-    """Create an ONNX Runtime session using the configured execution providers."""
-    providers = modules.globals.execution_providers
-    session = onnxruntime.InferenceSession(model_path, providers=providers)
+    """Create an ONNX Runtime session with optimised provider config.
+
+    On Apple Silicon, applies CoreML graph optimizations (Pad decomposition,
+    Shape/Gather folding, Split decomposition) to reduce CPU↔ANE partition
+    boundaries.
+    """
+    if IS_APPLE_SILICON:
+        from modules.onnx_optimize import optimize_for_coreml
+        # Infer input shape from the model for Shape/Gather folding
+        try:
+            import onnx
+            m = onnx.load(model_path)
+            inp = m.graph.input[0]
+            dims = inp.type.tensor_type.shape.dim
+            shape = tuple(d.dim_value for d in dims if d.dim_value > 0)
+            input_shape = shape if len(shape) == 4 else None
+        except Exception:
+            input_shape = None
+        model_path = optimize_for_coreml(model_path, input_shape=input_shape)
+
+    providers = build_provider_config()
+    session_options = onnxruntime.SessionOptions()
+    session_options.graph_optimization_level = (
+        onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+    )
+    session = onnxruntime.InferenceSession(
+        model_path, sess_options=session_options, providers=providers,
+    )
     return session
 
 
@@ -118,7 +212,8 @@ def enhance_face_onnx(
 
     blob = preprocess_face(face_crop, input_size)
     with THREAD_SEMAPHORE:
-        output = session.run(None, {session.get_inputs()[0].name: blob})[0]
+        input_name = session.get_inputs()[0].name
+        output = run_inference(session, input_name, blob)
     enhanced = postprocess_face(output)
 
     # Create mask for blending (feathered edges)

modules/processors/frame/core.py

@@ -1,12 +1,17 @@
+import os
+import subprocess
 import sys
 import importlib
 from concurrent.futures import ThreadPoolExecutor
 from types import ModuleType
 from typing import Any, List, Callable
+
+import numpy as np
 from tqdm import tqdm
 
 import modules
-import modules.globals                   
+import modules.globals
+from modules.face_analyser import get_one_face
 
 FRAME_PROCESSORS_MODULES: List[ModuleType] = []
 FRAME_PROCESSORS_INTERFACE = [
@@ -107,3 +112,295 @@ def process_video(source_path: str, frame_paths: list[str], process_frames: Call
     with tqdm(total=total, desc='Processing', unit='frame', dynamic_ncols=True, bar_format=progress_bar_format) as progress:
         progress.set_postfix({'execution_providers': modules.globals.execution_providers, 'execution_threads': modules.globals.execution_threads, 'max_memory': modules.globals.max_memory})
         multi_process_frame(source_path, frame_paths, process_frames, progress)
+
+
+def process_video_in_memory(source_path: str, target_path: str, fps: float) -> bool:
+    """Process video frames in-memory using FFmpeg pipes, eliminating disk I/O.
+
+    Reads raw frames from the source video via an FFmpeg decoder pipe, runs each
+    frame through all active frame processors sequentially, and writes the
+    result directly to an FFmpeg encoder pipe.  This avoids extracting frames to
+    PNG on disk, which is the biggest I/O bottleneck in the disk-based pipeline.
+
+    Returns True on success, False on failure (caller should fall back to the
+    disk-based pipeline).
+    """
+    import cv2
+    from modules.face_analyser import get_one_face
+    from modules.utilities import (
+        get_video_dimensions,
+        estimate_frame_count,
+        get_temp_output_path,
+    )
+
+    temp_output_path = get_temp_output_path(target_path)
+
+    # --- Pre-load source face (needed by face_swapper in simple mode) ---
+    source_face = None
+    if source_path and os.path.exists(source_path):
+        source_img = cv2.imread(source_path)
+        if source_img is not None:
+            source_face = get_one_face(source_img)
+            del source_img
+        if source_face is None:
+            print("[DLC.CORE] Warning: No face detected in source image. "
+                  "Face swapping will be skipped.")
+
+    # --- Collect frame processors & reset per-video state ---
+    frame_processors = get_frame_processors_modules(modules.globals.frame_processors)
+    for fp in frame_processors:
+        if hasattr(fp, 'PREVIOUS_FRAME_RESULT'):
+            fp.PREVIOUS_FRAME_RESULT = None
+
+    # --- Video metadata ---
+    try:
+        width, height = get_video_dimensions(target_path)
+    except Exception as e:
+        print(f"[DLC.CORE] Failed to get video dimensions: {e}")
+        return False
+
+    total_frames = estimate_frame_count(target_path, fps)
+    frame_size = width * height * 3
+
+    # --- Build encoder arguments ---
+    encoder = modules.globals.video_encoder
+    encoder_options: List[str] = []
+    is_hw_encoder = False
+
+    if 'CUDAExecutionProvider' in modules.globals.execution_providers:
+        if encoder == 'libx264':
+            encoder = 'h264_nvenc'
+            is_hw_encoder = True
+            encoder_options = [
+                '-preset', 'p4', '-tune', 'hq', '-rc', 'vbr',
+                '-cq', str(modules.globals.video_quality), '-b:v', '0',
+            ]
+        elif encoder == 'libx265':
+            encoder = 'hevc_nvenc'
+            is_hw_encoder = True
+            encoder_options = [
+                '-preset', 'p4', '-tune', 'hq', '-rc', 'vbr',
+                '-cq', str(modules.globals.video_quality), '-b:v', '0',
+            ]
+    elif 'DmlExecutionProvider' in modules.globals.execution_providers:
+        if encoder == 'libx264':
+            encoder = 'h264_amf'
+            is_hw_encoder = True
+            encoder_options = [
+                '-quality', 'quality', '-rc', 'vbr_latency',
+                '-qp_i', str(modules.globals.video_quality),
+                '-qp_p', str(modules.globals.video_quality),
+            ]
+        elif encoder == 'libx265':
+            encoder = 'hevc_amf'
+            is_hw_encoder = True
+            encoder_options = [
+                '-quality', 'quality', '-rc', 'vbr_latency',
+                '-qp_i', str(modules.globals.video_quality),
+                '-qp_p', str(modules.globals.video_quality),
+            ]
+
+    if not is_hw_encoder:
+        if encoder == 'libx264':
+            encoder_options = [
+                '-preset', 'medium',
+                '-crf', str(modules.globals.video_quality),
+                '-tune', 'film',
+            ]
+        elif encoder == 'libx265':
+            encoder_options = [
+                '-preset', 'medium',
+                '-crf', str(modules.globals.video_quality),
+                '-x265-params', 'log-level=error',
+            ]
+        elif encoder == 'libvpx-vp9':
+            encoder_options = [
+                '-crf', str(modules.globals.video_quality),
+                '-b:v', '0', '-cpu-used', '2',
+            ]
+
+    # --- Attempt pipeline (hw encoder first, then sw fallback) ---
+    encoders_to_try = [(encoder, encoder_options)]
+    if is_hw_encoder:
+        # Software fallback
+        sw_encoder = 'libx264'
+        sw_options = [
+            '-preset', 'medium',
+            '-crf', str(modules.globals.video_quality),
+            '-tune', 'film',
+        ]
+        encoders_to_try.append((sw_encoder, sw_options))
+
+    for attempt, (enc, enc_opts) in enumerate(encoders_to_try):
+        # Reset interpolation state on retry
+        if attempt > 0:
+            for fp in frame_processors:
+                if hasattr(fp, 'PREVIOUS_FRAME_RESULT'):
+                    fp.PREVIOUS_FRAME_RESULT = None
+
+        success = _run_pipe_pipeline(
+            target_path, temp_output_path, fps,
+            source_face, frame_processors,
+            width, height, frame_size, total_frames,
+            enc, enc_opts,
+        )
+        if success:
+            return True
+
+        if attempt == 0 and is_hw_encoder:
+            print(f"[DLC.CORE] Hardware encoder '{enc}' failed, "
+                  f"retrying with software encoder...")
+
+    return False
+
+
+def _run_pipe_pipeline(
+    target_path: str,
+    temp_output_path: str,
+    fps: float,
+    source_face: Any,
+    frame_processors: List[Any],
+    width: int,
+    height: int,
+    frame_size: int,
+    total_frames: int,
+    encoder: str,
+    encoder_options: List[str],
+) -> bool:
+    """Run the FFmpeg-pipe read → process → encode pipeline once."""
+
+    # --- Reader: decode source video to raw BGR24 on stdout ---
+    reader_cmd = [
+        'ffmpeg', '-hide_banner',
+        '-hwaccel', 'auto',
+        '-i', target_path,
+        '-f', 'rawvideo',
+        '-pix_fmt', 'bgr24',
+        '-v', 'error',
+        '-',
+    ]
+
+    # --- Writer: encode raw BGR24 from stdin ---
+    writer_cmd = [
+        'ffmpeg', '-hide_banner',
+        '-f', 'rawvideo',
+        '-pix_fmt', 'bgr24',
+        '-s', f'{width}x{height}',
+        '-r', str(fps),
+        '-i', '-',
+        '-c:v', encoder,
+    ]
+    writer_cmd.extend(encoder_options)
+    writer_cmd.extend([
+        '-pix_fmt', 'yuv420p',
+        '-movflags', '+faststart',
+        '-vf', 'colorspace=bt709:iall=bt601-6-625:fast=1',
+        '-v', 'error',
+        '-y', temp_output_path,
+    ])
+
+    reader = None
+    writer = None
+    try:
+        reader = subprocess.Popen(
+            reader_cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
+        )
+        writer = subprocess.Popen(
+            writer_cmd, stdin=subprocess.PIPE, stderr=subprocess.PIPE,
+        )
+    except Exception as e:
+        print(f"[DLC.CORE] Failed to start FFmpeg pipes: {e}")
+        for proc in (reader, writer):
+            if proc:
+                try:
+                    proc.kill()
+                except Exception:
+                    pass
+        return False
+
+    processed_count = 0
+    bar_fmt = ('{l_bar}{bar}| {n_fmt}/{total_fmt} '
+               '[{elapsed}<{remaining}, {rate_fmt}{postfix}]')
+
+    try:
+        with tqdm(total=total_frames, desc='Processing', unit='frame',
+                  dynamic_ncols=True, bar_format=bar_fmt) as progress:
+            progress.set_postfix({
+                'execution_providers': modules.globals.execution_providers,
+                'threads': modules.globals.execution_threads,
+                'mode': 'in-memory',
+            })
+
+            # Pipelined detection: while processing frame N (swap on
+            # ANE), start detecting the face in the next frame
+            # (detection on GPU).  They use different hardware units
+            # so the work overlaps.
+            detect_executor = ThreadPoolExecutor(max_workers=1)
+            pending_detect = None
+            use_pipeline = not modules.globals.many_faces
+
+            while True:
+                raw = reader.stdout.read(frame_size)
+                if len(raw) != frame_size:
+                    break
+
+                frame = np.frombuffer(raw, dtype=np.uint8).reshape(
+                    (height, width, 3)
+                ).copy()
+
+                # Get the detection result for THIS frame
+                if use_pipeline:
+                    if pending_detect is not None:
+                        target_face = pending_detect.result()
+                    else:
+                        target_face = get_one_face(frame)
+                    # Start detecting on THIS frame eagerly — the result
+                    # will be used for the next iteration.  At video
+                    # frame rates the face barely moves between frames.
+                    # Hand the detector its own copy: the frame processors
+                    # below mutate `frame` in place (paste-back), which
+                    # would otherwise race with detection.
+                    pending_detect = detect_executor.submit(
+                        get_one_face, frame.copy())
+                else:
+                    target_face = None
+
+                # Run frame through every active processor
+                for fp in frame_processors:
+                    try:
+                        frame = fp.process_frame(source_face, frame, target_face=target_face)
+                    except TypeError:
+                        frame = fp.process_frame(source_face, frame)
+
+                writer.stdin.write(frame.tobytes())
+                processed_count += 1
+                progress.update(1)
+
+            detect_executor.shutdown(wait=True)
+
+        # Graceful shutdown
+        writer.stdin.close()
+        writer.wait()
+        reader.wait()
+
+        if writer.returncode != 0:
+            stderr_out = writer.stderr.read().decode(errors='ignore').strip()
+            if stderr_out:
+                print(f"[DLC.CORE] FFmpeg encoder error: {stderr_out}")
+            return False
+
+        return processed_count > 0 and os.path.isfile(temp_output_path)
+
+    except BrokenPipeError:
+        print("[DLC.CORE] FFmpeg pipe broken (encoder may not be available).")
+        return False
+    except Exception as e:
+        print(f"[DLC.CORE] In-memory processing error: {e}")
+        return False
+    finally:
+        for proc in (reader, writer):
+            if proc:
+                try:
+                    proc.kill()
+                except Exception:
+                    pass

modules/processors/frame/face_enhancer.py

@@ -1,4 +1,3 @@
-# --- START OF FILE face_enhancer.py ---
 # Uses ONNX Runtime for GFPGAN face enhancement (no torch/gfpgan dependency)
 
 from typing import Any, List
@@ -81,18 +80,11 @@ def get_face_enhancer() -> onnxruntime.InferenceSession:
                 )
 
             try:
-                providers = modules.globals.execution_providers
-
-                session_options = onnxruntime.SessionOptions()
-                session_options.graph_optimization_level = (
-                    onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+                from modules.processors.frame._onnx_enhancer import (
+                    create_onnx_session,
                 )
 
-                FACE_ENHANCER = onnxruntime.InferenceSession(
-                    model_path,
-                    sess_options=session_options,
-                    providers=providers,
-                )
+                FACE_ENHANCER = create_onnx_session(model_path)
 
                 input_info = FACE_ENHANCER.get_inputs()[0]
                 output_info = FACE_ENHANCER.get_outputs()[0]
@@ -158,6 +150,18 @@ def _align_face(
     return aligned_face, affine_matrix
 
 
+_HAS_TORCH_CUDA = False
+try:
+    import torch
+    if torch.cuda.is_available():
+        _HAS_TORCH_CUDA = True
+except ImportError:
+    pass
+
+# Cache the feathered mask — it's the same for every call at a given size
+_enhancer_cache: dict = {'mask': None, 'mask_size': 0}
+
+
 def _paste_back(
     frame: Frame,
     enhanced_face: np.ndarray,
@@ -167,53 +171,77 @@ def _paste_back(
     """
     Paste an enhanced (aligned) face back onto the original frame using the
     inverse affine transform with feathered-edge blending.
+
+    Optimized: operates on a tight crop around the face bbox instead of the
+    full frame, and uses GPU for blending when available.
     """
     h, w = frame.shape[:2]
-
-    # Inverse the affine warp
     inv_matrix = cv2.invertAffineTransform(affine_matrix)
-    inv_restored = cv2.warpAffine(
-        enhanced_face,
-        inv_matrix,
-        (w, h),
-        borderMode=cv2.BORDER_CONSTANT,
-        borderValue=(0, 0, 0),
+
+    # Build or reuse cached feathered mask (uint8 — blended via cv2 SIMD ops)
+    if _enhancer_cache['mask_size'] != output_size:
+        face_mask_f = np.ones((output_size, output_size), dtype=np.float32)
+        border = max(1, int(output_size * 0.05))
+        ramp_up = np.linspace(0.0, 1.0, border, dtype=np.float32)
+        ramp_down = np.linspace(1.0, 0.0, border, dtype=np.float32)
+        face_mask_f[:border, :] *= ramp_up[:, None]
+        face_mask_f[-border:, :] *= ramp_down[:, None]
+        face_mask_f[:, :border] *= ramp_up[None, :]
+        face_mask_f[:, -border:] *= ramp_down[None, :]
+        _enhancer_cache['mask'] = (face_mask_f * 255.0).astype(np.uint8)
+        _enhancer_cache['mask_size'] = output_size
+
+    # Compute tight bbox from affine corners (avoids full-frame warpAffine scan)
+    corners = np.array([[0, 0], [output_size, 0],
+                        [output_size, output_size], [0, output_size]],
+                       dtype=np.float32)
+    transformed = (inv_matrix[:, :2] @ corners.T).T + inv_matrix[:, 2]
+    x1 = max(0, int(np.floor(transformed[:, 0].min())))
+    x2 = min(w, int(np.ceil(transformed[:, 0].max())))
+    y1 = max(0, int(np.floor(transformed[:, 1].min())))
+    y2 = min(h, int(np.ceil(transformed[:, 1].max())))
+    if x1 >= x2 or y1 >= y2:
+        return frame
+
+    # Pad a few pixels for feathering
+    pad = max(1, int(output_size * 0.05)) + 2
+    y1p, y2p = max(0, y1 - pad), min(h, y2 + pad)
+    x1p, x2p = max(0, x1 - pad), min(w, x2 + pad)
+    crop_w, crop_h = x2p - x1p, y2p - y1p
+
+    # Warp enhanced face and mask into crop space only
+    inv_crop = inv_matrix.copy()
+    inv_crop[0, 2] -= x1p
+    inv_crop[1, 2] -= y1p
+
+    inv_restored_crop = cv2.warpAffine(
+        enhanced_face, inv_crop, (crop_w, crop_h),
+        borderMode=cv2.BORDER_CONSTANT, borderValue=(0, 0, 0),
+    )
+    inv_mask_crop = cv2.warpAffine(
+        _enhancer_cache['mask'], inv_crop, (crop_w, crop_h),
+        borderMode=cv2.BORDER_CONSTANT, borderValue=0,
     )
 
-    # Build a soft feathered mask in aligned space for edge blending
-    face_mask = np.ones((output_size, output_size), dtype=np.float32)
+    target_crop = frame[y1p:y2p, x1p:x2p]
 
-    # Feather the border (5 % of the size on each edge)
-    border = max(1, int(output_size * 0.05))
-    ramp_up = np.linspace(0.0, 1.0, border, dtype=np.float32)
-    ramp_down = np.linspace(1.0, 0.0, border, dtype=np.float32)
+    if _HAS_TORCH_CUDA:
+        # Upload uint8 alpha — smaller transfer, scale on device.
+        mask_t = torch.from_numpy(inv_mask_crop).cuda().float().mul_(1.0 / 255.0).unsqueeze(2)
+        enhanced_t = torch.from_numpy(inv_restored_crop).float().cuda()
+        target_t = torch.from_numpy(target_crop).float().cuda()
+        blended = (mask_t * enhanced_t + (1.0 - mask_t) * target_t
+                   ).to(torch.uint8).cpu().numpy()
+        frame[y1p:y2p, x1p:x2p] = blended
+    else:
+        # Fused uint8 blend via cv2 SIMD — ~7× faster than the float32 round-trip.
+        alpha_3c = cv2.merge([inv_mask_crop, inv_mask_crop, inv_mask_crop])
+        inv_alpha = 255 - alpha_3c
+        a_enh = cv2.multiply(inv_restored_crop, alpha_3c, scale=1.0 / 255.0)
+        a_tgt = cv2.multiply(target_crop, inv_alpha, scale=1.0 / 255.0)
+        frame[y1p:y2p, x1p:x2p] = cv2.add(a_enh, a_tgt)
 
-    # Top / bottom rows
-    face_mask[:border, :] *= ramp_up[:, None]
-    face_mask[-border:, :] *= ramp_down[:, None]
-    # Left / right columns
-    face_mask[:, :border] *= ramp_up[None, :]
-    face_mask[:, -border:] *= ramp_down[None, :]
-
-    # Expand to 3-channel
-    face_mask_3c = np.stack([face_mask] * 3, axis=-1)
-
-    # Warp mask back to original frame space
-    inv_mask = cv2.warpAffine(
-        face_mask_3c,
-        inv_matrix,
-        (w, h),
-        borderMode=cv2.BORDER_CONSTANT,
-        borderValue=(0, 0, 0),
-    )
-    inv_mask = np.clip(inv_mask, 0.0, 1.0)
-
-    # Alpha-blend
-    result = (
-        frame.astype(np.float32) * (1.0 - inv_mask)
-        + inv_restored.astype(np.float32) * inv_mask
-    )
-    return np.clip(result, 0, 255).astype(np.uint8)
+    return frame
 
 
 def _preprocess_face(aligned_face: np.ndarray) -> np.ndarray:
@@ -221,14 +249,13 @@ def _preprocess_face(aligned_face: np.ndarray) -> np.ndarray:
     Convert an aligned BGR uint8 face image to the ONNX model input tensor.
     Format: NCHW float32, normalised to [-1, 1].
     """
-    # BGR -> RGB
-    rgb = cv2.cvtColor(aligned_face, cv2.COLOR_BGR2RGB).astype(np.float32)
-    # [0, 255] -> [0, 1] -> [-1, 1]
-    rgb = rgb / 255.0
-    rgb = (rgb - 0.5) / 0.5
-    # HWC -> CHW, add batch dim
-    chw = np.transpose(rgb, (2, 0, 1))
-    return np.expand_dims(chw, axis=0)  # shape: (1, 3, H, W)
+    # BGR -> RGB, normalize, and transpose in one pass
+    # Fused: (x / 255.0 - 0.5) / 0.5 = x / 127.5 - 1.0
+    rgb = aligned_face[:, :, ::-1]  # BGR->RGB zero-copy view
+    chw = np.transpose(rgb, (2, 0, 1)).astype(np.float32)
+    chw *= (1.0 / 127.5)
+    chw -= 1.0
+    return chw[np.newaxis, ...]  # shape: (1, 3, H, W)
 
 
 def _postprocess_face(output: np.ndarray) -> np.ndarray:
@@ -236,24 +263,42 @@ def _postprocess_face(output: np.ndarray) -> np.ndarray:
     Convert the ONNX model output tensor back to a BGR uint8 image.
     Expects input in NCHW format with values in [-1, 1].
     """
-    face = np.squeeze(output)  # remove batch dim -> (3, H, W)
-    face = np.transpose(face, (1, 2, 0))  # CHW -> HWC
-    # [-1, 1] -> [0, 1] -> [0, 255]
-    face = (face + 1.0) / 2.0
-    face = np.clip(face * 255.0, 0, 255).astype(np.uint8)
-    # RGB -> BGR
-    return cv2.cvtColor(face, cv2.COLOR_RGB2BGR)
+    # Fused: ((x + 1.0) / 2.0) * 255 = (x + 1.0) * 127.5
+    face = output[0]  # remove batch dim -> (3, H, W)
+    face = (face + 1.0) * 127.5
+    np.clip(face, 0, 255, out=face)
+    face = face.astype(np.uint8).transpose(1, 2, 0)  # CHW -> HWC
+    return face[:, :, ::-1].copy()  # RGB -> BGR
 
 
-def enhance_face(temp_frame: Frame) -> Frame:
-    """Enhances all faces in a frame using the GFPGAN ONNX model."""
+# Cache for temporal enhancement skipping in live mode.
+# GFPGAN output barely changes between consecutive frames (same face,
+# same position), so we run inference every _ENH_INTERVAL frames and
+# reuse the cached enhanced face + affine matrix in between.
+_enh_live_cache: dict = {
+    'enhanced_bgr': None,
+    'affine_matrix': None,
+    'align_size': 0,
+    'frame_count': 0,
+}
+_ENH_INTERVAL = 2  # run inference every N frames, paste cached result otherwise
+
+
+def enhance_face(temp_frame: Frame, detected_faces=None) -> Frame:
+    """Enhances all faces in a frame using the GFPGAN ONNX model.
+
+    Args:
+        detected_faces: Pre-detected face list. When provided, skips
+            the internal detection call (saves ~15-20ms per frame).
+            Also enables temporal caching — inference runs every
+            _ENH_INTERVAL frames, reusing the cached result otherwise.
+    """
     session = get_face_enhancer()
 
     # Determine model input resolution from the session metadata
     input_info = session.get_inputs()[0]
     input_name = input_info.name
     input_shape = input_info.shape  # e.g. [1, 3, 512, 512]
-    # Safely extract input size (handle dynamic / symbolic dimensions)
     try:
         align_size = int(input_shape[2])
         if align_size <= 0:
@@ -261,15 +306,25 @@ def enhance_face(temp_frame: Frame) -> Frame:
     except (ValueError, TypeError, IndexError):
         align_size = 512
 
-    # Detect faces using InsightFace (already a project dependency)
-    faces = get_many_faces(temp_frame)
+    # Use pre-detected faces if available, otherwise detect
+    faces = detected_faces if detected_faces is not None else get_many_faces(temp_frame)
     if not faces:
         return temp_frame
 
-    result_frame = temp_frame.copy()
+    # Temporal caching: only available when faces are pre-detected (live mode)
+    # AND we're in single-face mode — the cache holds exactly one enhancement,
+    # so reusing it in many_faces mode would paste the same face onto every
+    # detected target.
+    many_faces_mode = getattr(modules.globals, "many_faces", False)
+    use_cache = detected_faces is not None and not many_faces_mode
+    if use_cache:
+        _enh_live_cache['frame_count'] += 1
+        run_inference_this_frame = (_enh_live_cache['frame_count'] % _ENH_INTERVAL == 0
+                                   or _enh_live_cache['enhanced_bgr'] is None)
+    else:
+        run_inference_this_frame = True
 
     for face in faces:
-        # Need the 5-point key-points for alignment
         if not hasattr(face, "kps") or face.kps is None:
             continue
 
@@ -277,48 +332,68 @@ def enhance_face(temp_frame: Frame) -> Frame:
         if landmarks_5.shape[0] < 5:
             continue
 
-        # Align / crop the face at the model's INPUT resolution
-        aligned_face, affine_matrix = _align_face(
-            temp_frame, landmarks_5, output_size=align_size
-        )
-        if aligned_face is None or affine_matrix is None:
-            continue
-
-        try:
-            with THREAD_SEMAPHORE:
-                input_tensor = _preprocess_face(aligned_face)
-                output_tensor = session.run(None, {input_name: input_tensor})[0]
-                enhanced_bgr = _postprocess_face(output_tensor)
-
-            # The model may output at a different resolution than its input
-            # (e.g. input 512x512 → output 1024x1024).  Resize the enhanced
-            # face back to the alignment size so the inverse affine maps
-            # correctly.
-            eh, ew = enhanced_bgr.shape[:2]
-            if eh != align_size or ew != align_size:
-                enhanced_bgr = cv2.resize(
-                    enhanced_bgr,
-                    (align_size, align_size),
-                    interpolation=cv2.INTER_LANCZOS4,
-                )
-
-            # Paste enhanced face back onto the frame
-            result_frame = _paste_back(
-                result_frame, enhanced_bgr, affine_matrix, output_size=align_size
+        if run_inference_this_frame:
+            aligned_face, affine_matrix = _align_face(
+                temp_frame, landmarks_5, output_size=align_size
             )
-        except Exception as e:
-            print(f"{NAME}: Error enhancing a face: {e}")
-            continue
+            if aligned_face is None or affine_matrix is None:
+                continue
 
-    return result_frame
+            try:
+                with THREAD_SEMAPHORE:
+                    from modules.processors.frame._onnx_enhancer import (
+                        run_inference,
+                    )
+                    input_tensor = _preprocess_face(aligned_face)
+                    output_tensor = run_inference(session, input_name, input_tensor)
+                    enhanced_bgr = _postprocess_face(output_tensor)
 
+                eh, ew = enhanced_bgr.shape[:2]
+                if eh != align_size or ew != align_size:
+                    enhanced_bgr = cv2.resize(
+                        enhanced_bgr,
+                        (align_size, align_size),
+                        interpolation=cv2.INTER_LANCZOS4,
+                    )
+
+                # Cache for reuse on next frame
+                if use_cache:
+                    _enh_live_cache['enhanced_bgr'] = enhanced_bgr
+                    _enh_live_cache['affine_matrix'] = affine_matrix
+                    _enh_live_cache['align_size'] = align_size
+
+                _paste_back(
+                    temp_frame, enhanced_bgr, affine_matrix, output_size=align_size
+                )
+            except Exception as e:
+                print(f"{NAME}: Error enhancing a face: {e}")
+                continue
+        else:
+            # Reuse cached enhanced face — just paste back onto current frame
+            cached = _enh_live_cache
+            if cached['enhanced_bgr'] is not None:
+                _paste_back(
+                    temp_frame, cached['enhanced_bgr'],
+                    cached['affine_matrix'],
+                    output_size=cached['align_size'],
+                )
+        if not many_faces_mode:
+            break  # single-face live mode — only process first face
 
-def process_frame(source_face: Face | None, temp_frame: Frame) -> Frame:
-    """Processes a frame: enhances face if detected."""
-    temp_frame = enhance_face(temp_frame)
     return temp_frame
 
 
+def process_frame(source_face: Face | None, temp_frame: Frame,
+                   detected_faces=None) -> Frame:
+    """Processes a frame: enhances face if detected."""
+    return enhance_face(temp_frame, detected_faces=detected_faces)
+
+
+def process_frame_v2(temp_frame: Frame, detected_faces=None) -> Frame:
+    """Processes a frame without source face (used by live webcam preview)."""
+    return enhance_face(temp_frame, detected_faces=detected_faces)
+
+
 def process_frames(
     source_path: str | None, temp_frame_paths: List[str], progress: Any = None
 ) -> None:
@@ -367,6 +442,3 @@ def process_video(
     modules.processors.frame.core.process_video(
         source_path, temp_frame_paths, process_frames
     )
-
-
-# --- END OF FILE face_enhancer.py ---

modules/processors/frame/face_enhancer_gpen256.py

@@ -82,8 +82,11 @@ def enhance_face(temp_frame: Frame, face: Face) -> Frame:
         return temp_frame
 
 
-def process_frame(source_face: Face | None, temp_frame: Frame) -> Frame:
-    target_face = get_one_face(temp_frame)
+def process_frame(source_face: Face | None, temp_frame: Frame, detected_faces=None) -> Frame:
+    if detected_faces:
+        target_face = detected_faces[0]
+    else:
+        target_face = get_one_face(temp_frame)
     if target_face is None:
         return temp_frame
     return enhance_face(temp_frame, target_face)

modules/processors/frame/face_enhancer_gpen512.py

@@ -82,8 +82,11 @@ def enhance_face(temp_frame: Frame, face: Face) -> Frame:
         return temp_frame
 
 
-def process_frame(source_face: Face | None, temp_frame: Frame) -> Frame:
-    target_face = get_one_face(temp_frame)
+def process_frame(source_face: Face | None, temp_frame: Frame, detected_faces=None) -> Frame:
+    if detected_faces:
+        target_face = detected_faces[0]
+    else:
+        target_face = get_one_face(temp_frame)
     if target_face is None:
         return temp_frame
     return enhance_face(temp_frame, target_face)

modules/processors/frame/face_masking.py

@@ -82,8 +82,8 @@ def create_lower_mouth_mask(
 
     landmarks = face.landmark_2d_106
     if landmarks is not None:
-        # Use outer mouth landmarks (52-63) to capture the lips only
-        lower_lip_order = list(range(52, 64))
+        # Use outer mouth landmarks (52-71) to capture the full mouth area
+        lower_lip_order = list(range(52, 72))
         
         if max(lower_lip_order) >= landmarks.shape[0]:
             return mask, mouth_cutout, mouth_box, lower_lip_polygon
@@ -94,13 +94,16 @@ def create_lower_mouth_mask(
         center = np.mean(lower_lip_landmarks, axis=0)
 
         # Expand the landmarks outward using the mouth_mask_size
-        # Use a more conservative expansion to avoid affecting face shape
-        expansion_factor = (
-            1 + modules.globals.mask_down_size * modules.globals.mouth_mask_size
-        )
-        expanded_landmarks = (lower_lip_landmarks - center) * expansion_factor + center
+        mouth_mask_size = getattr(modules.globals, "mouth_mask_size", 0.0) # 0-100 slider
+        expansion_factor = 1 + (mouth_mask_size / 100.0) * 2.5
 
-        # Removed specific top/chin extensions to preserve face shape
+        # Expand with extra downward bias toward chin
+        offsets = lower_lip_landmarks - center
+        chin_bias = 1 + (mouth_mask_size / 100.0) * 1.5
+        scale_y = np.where(offsets[:, 1] > 0, expansion_factor * chin_bias, expansion_factor)
+        expanded_landmarks = lower_lip_landmarks.copy()
+        expanded_landmarks[:, 0] = center[0] + offsets[:, 0] * expansion_factor
+        expanded_landmarks[:, 1] = center[1] + offsets[:, 1] * scale_y
 
         # Convert back to integer coordinates
         expanded_landmarks = expanded_landmarks.astype(np.int32)

modules/processors/frame/face_swapper.py

@@ -1,6 +1,7 @@
-from typing import Any, List, Optional
+from typing import Any, List, Optional, Tuple
 import cv2
 import insightface
+import logging
 import threading
 import numpy as np
 import platform
@@ -28,6 +29,149 @@ NAME = "DLC.FACE-SWAPPER"
 PREVIOUS_FRAME_RESULT = None # Stores the final processed frame from the previous step
 # --- END: Added for Interpolation ---
 
+# --- Poisson blend (ported from deep-live-cam-gumroad-edition) ---
+# Root-cause fix for the "wobble": the blend mask is NOT built from the
+# independently-detected 106-pt landmarks (they jitter sub-pixel every frame
+# and seamlessClone is hyper-sensitive to its mask boundary). Instead it is
+# derived from the swap's OWN affine transform (M) + the swapped pixels
+# (bgr_fake), so the mask is locked exactly to where the swapped face was
+# placed — no independent jitter source, no EMA, no lag. The mask is cached
+# when the face is nearly still so an identical array is reused (zero wobble).
+_ELLIPTICAL_MASK_CACHE: dict = {}
+_poisson_cached_mask: Optional[np.ndarray] = None
+_poisson_cached_key: Optional[tuple] = None
+
+
+def _create_elliptical_mask(size: Tuple[int, int]) -> np.ndarray:
+    """Fixed, heavily-blurred elliptical mask in aligned-face space.
+
+    Geometry-based (not content-adaptive) and cached by size — identical
+    every frame for the same model input size, so it contributes no jitter.
+    """
+    global _ELLIPTICAL_MASK_CACHE
+    if size in _ELLIPTICAL_MASK_CACHE:
+        return _ELLIPTICAL_MASK_CACHE[size]
+    h, w = size
+    center = (w // 2, h // 2)
+    axes = (int(w * 0.44), int(h * 0.44))
+    mask = np.zeros((h, w), dtype=np.float32)
+    cv2.ellipse(mask, center, axes, 0, 0, 360, 1, -1)
+    if h * w < 65536:
+        mask = cv2.GaussianBlur(mask, (31, 31), 12)
+    else:
+        mask = gpu_gaussian_blur(mask, (31, 31), 12)
+    _ELLIPTICAL_MASK_CACHE[size] = mask
+    return mask
+
+
+def _apply_poisson_blend(swapped_frame: Frame, original_frame: Frame,
+                         target_face: Face, affine_matrix: np.ndarray = None,
+                         bgr_fake: np.ndarray = None) -> Frame:
+    """Poisson-blend the swapped face onto the original frame.
+
+    Preferred path derives the blend mask from the swap's inverse affine so
+    it tracks the swapped face exactly per-frame (no landmark jitter, no
+    smoothing). Falls back to a cached bbox-ellipse if the affine is absent.
+    Writes only the blended ellipse back so other faces are preserved.
+    """
+    global _poisson_cached_mask, _poisson_cached_key
+    try:
+        # ---- Preferred: blend ONLY the genuinely-swapped region ----
+        # Use the exact paste-back mask (warped elliptical mask), eroded so
+        # the Poisson seam sits on solidly-swapped pixels only.
+        if affine_matrix is not None and bgr_fake is not None:
+            try:
+                h, w = swapped_frame.shape[:2]
+                fh, fw = bgr_fake.shape[:2]
+                inv = cv2.invertAffineTransform(affine_matrix)
+                corners = np.array([[0, 0, 1], [fw, 0, 1], [fw, fh, 1], [0, fh, 1]],
+                                   dtype=np.float32)
+                t = corners @ inv.T
+                px1 = max(0, int(np.floor(t[:, 0].min())))
+                py1 = max(0, int(np.floor(t[:, 1].min())))
+                px2 = min(w, int(np.ceil(t[:, 0].max())))
+                py2 = min(h, int(np.ceil(t[:, 1].max())))
+                rw, rh = px2 - px1, py2 - py1
+                if rw > 8 and rh > 8:
+                    roi_aff = inv.copy()
+                    roi_aff[0, 2] -= px1
+                    roi_aff[1, 2] -= py1
+                    fm = _create_elliptical_mask((fh, fw))
+                    mroi = cv2.warpAffine(fm, roi_aff, (rw, rh),
+                                          flags=cv2.INTER_LINEAR,
+                                          borderMode=cv2.BORDER_CONSTANT, borderValue=0)
+                    bin_roi = np.where(mroi > 0.5, np.uint8(255), np.uint8(0))
+                    k = max(3, (min(rw, rh) // 20) | 1)
+                    bin_roi = cv2.erode(bin_roi,
+                                        cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (k, k)))
+                    bx, by, bw, bh = cv2.boundingRect(bin_roi)
+                    if bw > 0 and bh > 0:
+                        mx1, my1 = px1 + bx, py1 + by
+                        mx2, my2 = mx1 + bw - 1, my1 + bh - 1
+                        # seamlessClone needs the cloned region off the border
+                        if mx1 > 0 and my1 > 0 and mx2 < w - 1 and my2 < h - 1:
+                            mask = np.zeros((h, w), dtype=np.uint8)
+                            mask[py1:py2, px1:px2] = bin_roi
+                            center = (mx1 + bw // 2, my1 + bh // 2)
+                            blended = cv2.seamlessClone(swapped_frame, original_frame,
+                                                        mask, center, cv2.NORMAL_CLONE)
+                            np.copyto(swapped_frame[my1:my2 + 1, mx1:mx2 + 1],
+                                      blended[my1:my2 + 1, mx1:mx2 + 1],
+                                      where=mask[my1:my2 + 1, mx1:mx2 + 1, None].astype(bool))
+                            return swapped_frame
+            except Exception:
+                pass  # fall through to the robust bbox-ellipse path below
+        # ---- Fallback: bbox-ellipse (defensive, cached when still) ----
+        if not hasattr(target_face, 'bbox') or target_face.bbox is None:
+            return swapped_frame
+        x1, y1, x2, y2 = target_face.bbox.astype(int)
+        h, w = swapped_frame.shape[:2]
+        x1, y1 = (max(0, x1), max(0, y1))
+        x2, y2 = (min(w, x2), min(h, y2))
+        if x2 <= x1 or y2 <= y1 or x2 - x1 <= 10 or (y2 - y1 <= 10):
+            return swapped_frame
+        padding = int(min(x2 - x1, y2 - y1) * 0.1)
+        x1_p = max(0, x1 - padding)
+        y1_p = max(0, y1 - padding)
+        x2_p = min(w, x2 + padding)
+        y2_p = min(h, y2 + padding)
+        center_x = int(round((x1 + x2) / 2.0))
+        center_y = int(round((y1 + y2) / 2.0))
+        radius_x = max(1, int(round((x2_p - x1_p) / 2.0)))
+        radius_y = max(1, int(round((y2_p - y1_p) / 2.0)))
+        if not (0 <= center_x < w and 0 <= center_y < h):
+            return swapped_frame
+        center = (center_x, center_y)
+        if center_x - radius_x < 0 or center_x + radius_x >= w or center_y - radius_y < 0 or (center_y + radius_y >= h):
+            return swapped_frame
+        # Reuse cached mask when center/radius unchanged frame-to-frame
+        # (face nearly still) — saves the np.zeros + cv2.ellipse, and the
+        # identical array means literally zero wobble while still.
+        mask_key = (center_x, center_y, radius_x, radius_y, h, w)
+        if _poisson_cached_key == mask_key and _poisson_cached_mask is not None:
+            mask = _poisson_cached_mask
+        else:
+            mask = np.zeros((h, w), dtype=np.uint8)
+            cv2.ellipse(mask, center, (radius_x, radius_y), 0, 0, 360, 255, -1)
+            if np.sum(mask) == 0:
+                return swapped_frame
+            _poisson_cached_mask = mask
+            _poisson_cached_key = mask_key
+        blended = cv2.seamlessClone(swapped_frame, original_frame, mask, center, cv2.NORMAL_CLONE)
+        # Composite ONLY this face's ellipse back (ROI-bounded) so previously
+        # blended faces in multi-face mode are preserved.
+        rx0 = max(0, center_x - radius_x)
+        rx1 = min(w, center_x + radius_x + 1)
+        ry0 = max(0, center_y - radius_y)
+        ry1 = min(h, center_y + radius_y + 1)
+        roi_mask = mask[ry0:ry1, rx0:rx1]
+        np.copyto(swapped_frame[ry0:ry1, rx0:rx1],
+                  blended[ry0:ry1, rx0:rx1],
+                  where=roi_mask[:, :, None].astype(bool))
+        return swapped_frame
+    except Exception:
+        return swapped_frame
+
 # --- START: Mac M1-M5 Optimizations ---
 IS_APPLE_SILICON = platform.system() == 'Darwin' and platform.machine() == 'arm64'
 FRAME_CACHE = deque(maxlen=3)  # Cache for frame reuse
@@ -54,21 +198,22 @@ def pre_check() -> bool:
         logging.error(f"Failed to create directory {download_directory_path} due to permission error: {e}")
         return False
     
-    # Use the direct download URL from Hugging Face
+    # Use the direct download URL from Hugging Face (FP32 model for broad GPU compatibility)
     conditional_download(
         download_directory_path,
         [
-            "https://huggingface.co/hacksider/deep-live-cam/resolve/main/inswapper_128_fp16.onnx"
+            "https://huggingface.co/hacksider/deep-live-cam/resolve/main/inswapper_128.onnx"
         ],
     )
     return True
 
 
 def pre_start() -> bool:
-    # Simplified pre_start, assuming checks happen before calling process functions
-    model_path = os.path.join(models_dir, "inswapper_128_fp16.onnx")
-    if not os.path.exists(model_path):
-        update_status(f"Model not found: {model_path}. Please download it.", NAME)
+    # Check for either model variant
+    fp16_path = os.path.join(models_dir, "inswapper_128_fp16.onnx")
+    fp32_path = os.path.join(models_dir, "inswapper_128.onnx")
+    if not os.path.exists(fp16_path) and not os.path.exists(fp32_path):
+        update_status(f"Model not found in {models_dir}. Please download inswapper_128.onnx.", NAME)
         return False
 
     # Try to get the face swapper to ensure it loads correctly
@@ -76,7 +221,6 @@ def pre_start() -> bool:
         # Error message already printed within get_face_swapper
         return False
 
-    # Add other essential checks if needed, e.g., target/source path validity
     return True
 
 
@@ -85,13 +229,28 @@ def get_face_swapper() -> Any:
 
     with THREAD_LOCK:
         if FACE_SWAPPER is None:
-            model_name = "inswapper_128.onnx"
-            if "CUDAExecutionProvider" in modules.globals.execution_providers:
-                model_name = "inswapper_128_fp16.onnx"
-            model_path = os.path.join(models_dir, model_name)
+            # Prefer FP16 on GPUs with Tensor Cores (Turing+) — half the
+            # memory bandwidth, faster inference.  Fall back to FP32 for
+            # older GPUs (e.g. GTX 16xx) where FP16 can produce NaN.
+            fp32_path = os.path.join(models_dir, "inswapper_128.onnx")
+            fp16_path = os.path.join(models_dir, "inswapper_128_fp16.onnx")
+            use_fp16 = _HAS_TORCH_CUDA and os.path.exists(fp16_path)
+            if use_fp16:
+                model_path = fp16_path
+            elif os.path.exists(fp32_path):
+                model_path = fp32_path
+            else:
+                update_status(f"No inswapper model found in {models_dir}.", NAME)
+                return None
+            # On Apple Silicon, rewrite Pad(reflect) → Slice+Concat so
+            # CoreML can run the entire model in a single partition on
+            # the Neural Engine instead of bouncing between CPU and ANE.
+            if IS_APPLE_SILICON:
+                from modules.onnx_optimize import optimize_for_coreml
+                model_path = optimize_for_coreml(model_path)
+
             update_status(f"Loading face swapper model from: {model_path}", NAME)
             try:
-                # Optimized provider configuration for Apple Silicon
                 providers_config = []
                 for p in modules.globals.execution_providers:
                     if p == "CoreMLExecutionProvider" and IS_APPLE_SILICON:
@@ -104,17 +263,25 @@ def get_face_swapper() -> Any:
                                 "SpecializationStrategy": "FastPrediction",
                                 "AllowLowPrecisionAccumulationOnGPU": 1,
                                 "EnableOnSubgraphs": 1,
-                                "RequireStaticShapes": 0,
-                                "MaximumCacheSize": 1024 * 1024 * 512,  # 512MB cache
                             }
                         ))
+                    elif p == "CUDAExecutionProvider":
+                        # Use bare provider — ONNX Runtime defaults are
+                        # fastest on modern GPUs (Blackwell/sm_120).
+                        providers_config.append(p)
                     else:
                         providers_config.append(p)
-                
                 FACE_SWAPPER = insightface.model_zoo.get_model(
                     model_path,
                     providers=providers_config,
                 )
+                # Set up CUDA graph session for faster inference
+                if _HAS_TORCH_CUDA and any(
+                    p == "CUDAExecutionProvider" or
+                    (isinstance(p, tuple) and p[0] == "CUDAExecutionProvider")
+                    for p in providers_config
+                ):
+                    _init_cuda_graph_session(model_path, FACE_SWAPPER)
                 update_status("Face swapper model loaded successfully.", NAME)
             except Exception as e:
                 update_status(f"Error loading face swapper model: {e}", NAME)
@@ -123,6 +290,214 @@ def get_face_swapper() -> Any:
     return FACE_SWAPPER
 
 
+_HAS_TORCH_CUDA = False
+try:
+    import torch
+    if torch.cuda.is_available():
+        _HAS_TORCH_CUDA = True
+except ImportError:
+    pass
+
+# Cache for paste-back
+_paste_cache = {
+    'soft_alpha': None,  # feathered alpha mask in aligned-face space
+    'alpha_size': 0,
+}
+
+
+def _get_soft_alpha(size: int) -> np.ndarray:
+    """Feathered alpha template in aligned-face space, cached.
+
+    The legacy paste-back eroded and Gaussian-blurred the warped mask in
+    output coordinates with kernels scaled to the output face size, which
+    made the per-frame cost quartic in face linear size. Doing the same
+    erode+blur once in aligned space and then warping the *soft* mask
+    per-frame gives a visually equivalent feather at O(crop_area) cost —
+    the feather radius scales naturally with the affine transform.
+    """
+    if _paste_cache['alpha_size'] != size:
+        # Elliptical (not square) template — matches the gumroad edition's
+        # _create_elliptical_mask. A full/eroded square leaves the aligned
+        # crop's corners near-opaque, so the swapped square's straight edges
+        # show as a visible box on the face. An ellipse (axes 0.44*size) zeroes
+        # the corners and the heavy blur feathers smoothly into the original.
+        center = (size // 2, size // 2)
+        axes = (int(size * 0.44), int(size * 0.44))
+        mask = np.zeros((size, size), dtype=np.uint8)
+        cv2.ellipse(mask, center, axes, 0, 0, 360, 255, -1)
+        mask = cv2.GaussianBlur(mask, (31, 31), 12)
+        _paste_cache['soft_alpha'] = mask  # uint8 [0, 255] — blended via cv2 SIMD ops
+        _paste_cache['alpha_size'] = size
+    return _paste_cache['soft_alpha']
+
+# CUDA graph swap session cache
+_cuda_graph_session = {
+    'session': None,
+    'io_binding': None,
+    'ort_input': None,
+    'ort_latent': None,
+    'recorded': False,
+}
+# Serializes CUDA-graph replay. The io_binding + ort_input/ort_latent are
+# shared across threads and run_with_iobinding mutates GPU-side buffers;
+# concurrent calls would produce wrong output.
+_cuda_graph_lock = threading.Lock()
+
+
+class _CudaGraphSessionAdapter:
+    """Drop-in wrapper around an ONNX Runtime session.
+
+    Routes ``.run()`` through CUDA graph replay when a recorded graph is
+    available, and transparently proxies every other attribute to the
+    underlying session so insightface's INSwapper sees an unchanged API.
+    """
+
+    def __init__(self, underlying):
+        # Use object.__setattr__ to bypass our own __setattr__.
+        object.__setattr__(self, "_underlying", underlying)
+
+    def run(self, output_names, input_dict, **kwargs):
+        if _cuda_graph_session['recorded']:
+            try:
+                keys = list(input_dict.keys())
+                blob = input_dict[keys[0]]
+                latent = input_dict[keys[1]]
+                return [_cuda_graph_swap_inference(blob, latent)]
+            except Exception:
+                pass
+        return self._underlying.run(output_names, input_dict, **kwargs)
+
+    def __getattr__(self, name):
+        return getattr(self._underlying, name)
+
+    def __setattr__(self, name, value):
+        setattr(self._underlying, name, value)
+
+
+def _init_cuda_graph_session(model_path: str, swapper):
+    """Create a CUDA-graph-enabled ONNX session for the swap model.
+
+    CUDA graphs record the GPU kernel launch sequence once, then replay it
+    with near-zero CPU overhead on subsequent runs.  Requires static input
+    shapes (inswapper is always 1x3x128x128 + 1x512).
+    """
+    import onnxruntime as ort
+    try:
+        providers = [('CUDAExecutionProvider', {'enable_cuda_graph': '1'})]
+        sess = ort.InferenceSession(model_path, providers=providers)
+
+        # Pre-allocate GPU buffers with correct shapes
+        inp_shape = (1, 3, swapper.input_size[1], swapper.input_size[0])
+        latent_shape = (1, 512)
+        dummy_inp = np.zeros(inp_shape, dtype=np.float32)
+        dummy_lat = np.zeros(latent_shape, dtype=np.float32)
+
+        ort_input = ort.OrtValue.ortvalue_from_numpy(dummy_inp, 'cuda', 0)
+        ort_latent = ort.OrtValue.ortvalue_from_numpy(dummy_lat, 'cuda', 0)
+
+        io = sess.io_binding()
+        io.bind_ortvalue_input(swapper.input_names[0], ort_input)
+        io.bind_ortvalue_input(swapper.input_names[1], ort_latent)
+        io.bind_output(swapper.output_names[0], 'cuda', 0)
+
+        # First run records the CUDA graph
+        sess.run_with_iobinding(io)
+
+        _cuda_graph_session['session'] = sess
+        _cuda_graph_session['io_binding'] = io
+        _cuda_graph_session['ort_input'] = ort_input
+        _cuda_graph_session['ort_latent'] = ort_latent
+        _cuda_graph_session['recorded'] = True
+
+        # Wrap swapper.session in an adapter instead of rebinding
+        # session.run. insightface's INSwapper.get() reads .run via the
+        # session attribute, so either works; the adapter survives any
+        # later attribute reads on the session and keeps the original
+        # session object untouched.
+        if not isinstance(swapper.session, _CudaGraphSessionAdapter):
+            swapper.session = _CudaGraphSessionAdapter(swapper.session)
+
+        import sys
+        print(f"[{NAME}] CUDA graph session initialized (swap model)")
+        sys.stdout.flush()
+    except Exception as e:
+        print(f"[{NAME}] CUDA graph init failed, using standard session: {e}")
+        _cuda_graph_session['recorded'] = False
+
+
+def _cuda_graph_swap_inference(blob: np.ndarray, latent: np.ndarray) -> np.ndarray:
+    """Run swap model via CUDA graph replay — minimal CPU overhead."""
+    cg = _cuda_graph_session
+    with _cuda_graph_lock:
+        cg['ort_input'].update_inplace(blob)
+        cg['ort_latent'].update_inplace(latent)
+        cg['session'].run_with_iobinding(cg['io_binding'])
+        return cg['io_binding'].get_outputs()[0].numpy()
+
+
+def _fast_paste_back(target_img: Frame, bgr_fake: np.ndarray, aimg: np.ndarray, M: np.ndarray) -> Frame:
+    """Paste bgr_fake back onto target_img via the inverse affine of M.
+
+    Restricts work to the face bbox in output coordinates and warps a
+    precomputed feathered alpha template per-frame instead of running a
+    size-scaled erode+blur on the warped mask. Cost is O(crop_area) regardless
+    of how much of the frame the face occupies.
+    """
+    h, w = target_img.shape[:2]
+    face_h, face_w = aimg.shape[:2]
+    # inswapper's aligned-face space is square (128x128). _get_soft_alpha
+    # caches a single NxN template keyed by N, so fail loudly if that ever
+    # stops being true rather than silently mis-warping the alpha mask.
+    assert face_h == face_w, f"Expected square aligned face, got {face_h}x{face_w}"
+    IM = cv2.invertAffineTransform(M)
+
+    # Bbox in output coords from the affine corners of the aligned-face square.
+    corners = np.array(
+        [[0, 0], [face_w, 0], [face_w, face_h], [0, face_h]], dtype=np.float32
+    )
+    transformed = (IM[:, :2] @ corners.T).T + IM[:, 2]
+    x1 = int(np.floor(transformed[:, 0].min()))
+    x2 = int(np.ceil(transformed[:, 0].max()))
+    y1 = int(np.floor(transformed[:, 1].min()))
+    y2 = int(np.ceil(transformed[:, 1].max()))
+    if x1 >= x2 or y1 >= y2:
+        return target_img
+
+    # Small interpolation margin only — the feather is baked into the template.
+    pad = 2
+    y1p, y2p = max(0, y1 - pad), min(h, y2 + pad + 1)
+    x1p, x2p = max(0, x1 - pad), min(w, x2 + pad + 1)
+
+    IM_crop = IM.copy()
+    IM_crop[0, 2] -= x1p
+    IM_crop[1, 2] -= y1p
+    crop_w, crop_h = x2p - x1p, y2p - y1p
+
+    soft_alpha = _get_soft_alpha(face_h)
+    bgr_fake_crop = cv2.warpAffine(bgr_fake, IM_crop, (crop_w, crop_h), borderMode=cv2.BORDER_REPLICATE)
+    alpha_crop = cv2.warpAffine(soft_alpha, IM_crop, (crop_w, crop_h), borderValue=0)
+
+    target_crop = target_img[y1p:y2p, x1p:x2p]
+
+    if _HAS_TORCH_CUDA:
+        # Scale alpha to [0, 1] on device — cheaper to upload uint8 than float.
+        mask_t = torch.from_numpy(alpha_crop).cuda().float().mul_(1.0 / 255.0).unsqueeze(2)
+        fake_t = torch.from_numpy(bgr_fake_crop).float().cuda()
+        tgt_t = torch.from_numpy(target_crop).float().cuda()
+        blended = (mask_t * fake_t + (1.0 - mask_t) * tgt_t).to(torch.uint8).cpu().numpy()
+        target_img[y1p:y2p, x1p:x2p] = blended
+    else:
+        # Fused uint8 blend via cv2 SIMD — no float32 round-trip.
+        # Measured ~7-8× faster than the old numpy float32 path on a 1000×1000 crop.
+        alpha_3c = cv2.merge([alpha_crop, alpha_crop, alpha_crop])
+        inv_alpha = 255 - alpha_3c
+        a_fake = cv2.multiply(bgr_fake_crop, alpha_3c, scale=1.0 / 255.0)
+        a_tgt = cv2.multiply(target_crop, inv_alpha, scale=1.0 / 255.0)
+        target_img[y1p:y2p, x1p:x2p] = cv2.add(a_fake, a_tgt)
+
+    return target_img
+
+
 def swap_face(source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
     """Optimized face swapping with better memory management and performance."""
     face_swapper = get_face_swapper()
@@ -136,112 +511,92 @@ def swap_face(source_face: Face, target_face: Face, temp_frame: Frame) -> Frame:
     if not hasattr(source_face, 'normed_embedding') or source_face.normed_embedding is None:
         return temp_frame
 
-    # Store a copy of the original frame before swapping for opacity blending
+    # _fast_paste_back writes in-place on the GPU path.  Only copy when
+    # mouth_mask or opacity < 1 need an unmodified original.
     opacity = getattr(modules.globals, "opacity", 1.0)
     opacity = max(0.0, min(1.0, opacity))
-    original_frame = temp_frame if opacity >= 1.0 else temp_frame.copy()
+    mouth_mask_enabled = getattr(modules.globals, "mouth_mask", False)
+    poisson_blend_enabled = getattr(modules.globals, "poisson_blend", False)
+    # Poisson blend's seamlessClone needs the genuine pre-swap frame as its
+    # destination. Without this, original_frame aliases temp_frame, which
+    # _fast_paste_back mutates in place — so seamlessClone would blend the
+    # swapped face onto the already-swapped frame (no visible effect).
+    needs_original = opacity < 1.0 or mouth_mask_enabled or poisson_blend_enabled
+    if needs_original:
+        original_frame = temp_frame.copy()
+    else:
+        original_frame = temp_frame
 
-    # Pre-swap Input Check with optimization
     if temp_frame.dtype != np.uint8:
         temp_frame = np.clip(temp_frame, 0, 255).astype(np.uint8)
 
-    # Apply the face swap with optimized memory handling
     try:
-        # Ensure contiguous memory layout for better performance on all platforms
         if not temp_frame.flags['C_CONTIGUOUS']:
             temp_frame = np.ascontiguousarray(temp_frame)
-        
-        swapped_frame_raw = face_swapper.get(
-            temp_frame, target_face, source_face, paste_back=True
-        )
 
-        # --- START: CRITICAL FIX FOR ORT 1.17 ---
-        # Check the output type and range from the model
-        if swapped_frame_raw is None:
-             # print("Warning: face_swapper.get returned None.") # Debug
-             return original_frame # Return original if swap somehow failed internally
+        # Use paste_back=False and our optimized paste-back
+        if any("DmlExecutionProvider" in p for p in modules.globals.execution_providers):
+            with modules.globals.dml_lock:
+                bgr_fake, M = face_swapper.get(
+                    temp_frame, target_face, source_face, paste_back=False
+                )
+        else:
+            bgr_fake, M = face_swapper.get(
+                temp_frame, target_face, source_face, paste_back=False
+            )
 
-        # Ensure the output is a numpy array
-        if not isinstance(swapped_frame_raw, np.ndarray):
-            # print(f"Warning: face_swapper.get returned type {type(swapped_frame_raw)}, expected numpy array.") # Debug
+        if bgr_fake is None:
             return original_frame
 
-        # Ensure the output has the correct shape (like the input frame)
-        if swapped_frame_raw.shape != temp_frame.shape:
-             # print(f"Warning: Swapped frame shape {swapped_frame_raw.shape} differs from input {temp_frame.shape}.") # Debug
-             # Attempt resize (might distort if aspect ratio changed, but better than crashing)
-             try:
-                 swapped_frame_raw = gpu_resize(swapped_frame_raw, (temp_frame.shape[1], temp_frame.shape[0]))
-             except Exception as resize_e:
-                 # print(f"Error resizing swapped frame: {resize_e}") # Debug
-                 return original_frame
+        if not isinstance(bgr_fake, np.ndarray):
+            return original_frame
 
-        # Explicitly clip values to 0-255 and convert to uint8
-        # This handles cases where the model might output floats or values outside the valid range
-        swapped_frame = np.clip(swapped_frame_raw, 0, 255).astype(np.uint8)
-        # --- END: CRITICAL FIX FOR ORT 1.17 ---
+        # Pass a dummy aimg with correct shape — _fast_paste_back only uses aimg.shape
+        # to create the white mask. Avoids redundant norm_crop2 (~0.6ms).
+        _face_size = face_swapper.input_size[0]
+        _aimg_dummy = np.empty((_face_size, _face_size, 3), dtype=np.uint8)
+
+        swapped_frame = _fast_paste_back(temp_frame, bgr_fake, _aimg_dummy, M)
 
     except Exception as e:
-        print(f"Error during face swap using face_swapper.get: {e}") # More specific error
-        # import traceback
-        # traceback.print_exc() # Print full traceback for debugging
-        return original_frame # Return original if swap fails
+        print(f"Error during face swap: {e}")
+        return original_frame
 
     # --- Post-swap Processing (Masking, Opacity, etc.) ---
     # Now, work with the guaranteed uint8 'swapped_frame'
 
-    if getattr(modules.globals, "mouth_mask", False): # Check if mouth_mask is enabled
+    if mouth_mask_enabled: # Check if mouth_mask is enabled
         # Create a mask for the target face
-        face_mask = create_face_mask(target_face, temp_frame) # Use temp_frame (original shape) for mask creation geometry
+        face_mask = create_face_mask(target_face, original_frame) # Use original_frame for mask creation geometry
 
-        # Create the mouth mask using original geometry
+        # Create the mouth mask using the ORIGINAL frame (before swap) for cutout
         mouth_mask, mouth_cutout, mouth_box, lower_lip_polygon = (
-            create_lower_mouth_mask(target_face, temp_frame) # Use temp_frame (original) for cutout
+            create_lower_mouth_mask(target_face, original_frame) # Use original_frame for real mouth cutout
         )
 
         # Apply the mouth area only if mouth_cutout exists
-        if mouth_cutout is not None and mouth_box != (0,0,0,0): # Add check for valid box
-             # Apply mouth area (from original) onto the 'swapped_frame'
+        if mouth_cutout is not None and mouth_box != (0,0,0,0):
+            # Apply mouth area (from original) onto the 'swapped_frame'
             swapped_frame = apply_mouth_area(
                 swapped_frame, mouth_cutout, mouth_box, face_mask, lower_lip_polygon
             )
 
+            # Draw bounding box only while slider is being dragged
             if getattr(modules.globals, "show_mouth_mask_box", False):
-                        mouth_mask_data = (mouth_mask, mouth_cutout, mouth_box, lower_lip_polygon)
-                        # Draw visualization on the swapped_frame *before* opacity blending
-                        swapped_frame = draw_mouth_mask_visualization(
-                            swapped_frame, target_face, mouth_mask_data
-                        )
+                mouth_mask_data = (mouth_mask, mouth_cutout, mouth_box, lower_lip_polygon)
+                swapped_frame = draw_mouth_mask_visualization(
+                    swapped_frame, target_face, mouth_mask_data
+                )
         
     # --- Poisson Blending ---
+    # Mask derived from the swap's own affine (M) + swapped pixels (bgr_fake),
+    # so it tracks the swapped face exactly per-frame — no landmark jitter,
+    # no EMA, no lag. See _apply_poisson_blend.
     if getattr(modules.globals, "poisson_blend", False):
-        face_mask = create_face_mask(target_face, temp_frame)
-        if face_mask is not None:
-            # Find bounding box of the mask
-            y_indices, x_indices = np.where(face_mask > 0)
-            if len(x_indices) > 0 and len(y_indices) > 0:
-                x_min, x_max = np.min(x_indices), np.max(x_indices)
-                y_min, y_max = np.min(y_indices), np.max(y_indices)
+        swapped_frame = _apply_poisson_blend(
+            swapped_frame, original_frame, target_face, M, bgr_fake
+        )
 
-                # Calculate center
-                center = (int((x_min + x_max) / 2), int((y_min + y_max) / 2))
-
-                # Crop src and mask
-                src_crop = swapped_frame[y_min : y_max + 1, x_min : x_max + 1]
-                mask_crop = face_mask[y_min : y_max + 1, x_min : x_max + 1]
-
-                try:
-                    # Use original_frame as destination to blend the swapped face onto it
-                    swapped_frame = cv2.seamlessClone(
-                        src_crop,
-                        original_frame,
-                        mask_crop,
-                        center,
-                        cv2.NORMAL_CLONE,
-                    )
-                except Exception as e:
-                    print(f"Poisson blending failed: {e}")
-        
     # Apply opacity blend between the original frame and the swapped frame
     if opacity >= 1.0:
         return swapped_frame.astype(np.uint8)
@@ -292,6 +647,14 @@ def apply_post_processing(current_frame: Frame, swapped_face_bboxes: List[np.nda
     """Applies sharpening and interpolation with Apple Silicon optimizations."""
     global PREVIOUS_FRAME_RESULT
 
+    sharpness_value = getattr(modules.globals, "sharpness", 0.0)
+    enable_interpolation = getattr(modules.globals, "enable_interpolation", False)
+
+    # Skip copy when no post-processing is active
+    if sharpness_value <= 0.0 and not enable_interpolation:
+        PREVIOUS_FRAME_RESULT = None
+        return current_frame
+
     processed_frame = current_frame.copy()
 
     # 1. Apply Sharpening (if enabled) with optimized kernel for Apple Silicon
@@ -367,42 +730,40 @@ def apply_post_processing(current_frame: Frame, swapped_face_bboxes: List[np.nda
 # --- END: Helper function for interpolation and sharpening ---
 
 
-def process_frame(source_face: Face, temp_frame: Frame) -> Frame:
-    """
-    DEPRECATED / SIMPLER VERSION - Processes a single frame using one source face.
-    Consider using process_frame_v2 for more complex scenarios.
+def process_frame(source_face: Face, temp_frame: Frame, target_face: Face = None) -> Frame:
+    """Process a single frame, swapping source_face onto detected target(s).
+
+    Args:
+        target_face: Pre-detected target face. When provided, skips the
+            internal face detection call (saves ~30-40ms per frame).
+            Ignored when many_faces mode is active.
     """
     if getattr(modules.globals, "opacity", 1.0) == 0:
-        # If opacity is 0, no swap happens, so no post-processing needed.
-        # Also reset interpolation state if it was active.
         global PREVIOUS_FRAME_RESULT
         PREVIOUS_FRAME_RESULT = None
         return temp_frame
 
-    # Color correction removed from here (better applied before swap if needed)
-
-    processed_frame = temp_frame # Start with the input frame
-    swapped_face_bboxes = [] # Keep track of where swaps happened
+    processed_frame = temp_frame
+    swapped_face_bboxes = []
 
     if modules.globals.many_faces:
         many_faces = get_many_faces(processed_frame)
         if many_faces:
-            current_swap_target = processed_frame.copy() # Apply swaps sequentially on a copy
-            for target_face in many_faces:
-                current_swap_target = swap_face(source_face, target_face, current_swap_target)
-                if target_face is not None and hasattr(target_face, "bbox") and target_face.bbox is not None:
-                    swapped_face_bboxes.append(target_face.bbox.astype(int))
-            processed_frame = current_swap_target # Assign the final result after all swaps
+            current_swap_target = processed_frame.copy()
+            for face in many_faces:
+                current_swap_target = swap_face(source_face, face, current_swap_target)
+                if face is not None and hasattr(face, "bbox") and face.bbox is not None:
+                    swapped_face_bboxes.append(face.bbox.astype(int))
+            processed_frame = current_swap_target
     else:
-        target_face = get_one_face(processed_frame)
+        if target_face is None:
+            target_face = get_one_face(processed_frame)
         if target_face:
             processed_frame = swap_face(source_face, target_face, processed_frame)
-            if target_face is not None and hasattr(target_face, "bbox") and target_face.bbox is not None:
-                    swapped_face_bboxes.append(target_face.bbox.astype(int))
+            if hasattr(target_face, "bbox") and target_face.bbox is not None:
+                swapped_face_bboxes.append(target_face.bbox.astype(int))
 
-    # Apply sharpening and interpolation
     final_frame = apply_post_processing(processed_frame, swapped_face_bboxes)
-
     return final_frame
 
 
@@ -750,13 +1111,20 @@ def create_lower_mouth_mask(
         return mask, mouth_cutout, mouth_box, lower_lip_polygon
 
     try: # Wrap main logic in try-except
-        # Use outer mouth landmarks (52-63) to capture the lips only
-        # This avoids including the chin/jawline, preserving the face shape from the swap
+        # Outer mouth/lip landmarks (52-63) — the lip outline only. In this
+        # repo's insightface 2d106 convention these 12 points, taken in index
+        # order, form a SIMPLE (non-self-intersecting) closed polygon that
+        # cv2.fillPoly fills as one solid region directly over the mouth.
+        # This is the last shipped, known-good landmark set; range(52,72)
+        # (the regression) added the inner-lip points and made the path
+        # self-intersect, and the ancient [65,66,62,...,0,8,7...] indices
+        # belong to a different/older landmark convention (they land on the
+        # inner lip + random jaw points, so the mask never covers the mouth).
         lower_lip_order = list(range(52, 64))
 
-        # Check if all indices are valid for the loaded landmarks (already partially done by < 106 check)
+        # All indices must be valid for the loaded landmark set
         if max(lower_lip_order) >= landmarks.shape[0]:
-            # print(f"Warning: Landmark index {max(lower_lip_order)} out of bounds for shape {landmarks.shape[0]}.")
+            # print(f"Warning: Landmark index out of bounds for shape {landmarks.shape[0]}.")
             return mask, mouth_cutout, mouth_box, lower_lip_polygon
 
         lower_lip_landmarks = landmarks[lower_lip_order].astype(np.float32)
@@ -771,10 +1139,25 @@ def create_lower_mouth_mask(
             # print("Warning: Could not calculate valid center for mouth mask.")
             return mask, mouth_cutout, mouth_box, lower_lip_polygon
 
+        # Drive expansion from the Mouth Mask slider so it actually responds.
+        # The known-good version expanded by the now-unused mask_down_size
+        # constant, which is why the slider had no effect.
+        # s: 0.0 (slider ~0, tight lip outline) -> 1.0 (slider 100, mouth->chin).
+        mouth_mask_size = getattr(modules.globals, "mouth_mask_size", 0.0) # 0-100 slider
+        s = max(0.0, min(1.0, mouth_mask_size / 100.0))
 
-        mask_down_size = getattr(modules.globals, "mask_down_size", 0.1) # Default 0.1
-        expansion_factor = 1 + mask_down_size
-        expanded_landmarks = (lower_lip_landmarks - center) * expansion_factor + center
+        # Uniformly scaling a simple polygon about its centroid keeps it simple
+        # (no self-intersection). x grows with expansion_factor; points below
+        # centre (toward the chin) also get an extra downward stretch so high
+        # slider values reach from the mouth down to the chin.
+        expansion_factor = 1.0 + s * 2.0          # 1.0x -> 3.0x
+        chin_bias = 1.0 + s * 2.0                  # extra downward stretch
+        offsets = lower_lip_landmarks - center
+        scale_y = np.where(offsets[:, 1] > 0,
+                           expansion_factor * chin_bias, expansion_factor)
+        expanded_landmarks = lower_lip_landmarks.copy()
+        expanded_landmarks[:, 0] = center[0] + offsets[:, 0] * expansion_factor
+        expanded_landmarks[:, 1] = center[1] + offsets[:, 1] * scale_y
 
         # Ensure landmarks are finite after adjustments
         if not np.all(np.isfinite(expanded_landmarks)):
@@ -881,8 +1264,8 @@ def draw_mouth_mask_visualization(
         print(f"Error drawing polygon for visualization: {e}") # Optional debug
         pass
 
-    # Optional: Draw bounding box (red rectangle)
-    # cv2.rectangle(vis_frame, (min_x, min_y), (max_x, max_y), (0, 0, 255), 1)
+    # Draw bounding box (red rectangle)
+    cv2.rectangle(vis_frame, (min_x, min_y), (max_x, max_y), (0, 0, 255), 2)
 
     # Optional: Add labels
     label_pos_y = min_y - 10 if min_y > 20 else max_y + 15 # Adjust position based on box location
@@ -962,85 +1345,34 @@ def apply_mouth_area(
             # print("Warning: Mouth cutout is invalid after resize attempt.")
             return frame
 
-        # --- Color Correction Step ---
-        # Apply color transfer from ROI (swapped face region) to the original mouth cutout
-        # This helps match lighting/color before blending
-        color_corrected_mouth = resized_mouth_cutout # Default to resized if correction fails
-        try:
-           # Ensure both images are 3 channels for color transfer
-           if len(resized_mouth_cutout.shape) == 3 and resized_mouth_cutout.shape[2] == 3 and \
-              len(roi.shape) == 3 and roi.shape[2] == 3:
-                 color_corrected_mouth = apply_color_transfer(resized_mouth_cutout, roi)
-           else:
-               # print("Warning: Cannot apply color transfer, images not BGR.")
-               pass
-        except cv2.error as ct_e: # Handle potential errors in color transfer
-           # print(f"Warning: Color transfer failed: {ct_e}. Using uncorrected mouth cutout.") # Optional debug
-           pass
-        except Exception as ct_gen_e:
-           # print(f"Warning: Unexpected error during color transfer: {ct_gen_e}")
-           pass
-        # --- End Color Correction ---
-
-
         # --- Mask Creation ---
-        # Create a mask based *specifically* on the mouth_polygon, relative to the ROI
+        # Create a mask based on the mouth_polygon, relative to the ROI
         polygon_mask_roi = np.zeros(roi.shape[:2], dtype=np.uint8)
-        # Adjust polygon coordinates relative to the ROI's top-left corner
         adjusted_polygon = mouth_polygon - [min_x, min_y]
-        # Draw the filled polygon on the ROI mask
         cv2.fillPoly(polygon_mask_roi, [adjusted_polygon.astype(np.int32)], 255)
 
-        # Feather the polygon mask (Gaussian blur)
-        mask_feather_ratio = getattr(modules.globals, "mask_feather_ratio", 12) # Default 12
-        # Calculate feather amount based on the smaller dimension of the box
-        feather_base_dim = min(box_width, box_height)
-        feather_amount = max(1, min(30, feather_base_dim // max(1, mask_feather_ratio))) # Avoid div by zero
-        # Ensure kernel size is odd and positive
+        # Feather the edges with Gaussian blur for smooth blending
+        feather_amount = max(1, min(30, min(box_width, box_height) // 8))
         kernel_size = 2 * feather_amount + 1
-        feathered_polygon_mask = cv2.GaussianBlur(polygon_mask_roi.astype(np.float32), (kernel_size, kernel_size), 0)
+        feathered_mask = cv2.GaussianBlur(polygon_mask_roi.astype(np.float32), (kernel_size, kernel_size), 0)
 
-        # Normalize feathered mask to [0.0, 1.0] range
-        max_val = feathered_polygon_mask.max()
-        if max_val > 1e-6: # Avoid division by zero
-           feathered_polygon_mask = feathered_polygon_mask / max_val
+        # Normalize to [0.0, 1.0]
+        max_val = feathered_mask.max()
+        if max_val > 1e-6:
+            feathered_mask = feathered_mask / max_val
         else:
-           feathered_polygon_mask.fill(0.0) # Mask is all black if max is near zero
-        # --- End Mask Creation ---
+            feathered_mask.fill(0.0)
 
-
-        # --- Refined Blending ---
-        # Get the corresponding ROI from the *full face mask* (already blurred)
-        # Ensure face_mask is float and normalized [0.0, 1.0]
-        if face_mask.dtype != np.float64 and face_mask.dtype != np.float32:
-            face_mask_float = face_mask.astype(np.float32) / 255.0
-        else: # Assume already float [0,1] if type is float
-            face_mask_float = face_mask.astype(np.float32) if face_mask.dtype == np.float64 else face_mask
-        face_mask_roi = face_mask_float[min_y:max_y, min_x:max_x]
-
-        # Combine the feathered mouth polygon mask with the face mask ROI
-        # Use minimum to ensure we only affect area inside both masks (mouth area within face)
-        # This helps blend the edges smoothly with the surrounding swapped face region
-        combined_mask = np.minimum(feathered_polygon_mask, face_mask_roi)
-
-        # Expand mask to 3 channels for blending (ensure it matches image channels)
+        # --- Blending: paste original mouth onto swapped face ---
         if len(frame.shape) == 3 and frame.shape[2] == 3:
-            combined_mask_3channel = combined_mask[:, :, np.newaxis]
+            mask_3ch = feathered_mask[:, :, np.newaxis].astype(np.float32)
+            inv_mask = 1.0 - mask_3ch
 
-            # Ensure data types are compatible for blending
-            # float32 provides sufficient precision for 8-bit image blending
-            combined_mask_f32 = combined_mask_3channel.astype(np.float32)
-            inv_mask = np.float32(1.0) - combined_mask_f32
+            # Blend: (original_mouth * mask) + (swapped_face * (1 - mask))
+            blended_roi = (resized_mouth_cutout.astype(np.float32) * mask_3ch +
+                           roi.astype(np.float32) * inv_mask)
 
-            # Blend: (original_mouth * combined_mask) + (swapped_face_roi * (1 - combined_mask))
-            blended_roi = (color_corrected_mouth * combined_mask_f32 +
-                           roi * inv_mask)
-
-            # Place the blended ROI back into the frame
-            frame[min_y:max_y, min_x:max_x] = blended_roi.astype(np.uint8)
-        else:
-            # print("Warning: Cannot apply mouth mask blending, frame is not 3-channel BGR.")
-            pass # Don't modify frame if it's not BGR
+            frame[min_y:max_y, min_x:max_x] = np.clip(blended_roi, 0, 255).astype(np.uint8)
 
     except Exception as e:
         print(f"Error applying mouth area: {e}") # Optional debug
@@ -1053,10 +1385,13 @@ def apply_mouth_area(
 
 def create_face_mask(face: Face, frame: Frame) -> np.ndarray:
     """Creates a feathered mask covering the whole face area based on landmarks."""
+    if frame is None or not hasattr(frame, "shape") or len(frame.shape) < 2:
+        return np.zeros((0, 0), dtype=np.uint8)
+
     mask = np.zeros(frame.shape[:2], dtype=np.uint8) # Start with uint8
 
     # Validate inputs
-    if face is None or not hasattr(face, 'landmark_2d_106') or frame is None:
+    if face is None or not hasattr(face, 'landmark_2d_106'):
         # print("Warning: Invalid face or frame for create_face_mask.")
         return mask # Return empty mask
 
@@ -1223,4 +1558,4 @@ def apply_color_transfer(source, target):
          # traceback.print_exc()
          return source
 
-    return result_bgr
+    return result_bgr

modules/utilities.py

@@ -30,8 +30,12 @@ def run_ffmpeg(args: List[str]) -> bool:
     try:
         subprocess.check_output(commands, stderr=subprocess.STDOUT)
         return True
-    except Exception:
-        pass
+    except subprocess.CalledProcessError as error:
+        output = error.output.decode(errors="ignore").strip()
+        if output:
+            print(output)
+    except Exception as error:
+        print(f"ffmpeg execution failed: {error}")
     return False
 
 
@@ -61,19 +65,19 @@ def extract_frames(target_path: str) -> None:
     """Extract frames with hardware acceleration and optimized settings."""
     temp_directory_path = get_temp_directory_path(target_path)
     
-    # Use hardware-accelerated decoding and optimized pixel format
+    # Write a contiguous image sequence so the later "%04d.png" input pattern
+    # used during encoding can consume every frame reliably.
     run_ffmpeg(
         [
             "-i", target_path,
             "-vf", "format=rgb24",  # Use video filter for format conversion (faster)
             "-vsync", "0",  # Prevent frame duplication
-            "-frame_pts", "1",  # Preserve frame timing
             os.path.join(temp_directory_path, "%04d.png"),
         ]
     )
 
 
-def create_video(target_path: str, fps: float = 30.0) -> None:
+def create_video(target_path: str, fps: float = 30.0) -> bool:
     """Create video with hardware-accelerated encoding and optimized settings."""
     temp_output_path = get_temp_output_path(target_path)
     temp_directory_path = get_temp_directory_path(target_path)
@@ -182,7 +186,8 @@ def create_video(target_path: str, fps: float = 30.0) -> None:
             "-y",
             temp_output_path,
         ]
-        run_ffmpeg(ffmpeg_args_fallback)
+        success = run_ffmpeg(ffmpeg_args_fallback)
+    return success and os.path.isfile(temp_output_path)
 
 
 def restore_audio(target_path: str, output_path: str) -> None:
@@ -309,3 +314,35 @@ def conditional_download(download_directory_path: str, urls: List[str]) -> None:
 
 def resolve_relative_path(path: str) -> str:
     return os.path.abspath(os.path.join(os.path.dirname(__file__), path))
+
+
+def get_video_dimensions(target_path: str) -> tuple:
+    """Get video width and height using ffprobe."""
+    command = [
+        "ffprobe", "-v", "error",
+        "-select_streams", "v:0",
+        "-show_entries", "stream=width,height",
+        "-of", "csv=p=0:s=x",
+        target_path,
+    ]
+    output = subprocess.check_output(command).decode().strip()
+    width, height = map(int, output.split("x"))
+    return width, height
+
+
+def estimate_frame_count(target_path: str, fps: float = None) -> int:
+    """Estimate total frame count from video duration and fps."""
+    if fps is None:
+        fps = detect_fps(target_path)
+    command = [
+        "ffprobe", "-v", "error",
+        "-show_entries", "format=duration",
+        "-of", "csv=p=0",
+        target_path,
+    ]
+    try:
+        output = subprocess.check_output(command).decode().strip()
+        duration = float(output)
+        return int(duration * fps)
+    except Exception:
+        return 0

modules/video_capture.py

@@ -1,5 +1,7 @@
 import cv2
 import numpy as np
+import sys
+import time
 from typing import Optional, Tuple, Callable
 import platform
 import threading
@@ -17,6 +19,10 @@ class VideoCapturer:
         self._frame_ready = threading.Event()
         self.is_running = False
         self.cap = None
+        # Actual values reported by the camera after configuration
+        self.actual_width: int = 0
+        self.actual_height: int = 0
+        self.actual_fps: float = 0.0
 
         # Initialize Windows-specific components if on Windows
         if platform.system() == "Windows":
@@ -32,17 +38,35 @@ class VideoCapturer:
         """Initialize and start video capture"""
         try:
             if platform.system() == "Windows":
-                # Windows-specific capture methods
+                # device_index comes from pygrabber.FilterGraph (DirectShow
+                # enumeration), so open with DSHOW first to preserve mapping.
+                # MSMF and DirectShow enumerate cameras in different orders, so
+                # opening MSMF with a DSHOW index silently selects the wrong
+                # camera. MSMF/ANY remain as fallbacks for cameras DSHOW can't
+                # open.
+                #
+                # Pass codec + resolution + fps as construction params (OpenCV
+                # 4.6+). DSHOW locks the pixel format at open time and ignores
+                # later cap.set(CAP_PROP_FOURCC, ...) — without this, DSHOW
+                # falls back to uncompressed YUYV at 1080p, which is USB-
+                # bandwidth-limited to ~5 fps. Setting MJPG at construction
+                # negotiates compressed frames from the first read.
+                mjpg = cv2.VideoWriter_fourcc(*'MJPG')
+                open_params = [
+                    cv2.CAP_PROP_FOURCC, mjpg,
+                    cv2.CAP_PROP_FRAME_WIDTH, width,
+                    cv2.CAP_PROP_FRAME_HEIGHT, height,
+                    cv2.CAP_PROP_FPS, fps,
+                ]
                 capture_methods = [
-                    (self.device_index, cv2.CAP_DSHOW),  # Try DirectShow first
-                    (self.device_index, cv2.CAP_ANY),  # Then try default backend
-                    (-1, cv2.CAP_ANY),  # Try -1 as fallback
-                    (0, cv2.CAP_ANY),  # Finally try 0 without specific backend
+                    (self.device_index, cv2.CAP_DSHOW),
+                    (self.device_index, cv2.CAP_MSMF),
+                    (self.device_index, cv2.CAP_ANY),
                 ]
 
                 for dev_id, backend in capture_methods:
                     try:
-                        self.cap = cv2.VideoCapture(dev_id, backend)
+                        self.cap = cv2.VideoCapture(dev_id, backend, open_params)
                         if self.cap.isOpened():
                             break
                         self.cap.release()
@@ -55,10 +79,29 @@ class VideoCapturer:
             if not self.cap or not self.cap.isOpened():
                 raise RuntimeError("Failed to open camera")
 
-            # Configure format
-            self.cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
-            self.cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
-            self.cap.set(cv2.CAP_PROP_FPS, fps)
+            # Belt-and-braces: also set via cap.set() for backends that honor
+            # post-open changes (MSMF, V4L2). DSHOW ignores these, but the
+            # construction params above already handled it.
+            if platform.system() != "Windows":
+                self.cap.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc(*'MJPG'))
+                self.cap.set(cv2.CAP_PROP_FRAME_WIDTH, width)
+                self.cap.set(cv2.CAP_PROP_FRAME_HEIGHT, height)
+                self.cap.set(cv2.CAP_PROP_FPS, fps)
+
+            # Read back resolution (usually reliable)
+            self.actual_width = int(self.cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+            self.actual_height = int(self.cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+            # CAP_PROP_FPS is unreliable on DirectShow — often reports 30
+            # even when the camera delivers 60.  Measure empirically by
+            # timing a burst of frames.
+            reported_fps = self.cap.get(cv2.CAP_PROP_FPS)
+            self.actual_fps = self._measure_fps(warmup=10, sample=30,
+                                                fallback=reported_fps or fps)
+
+            print(f"[VideoCapturer] {self.actual_width}x{self.actual_height} "
+                  f"@ {self.actual_fps:.1f}fps (reported={reported_fps:.0f})",
+                  flush=True)
 
             self.is_running = True
             return True
@@ -89,6 +132,29 @@ class VideoCapturer:
             self.is_running = False
             self.cap = None
 
+    def _measure_fps(self, warmup: int = 10, sample: int = 30,
+                     fallback: float = 30.0) -> float:
+        """Read warmup+sample frames and return measured FPS.
+
+        This is more reliable than CAP_PROP_FPS which often lies on
+        DirectShow.  Takes ~0.5-1s at startup but gives a ground-truth
+        number for adaptive polling/detection intervals.
+        """
+        try:
+            for _ in range(warmup):
+                self.cap.read()
+            t0 = time.perf_counter()
+            for _ in range(sample):
+                ret, _ = self.cap.read()
+                if not ret:
+                    return fallback
+            elapsed = time.perf_counter() - t0
+            if elapsed <= 0:
+                return fallback
+            return sample / elapsed
+        except Exception:
+            return fallback
+
     def set_frame_callback(self, callback: Callable[[np.ndarray], None]) -> None:
         """Set callback for frame processing"""
         self.frame_callback = callback

requirements.txt

@@ -5,12 +5,13 @@ cv2_enumerate_cameras==1.1.15
 onnx==1.18.0
 insightface==0.7.3
 psutil==5.9.8
-tk==0.1.0
-customtkinter==5.2.2
+PySide6>=6.7,<7
 pillow==12.1.1
+tqdm>=4.65.0
 onnxruntime-silicon==1.16.3; sys_platform == 'darwin' and platform_machine == 'arm64'
-onnxruntime-gpu==1.24.2; sys_platform != 'darwin'
-tensorflow; sys_platform != 'darwin'
+onnxruntime-gpu==1.23.2; sys_platform != 'darwin'
+tensorflow>=2.15.0; sys_platform != 'darwin'
+tensorflow>=2.15.0; sys_platform == 'darwin' and python_version < '3.13'
 opennsfw2==0.10.2
 protobuf==4.25.1
-pygrabber
+pygrabber; sys_platform == 'win32'

run.py

@@ -1,7 +1,73 @@
 #!/usr/bin/env python3
 
-# Import the tkinter fix to patch the ScreenChanged error
-import tkinter_fix
+import os
+import sys
+
+# Add the project root to PATH so bundled ffmpeg/ffprobe are found
+project_root = os.path.dirname(os.path.abspath(__file__))
+os.environ["PATH"] = project_root + os.pathsep + os.environ.get("PATH", "")
+
+# On Windows, register NVIDIA CUDA DLL directories so onnxruntime-gpu can
+# find cuDNN/cublas. Python 3.8+ ignores PATH for extension-module native deps —
+# os.add_dll_directory() is required. Also keep PATH for child processes/ffmpeg.
+if sys.platform == "win32":
+    _site_packages = os.path.join(sys.prefix, "Lib", "site-packages")
+    _venv_site_packages = os.path.join(project_root, "venv", "Lib", "site-packages")
+    for _sp in (_site_packages, _venv_site_packages):
+        _candidate_dirs = []
+        _torch_lib = os.path.join(_sp, "torch", "lib")
+        if os.path.isdir(_torch_lib):
+            _candidate_dirs.append(_torch_lib)
+        _nvidia_dir = os.path.join(_sp, "nvidia")
+        if os.path.isdir(_nvidia_dir):
+            for _pkg in os.listdir(_nvidia_dir):
+                _bin_dir = os.path.join(_nvidia_dir, _pkg, "bin")
+                if os.path.isdir(_bin_dir):
+                    _candidate_dirs.append(_bin_dir)
+        for _d in _candidate_dirs:
+            os.environ["PATH"] = _d + os.pathsep + os.environ["PATH"]
+            try:
+                os.add_dll_directory(_d)
+            except (OSError, AttributeError):
+                pass
+
+# On Linux, pre-load NVIDIA shared libraries (cuDNN, cuBLAS, nvrtc...) shipped
+# inside the venv via pip wheels (nvidia-cudnn-cu12, etc.). LD_LIBRARY_PATH
+# cannot be set after Python starts, so we use ctypes.CDLL with RTLD_GLOBAL
+# instead. This makes symbols available to onnxruntime when it dlopens its
+# CUDA provider.
+if sys.platform.startswith("linux"):
+    import ctypes
+    import glob
+    _py_lib = f"python{sys.version_info.major}.{sys.version_info.minor}"
+    _site_packages_candidates = [
+        os.path.join(project_root, "venv", "lib", _py_lib, "site-packages"),
+        os.path.join(sys.prefix, "lib", _py_lib, "site-packages"),
+    ]
+    for _sp in _site_packages_candidates:
+        _nvidia_dir = os.path.join(_sp, "nvidia")
+        if not os.path.isdir(_nvidia_dir):
+            continue
+        for _pkg in os.listdir(_nvidia_dir):
+            _lib_dir = os.path.join(_nvidia_dir, _pkg, "lib")
+            if not os.path.isdir(_lib_dir):
+                continue
+            # Also expose the directory to child processes, without
+            # duplicating an entry that is already present.
+            _ldp = os.environ.get("LD_LIBRARY_PATH", "")
+            if _lib_dir not in _ldp.split(os.pathsep):
+                os.environ["LD_LIBRARY_PATH"] = (
+                    _lib_dir + (os.pathsep + _ldp if _ldp else "")
+                )
+            for _so in sorted(glob.glob(os.path.join(_lib_dir, "lib*.so*"))):
+                try:
+                    ctypes.CDLL(_so, mode=ctypes.RTLD_GLOBAL)
+                except OSError:
+                    pass
+        break
+
+from modules import platform_info
+platform_info.print_banner()
 
 from modules import core
 

tests/test_face_analyser_get_one_face.py

@@ -0,0 +1,97 @@
+import importlib
+import sys
+import types
+import unittest
+from unittest.mock import patch
+
+
+def _install_import_stubs():
+    sys.modules.setdefault(
+        "insightface",
+        types.SimpleNamespace(app=types.SimpleNamespace(FaceAnalysis=object)),
+    )
+    sys.modules.setdefault(
+        "cv2",
+        types.SimpleNamespace(
+            IMREAD_COLOR=1,
+            imread=lambda *_args, **_kwargs: None,
+            imdecode=lambda *_args, **_kwargs: None,
+            imencode=lambda *_args, **_kwargs: (
+                True,
+                types.SimpleNamespace(tofile=lambda *_a, **_k: None),
+            ),
+        ),
+    )
+    sys.modules.setdefault(
+        "numpy",
+        types.SimpleNamespace(uint8=object, fromfile=lambda *_args, **_kwargs: b""),
+    )
+    sys.modules.setdefault(
+        "tqdm",
+        types.SimpleNamespace(tqdm=lambda iterable, **_kwargs: iterable),
+    )
+    sys.modules["modules.typing"] = types.SimpleNamespace(Frame=object)
+    sys.modules["modules.cluster_analysis"] = types.SimpleNamespace(
+        find_cluster_centroids=lambda *args, **kwargs: [],
+        find_closest_centroid=lambda *args, **kwargs: (0, None),
+    )
+    sys.modules["modules.utilities"] = types.SimpleNamespace(
+        get_temp_directory_path=lambda path: path,
+        create_temp=lambda path: None,
+        extract_frames=lambda path: None,
+        clean_temp=lambda path: None,
+        get_temp_frame_paths=lambda path: [],
+    )
+
+
+def _load_face_analyser():
+    _install_import_stubs()
+    sys.modules.pop("modules.face_analyser", None)
+    return importlib.import_module("modules.face_analyser")
+
+
+class Face:
+    def __init__(self, left):
+        self.bbox = [left, 0, 10, 10]
+
+
+class GetOneFaceTests(unittest.TestCase):
+    def test_uses_supplied_detected_faces_without_reanalysing_frame(self):
+        face_analyser = _load_face_analyser()
+        right = Face(20)
+        left = Face(5)
+
+        with patch.object(
+            face_analyser,
+            "_analyse_faces",
+            side_effect=AssertionError("should not analyse"),
+        ):
+            self.assertIs(face_analyser.get_one_face("frame", [right, left]), left)
+
+    def test_supplied_empty_detected_faces_returns_none(self):
+        face_analyser = _load_face_analyser()
+
+        with patch.object(
+            face_analyser,
+            "_analyse_faces",
+            side_effect=AssertionError("should not analyse"),
+        ):
+            self.assertIsNone(face_analyser.get_one_face("frame", []))
+
+    def test_without_supplied_faces_preserves_existing_detection_path(self):
+        face_analyser = _load_face_analyser()
+        right = Face(30)
+        left = Face(3)
+
+        with patch.object(face_analyser, "_is_dml", return_value=False), patch.object(
+            face_analyser,
+            "_analyse_faces",
+            return_value=[right, left],
+        ) as analyse:
+            self.assertIs(face_analyser.get_one_face("frame"), left)
+
+        analyse.assert_called_once_with("frame")
+
+
+if __name__ == "__main__":
+    unittest.main()

tkinter_fix.py

@@ -1,3 +1,6 @@
+import os
+os.environ.setdefault('TK_SILENCE_DEPRECATION', '1')
+
 import tkinter
 
 # Only needs to be imported once at the beginning of the application