Sync dependency updates with current main

src/remove_ai_watermarks/__init__.py

@@ -25,7 +25,7 @@ _os.environ.setdefault("TRANSFORMERS_VERBOSITY", "error")
 _warnings.filterwarnings("ignore", message=r".*ImageProcessorFast.*")
 
 
-__version__ = "0.19.0"
+__version__ = "0.20.0"
 
 __all__ = ["__version__", "remove_visible", "visible_provenance"]
 

src/remove_ai_watermarks/cli.py

@@ -21,6 +21,8 @@ import click
 from remove_ai_watermarks import __version__, image_io, watermark_registry
 from remove_ai_watermarks.noai.constants import SUPPORTED_FORMATS
 from remove_ai_watermarks.noai.watermark_profiles import (
+    resolve_seed,
+    resolve_steps,
     resolve_strength,
     strength_default_help,
     vendor_for_strength,
@@ -137,6 +139,23 @@ def _validate_image(path: Path) -> Path:
     return path
 
 
+def _resolved_strength_for_display(
+    source: Path,
+    strength: float | None,
+    vendor: str | None,
+    pipeline: str,
+) -> float:
+    """Resolve the same profile-specific strength the engine will execute."""
+    if pipeline == "qwen-zimage" and strength is None:
+        from PIL import Image
+
+        from remove_ai_watermarks.noai.qwen_zimage_pipeline import resolution_adaptive_denoise
+
+        with Image.open(source) as image:
+            return resolution_adaptive_denoise(image.width, image.height)
+    return resolve_strength(strength, vendor, pipeline)
+
+
 # Shared option decorator for commands that run the invisible-watermark pipeline.
 # Both cmd_invisible and cmd_all expose this flag; defining it once avoids
 # copy-paste drift.
@@ -184,8 +203,9 @@ _adaptive_polish_option = click.option(
     default=True,
     help="Restore the input's detail level after removal (capped unsharp + edge-masked grain "
     "targeting the input's sharpness, sparing text), countering the over-smoothed look. ON by "
-    "default; it self-limits where there is no detail deficit (text/flat graphics), so it is a "
-    "no-op there. Pass --no-adaptive-polish to disable. Independent of --unsharp/--humanize.",
+    "default except for qwen-zimage, whose upstream-matching output is left unchanged; it "
+    "self-limits where there is no detail deficit (text/flat graphics). Pass --adaptive-polish "
+    "or --no-adaptive-polish to override. Independent of --unsharp/--humanize.",
 )
 
 
@@ -204,7 +224,7 @@ def _tile_options(f: Any) -> Any:
         "--tile-size",
         type=int,
         default=1024,
-        help="Tile dimension in px for --tile (SDXL's training size). Default 1024.",
+        help="Tile dimension in px for --tile. Default 1024.",
     )(f)
     return click.option(
         "--tile/--no-tile",
@@ -227,8 +247,8 @@ _guidance_scale_option = click.option(
     "--guidance-scale",
     type=float,
     default=None,
-    help="Classifier-free guidance scale (CFG). Default: 7.5 (the library default). "
-    "Lower = follow the prompt less / stay closer to the input.",
+    help="Classifier-free guidance scale (CFG). Default: 7.5, except qwen-zimage "
+    "fixes CFG at 1.0. Lower = follow the prompt less / stay closer to the input.",
 )
 
 
@@ -255,13 +275,15 @@ def _normalize_pipeline(ctx: click.Context, param: click.Parameter, value: str |
 # ``controlnet`` (the default-SELECTED value), ``sdxl`` (plain SDXL img2img) and
 # ``qwen`` (Qwen-Image, CUDA/cloud-class) are the current profiles; ``default`` is an
 # OUTDATED back-compat alias for ``sdxl`` (warned + normalized away by _normalize_pipeline).
-_PIPELINE_CHOICES = ["sdxl", "controlnet", "qwen", "default"]
+_PIPELINE_CHOICES = ["sdxl", "controlnet", "qwen", "qwen-zimage", "default"]
 _PIPELINE_HELP = (
     "Pipeline profile. controlnet (DEFAULT) = SDXL + canny ControlNet that preserves "
     "text/faces via edge conditioning while removing SynthID; sdxl = plain SDXL img2img "
     "(lighter, no extra model download, but leaves SynthID on flat-graphic content); "
     "qwen = Qwen-Image (20B, Apache-2.0) img2img, best text/structure preservation but "
-    "CUDA/cloud-class (does not fit MPS). ('default' is an OUTDATED alias for 'sdxl'.)"
+    "CUDA/cloud-class; qwen-zimage = Qwen-Image-2512 + Lightning + Canny, followed by "
+    "SAM-masked Z-Image face repair (CUDA-only; install the qwen-zimage extra). "
+    "('default' is an OUTDATED alias for 'sdxl'.)"
 )
 
 # Shared --pipeline / --strength decorators so the three diffusion commands
@@ -336,6 +358,19 @@ def _resolve_auto_polish(auto: bool, adaptive_polish: bool) -> bool:
     return adaptive_polish
 
 
+def _resolve_profile_polish(auto: bool, adaptive_polish: bool, pipeline: str) -> bool:
+    """Keep the upstream qwen-zimage output unchanged unless polish was explicit."""
+    adaptive_polish = _resolve_auto_polish(auto, adaptive_polish)
+    if pipeline != "qwen-zimage":
+        return adaptive_polish
+    ctx = click.get_current_context(silent=True)
+    if ctx is None:
+        return adaptive_polish
+    if ctx.get_parameter_source("adaptive_polish") == click.core.ParameterSource.DEFAULT:
+        return False
+    return adaptive_polish
+
+
 def _warn_if_esrgan_unavailable(upscaler: str) -> None:
     """Tell the user once if ``--upscaler esrgan`` will silently fall back to Lanczos.
 
@@ -827,7 +862,12 @@ def cmd_erase(
     "-o", "--output", type=click.Path(path_type=Path), default=None, help="Output path (default: <source>_clean.<ext>)."
 )
 @_strength_option
-@click.option("--steps", type=int, default=50, help="Number of denoising steps. Default: 50.")
+@click.option(
+    "--steps",
+    type=int,
+    default=None,
+    help="Number of denoising steps. Default: 4 for qwen-zimage, 50 otherwise.",
+)
 @_pipeline_option
 @click.option(
     "--device",
@@ -835,7 +875,12 @@ def cmd_erase(
     default="auto",
     help="Inference device.",
 )
-@click.option("--seed", type=int, default=None, help="Random seed for reproducibility.")
+@click.option(
+    "--seed",
+    type=int,
+    default=None,
+    help="Random seed for reproducibility. Default: 0 for qwen-zimage, random otherwise.",
+)
 @click.option("--hf-token", type=str, default=None, help="HuggingFace API token.")
 @click.option(
     "--humanize", type=float, default=0.0, help="Analog Humanizer film grain intensity (0 = off, typical: 2.0-6.0)."
@@ -862,7 +907,7 @@ def cmd_invisible(
     source: Path,
     output: Path | None,
     strength: float | None,
-    steps: int,
+    steps: int | None,
     pipeline: str,
     device: str,
     seed: int | None,
@@ -898,8 +943,10 @@ def cmd_invisible(
     from remove_ai_watermarks.invisible_engine import InvisibleEngine
 
     source = _validate_image(source)
+    steps = resolve_steps(steps, pipeline)
+    seed = resolve_seed(seed, pipeline)
     _warn_if_esrgan_unavailable(upscaler)
-    adaptive_polish = _resolve_auto_polish(auto, adaptive_polish)
+    adaptive_polish = _resolve_profile_polish(auto, adaptive_polish, pipeline)
     if output is None:
         output = source.with_stem(source.stem + "_clean")
 
@@ -929,7 +976,7 @@ def cmd_invisible(
     vendor = vendor_for_strength(source)
     console.print(f"  Input:    {source.name}")
     console.print(f"  Pipeline: {pipeline}")
-    console.print(f"  Strength: {resolve_strength(strength, vendor, pipeline)}  Steps: {steps}")
+    console.print(f"  Strength: {_resolved_strength_for_display(source, strength, vendor, pipeline)}  Steps: {steps}")
 
     t0 = time.monotonic()
     result_path = engine.remove_watermark(
@@ -1101,7 +1148,12 @@ def cmd_identify(ctx: click.Context, source: Path, no_visible: bool, as_json: bo
 @_visible_backend_option
 @_visible_sensitivity_option
 @_strength_option
-@click.option("--steps", type=int, default=50, help="Number of denoising steps for invisible removal.")
+@click.option(
+    "--steps",
+    type=int,
+    default=None,
+    help="Number of denoising steps. Default: 4 for qwen-zimage, 50 otherwise.",
+)
 @_pipeline_option
 @_model_option
 @click.option(
@@ -1110,7 +1162,12 @@ def cmd_identify(ctx: click.Context, source: Path, no_visible: bool, as_json: bo
     default="auto",
     help="Inference device.",
 )
-@click.option("--seed", type=int, default=None, help="Random seed for reproducibility.")
+@click.option(
+    "--seed",
+    type=int,
+    default=None,
+    help="Random seed for reproducibility. Default: 0 for qwen-zimage, random otherwise.",
+)
 @click.option("--hf-token", type=str, default=None, help="HuggingFace API token.")
 @click.option(
     "--humanize", type=float, default=0.0, help="Analog Humanizer film grain intensity (0 = off, typical: 2.0-6.0)."
@@ -1138,7 +1195,7 @@ def cmd_all(
     backend: str,
     sensitivity: str,
     strength: float | None,
-    steps: int,
+    steps: int | None,
     pipeline: str,
     model: str | None,
     device: str,
@@ -1169,8 +1226,10 @@ def cmd_all(
     """
     _banner()
     source = _validate_image(source)
+    steps = resolve_steps(steps, pipeline)
+    seed = resolve_seed(seed, pipeline)
     _warn_if_esrgan_unavailable(upscaler)
-    adaptive_polish = _resolve_auto_polish(auto, adaptive_polish)
+    adaptive_polish = _resolve_profile_polish(auto, adaptive_polish, pipeline)
 
     if output is None:
         output = source.with_stem(source.stem + "_clean")
@@ -1260,7 +1319,9 @@ def cmd_all(
             # already lost its C2PA to the visible-removal pass, so reading it would
             # always resolve to the unknown-vendor default.
             vendor = vendor_for_strength(source)
-            console.print(f"    Strength: {resolve_strength(strength, vendor, pipeline)}  Steps: {steps}")
+            console.print(
+                f"    Strength: {_resolved_strength_for_display(source, strength, vendor, pipeline)}  Steps: {steps}"
+            )
             inv_engine.remove_watermark(
                 image_path=tmp_path,
                 output_path=tmp_path,
@@ -1521,7 +1582,12 @@ def _process_batch_image(
     "--mode", type=click.Choice(["visible", "invisible", "metadata", "all"]), default="visible", help="Processing mode."
 )
 @_strength_option
-@click.option("--steps", type=int, default=50, help="Number of denoising steps (invisible mode).")
+@click.option(
+    "--steps",
+    type=int,
+    default=None,
+    help="Number of denoising steps. Default: 4 for qwen-zimage, 50 otherwise.",
+)
 @_visible_backend_option
 @_visible_sensitivity_option
 @click.option(
@@ -1534,7 +1600,12 @@ def _process_batch_image(
     default="auto",
     help="Inference device.",
 )
-@click.option("--seed", type=int, default=None, help="Random seed for reproducibility.")
+@click.option(
+    "--seed",
+    type=int,
+    default=None,
+    help="Random seed for reproducibility. Default: 0 for qwen-zimage, random otherwise.",
+)
 @click.option("--hf-token", type=str, default=None, help="HuggingFace API token.")
 @click.option(
     "--max-resolution",
@@ -1559,7 +1630,7 @@ def cmd_batch(
     mode: str,
     output_dir: Path | None,
     strength: float | None,
-    steps: int,
+    steps: int | None,
     pipeline: str,
     device: str,
     seed: int | None,
@@ -1599,7 +1670,9 @@ def cmd_batch(
     console.print(f"  Mode: {mode}")
     if mode in ("invisible", "all"):
         _warn_if_esrgan_unavailable(upscaler)
-    adaptive_polish = _resolve_auto_polish(auto, adaptive_polish)
+    adaptive_polish = _resolve_profile_polish(auto, adaptive_polish, pipeline)
+    steps = resolve_steps(steps, pipeline)
+    seed = resolve_seed(seed, pipeline)
     options = _BatchOptions(
         strength=strength,
         steps=steps,

src/remove_ai_watermarks/invisible_engine.py

@@ -19,7 +19,12 @@ import warnings
 from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
-from .noai.watermark_profiles import DEFAULT_MODEL_ID as DEFAULT_SDXL_MODEL_ID
+from .noai.watermark_profiles import (
+    DEFAULT_MODEL_ID as DEFAULT_SDXL_MODEL_ID,
+)
+from .noai.watermark_profiles import (
+    resolve_seed,
+)
 
 if TYPE_CHECKING:
     from collections.abc import Callable
@@ -105,7 +110,9 @@ class InvisibleEngine:
                 that preserves text/face structure via edge conditioning while removing
                 SynthID), "sdxl" (plain SDXL img2img, lighter but leaves SynthID on
                 flat-graphic content), or "qwen" (Qwen-Image 20B img2img, best text/
-                structure preservation but CUDA/cloud-class). "default" aliases "sdxl".
+                structure preservation but CUDA/cloud-class), or "qwen-zimage"
+                (Qwen-Image-2512 Lightning + Canny, then SAM-masked Z-Image face
+                repair; CUDA-only). "default" aliases "sdxl".
             hf_token: HuggingFace API token.
             progress_callback: Optional callback for progress messages.
             controlnet_conditioning_scale: ControlNet structure-preservation
@@ -161,7 +168,7 @@ class InvisibleEngine:
         image_path: Path,
         output_path: Path | None = None,
         strength: float | None = None,
-        num_inference_steps: int = 100,
+        num_inference_steps: int | None = None,
         guidance_scale: float | None = None,
         seed: int | None = None,
         humanize: float = 0.0,
@@ -182,9 +189,12 @@ class InvisibleEngine:
             output_path: Output path (None = overwrite source).
             strength: Denoising strength (0.0-1.0). None -> the vendor-adaptive
                 default.
-            num_inference_steps: Number of denoising steps.
+            num_inference_steps: Number of denoising steps. None keeps the existing
+                100-step library default, except qwen-zimage uses its required
+                four-step Lightning schedule.
             guidance_scale: Classifier-free guidance scale.
-            seed: Random seed for reproducibility.
+            seed: Random seed for reproducibility. None resolves to 0 for
+                qwen-zimage and stays random for the other profiles.
             humanize: Intensity of Analog Humanizer film grain (0 = off).
             unsharp: Final unsharp-mask sharpening strength (0 = off, default).
                 Applied last to counter the soft / over-smoothed look of the
@@ -222,6 +232,12 @@ class InvisibleEngine:
         """
         import tempfile
 
+        if num_inference_steps is None:
+            profile = getattr(self._remover, "model_profile", None)
+            num_inference_steps = 4 if profile == "qwen-zimage" else 100
+        profile = getattr(self._remover, "model_profile", "controlnet")
+        seed = resolve_seed(seed, profile)
+
         from PIL import Image, ImageOps
 
         # Resolution policy: a max_resolution cap (0 = none) bounds memory on huge
@@ -242,7 +258,18 @@ class InvisibleEngine:
         # reassigned to the resized copy below; PIL resize returns a new object).
         reference_pil = image
 
-        target = _target_size(image.width, image.height, max_resolution, min_resolution)
+        # qwen-zimage operates at the input's native geometry in its reference graph.
+        # Keep an explicit max cap available for callers, but do not apply the SDXL
+        # 1024px minimum-resolution floor to this profile.
+        effective_min_resolution = (
+            0 if getattr(self._remover, "model_profile", None) == "qwen-zimage" else min_resolution
+        )
+        target = _target_size(
+            image.width,
+            image.height,
+            max_resolution,
+            effective_min_resolution,
+        )
         if target is not None:
             upscaling = max(target) > max(image.width, image.height)
             if self._progress_callback:
@@ -356,9 +383,12 @@ class InvisibleEngine:
         input_dir: Path,
         output_dir: Path,
         strength: float | None = None,
-        steps: int = 50,
+        steps: int | None = None,
     ) -> list[Path]:
         """Remove invisible watermarks from all images in a directory."""
+        if steps is None:
+            profile = getattr(self._remover, "model_profile", None)
+            steps = 4 if profile == "qwen-zimage" else 50
         return self._remover.remove_watermark_batch(
             input_dir=input_dir,
             output_dir=output_dir,

src/remove_ai_watermarks/noai/qwen_zimage_pipeline.py

@@ -0,0 +1,903 @@
+"""Qwen 2512 Canny regeneration followed by masked Z-Image face repair.
+
+This profile ports the two-stage architecture used by cebeuq/Synthid-Bypass:
+
+1. Qwen-Image-2512 img2img with the 4-step Lightning LoRA and the DiffSynth
+   blockwise Canny ControlNet regenerates the whole image, optionally as
+   overlapping feather-blended tiles for large inputs.
+2. Faces are detected on the original image, refined to masks with SAM, regenerated
+   from the original face crops with Z-Image Turbo, and feathered into stage 1.
+
+The runtime intentionally uses permissively licensed YuNet instead of the reference
+workflow's Ultralytics detector. All diffusion and segmentation models remain the same
+model families and the denoise formulas are direct ports of the reference custom node.
+"""
+
+# DiffSynth, torch, transformers, and cv2 expose mostly untyped tensor/array APIs.
+# Keep the relaxation local to this optional ML boundary.
+# pyright: reportUnknownMemberType=false, reportUnknownArgumentType=false, reportUnknownVariableType=false, reportUnknownParameterType=false, reportMissingTypeArgument=false, reportMissingTypeStubs=false, reportMissingImports=false, reportArgumentType=false, reportAssignmentType=false, reportReturnType=false, reportCallIssue=false, reportIndexIssue=false, reportOperatorIssue=false, reportOptionalMemberAccess=false, reportOptionalCall=false, reportOptionalSubscript=false, reportOptionalOperand=false, reportAttributeAccessIssue=false, reportPrivateImportUsage=false, reportPrivateUsage=false, reportUnnecessaryComparison=false
+from __future__ import annotations
+
+import contextlib
+import hashlib
+import logging
+import math
+import os
+import tempfile
+import urllib.request
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+import numpy as np
+from PIL import Image
+
+from remove_ai_watermarks.noai.watermark_profiles import resolve_seed
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+
+log = logging.getLogger(__name__)
+
+QWEN_IMAGE_2512_MODEL_ID = "Qwen/Qwen-Image-2512"
+QWEN_CANNY_CONTROLNET_MODEL_ID = "DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Canny"
+QWEN_LIGHTNING_MODEL_ID = "lightx2v/Qwen-Image-2512-Lightning"
+QWEN_LIGHTNING_PATTERN = "Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors"
+ZIMAGE_TURBO_MODEL_ID = "Tongyi-MAI/Z-Image-Turbo"
+SAM_MODEL_ID = "facebook/sam-vit-base"
+
+YUNET_MODEL_URL = (
+    "https://media.githubusercontent.com/media/opencv/opencv_zoo/main/"
+    "models/face_detection_yunet/face_detection_yunet_2023mar.onnx"
+)
+YUNET_MODEL_NAME = "face_detection_yunet_2023mar.onnx"
+YUNET_MODEL_SHA256 = "8f2383e4dd3cfbb4553ea8718107fc0423210dc964f9f4280604804ed2552fa4"
+# The upstream graph's 0.2 threshold belongs to YOLO and does not transfer to
+# YuNet's score calibration. At 0.2 YuNet admitted background and decorative
+# false positives, multiplying the serial Z-Image face-stage cost. A 0.5 gate
+# retained all visible faces in the public and upstream comparison fixtures.
+YUNET_SCORE_THRESHOLD = 0.5
+
+GLOBAL_STEPS = 4
+FACE_STEPS = 8
+GLOBAL_CFG = 1.0
+FACE_CFG = 1.0
+GLOBAL_CONTROLNET_SCALE = 1.0
+RESIDENT_FACE_MODEL_MIN_VRAM_GIB = 64.0
+
+# The source graph uses normalized Canny thresholds 0.05 and 0.25. OpenCV takes
+# byte thresholds, so round 255*x to the matching integer values.
+_CANNY_LOW = 13
+_CANNY_HIGH = 64
+
+# These strings intentionally preserve the reference workflow spelling. They are
+# model inputs, not user-facing copy, and changing them would change the port.
+_GLOBAL_PROMPT = "ultra clear and smoothe skin, spotless skin"
+_GLOBAL_NEGATIVE = "moles, freckes, high detail skin"
+_FACE_PROMPT = ""
+_FACE_NEGATIVE = "blurry, ugly, bad quality,"
+
+
+def resolve_face_model_residency(
+    requested: bool | None,
+    *,
+    total_memory_gib: float,
+) -> bool:
+    """Keep the Z-Image stack resident when explicitly requested or safely sized."""
+    if requested is not None:
+        return requested
+    return total_memory_gib >= RESIDENT_FACE_MODEL_MIN_VRAM_GIB
+
+
+def _pin_vram_managed_models(pipe: Any) -> None:
+    """Move the managed Z-Image stack to CUDA once and make offload a no-op."""
+    model_names = ["text_encoder", "dit", "vae_encoder", "vae_decoder"]
+    for name in model_names:
+        model = getattr(pipe, name, None)
+        if model is None:
+            continue
+        for module in model.modules():
+            if not all(
+                hasattr(module, attribute)
+                for attribute in (
+                    "offload_dtype",
+                    "offload_device",
+                    "onload_dtype",
+                    "onload_device",
+                    "preparing_dtype",
+                    "preparing_device",
+                    "computation_dtype",
+                    "computation_device",
+                )
+            ):
+                continue
+            module.offload_dtype = module.computation_dtype
+            module.offload_device = module.computation_device
+            module.onload_dtype = module.computation_dtype
+            module.onload_device = module.computation_device
+            module.preparing_dtype = module.computation_dtype
+            module.preparing_device = module.computation_device
+    pipe.load_models_to_device(model_names)
+
+
+def _cached_prompt_process(original_process: Any) -> Any:
+    cache: dict[str, dict[str, Any]] = {}
+
+    def cached_process(
+        runtime_pipe: Any,
+        prompt: str,
+        edit_image: Any = None,
+    ) -> dict[str, Any]:
+        if edit_image is not None:
+            return original_process(runtime_pipe, prompt, edit_image=edit_image)
+        if prompt not in cache:
+            cache[prompt] = original_process(runtime_pipe, prompt, edit_image=None)
+        return cache[prompt]
+
+    return cached_process
+
+
+def _cache_static_prompt_embeddings(
+    pipe: Any,
+    output_params: tuple[str, ...],
+) -> bool:
+    """Memoize a prompt unit when its embedding depends only on static text."""
+    for unit in pipe.units:
+        if tuple(getattr(unit, "output_params", ())) == output_params:
+            unit.process = _cached_prompt_process(unit.process)
+            return True
+    return False
+
+
+def _clamp(value: float, minimum: float, maximum: float) -> float:
+    if maximum < minimum:
+        minimum, maximum = maximum, minimum
+    return max(minimum, min(maximum, value))
+
+
+def resolution_adaptive_denoise(
+    width: int,
+    height: int,
+    *,
+    adaptive_level: int = 6,
+    denoise_min: float = 0.08,
+    denoise_max: float = 0.15,
+) -> float:
+    """Port the reference resolution-based adaptive denoise calculation.
+
+    At neutral level 5, 0.30 MP maps to ``denoise_min`` and 3.70 MP maps to
+    ``denoise_max``. Levels above or below 5 add the same asymmetric spread as the
+    reference custom node.
+    """
+    image_mp = max(1.0, float(width) * float(height)) / 1_000_000.0
+    normalized = _clamp((image_mp - 0.30) / (3.70 - 0.30), 0.0, 1.0)
+
+    minimum = float(denoise_min)
+    maximum = float(denoise_max)
+    if maximum < minimum:
+        minimum, maximum = maximum, minimum
+    denoise_range = maximum - minimum
+    base = minimum + denoise_range * normalized
+
+    level = int(adaptive_level)
+    if level >= 5:
+        offset = ((float(level) - 5.0) / 5.0) * denoise_range * 0.285714
+    else:
+        offset = -((5.0 - float(level)) / 4.0) * denoise_range * 0.257143
+    return _clamp(base + offset, 0.0001, 1.0)
+
+
+def largest_face_denoise(
+    boxes: list[tuple[int, int, int, int]],
+    image_size: tuple[int, int],
+    *,
+    base_denoise: float = 0.10,
+    adaptive_ratio: float = 0.03,
+    denoise_min: float = 0.05,
+    denoise_max: float = 0.28,
+) -> float:
+    """Scale face denoise from the largest face area, matching reference mode."""
+    width, height = image_size
+    image_area = max(1.0, float(width) * float(height))
+    largest_ratio = 0.0
+    for x1, y1, x2, y2 in boxes:
+        box_area = max(0.0, float(x2 - x1)) * max(0.0, float(y2 - y1))
+        largest_ratio = max(largest_ratio, box_area / image_area)
+    if largest_ratio <= 0.0:
+        return _clamp(base_denoise, denoise_min, denoise_max)
+    scaled = float(base_denoise) * (largest_ratio / max(1e-6, float(adaptive_ratio)))
+    return _clamp(scaled, denoise_min, denoise_max)
+
+
+def _target_size(width: int, height: int) -> tuple[int, int]:
+    """Floor image dimensions to the /16 latent grid without changing aspect."""
+    return max(16, (width // 16) * 16), max(16, (height // 16) * 16)
+
+
+def _resize_to_target(image: Image.Image) -> Image.Image:
+    """Resize pixels to the exact latent-grid dimensions passed to DiffSynth."""
+    target = _target_size(image.width, image.height)
+    if image.size == target:
+        return image
+    return image.resize(target, Image.Resampling.LANCZOS)
+
+
+def build_canny_control_image(image: Image.Image) -> Image.Image:
+    """Build the three-channel Canny conditioning image used by stage 1."""
+    import cv2
+
+    rgb = np.asarray(image.convert("RGB"))
+    gray = cv2.cvtColor(rgb, cv2.COLOR_RGB2GRAY)
+    edges = cv2.Canny(gray, _CANNY_LOW, _CANNY_HIGH)
+    return Image.fromarray(np.repeat(edges[:, :, None], 3, axis=2))
+
+
+def build_global_kwargs(
+    image: Image.Image,
+    *,
+    strength: float,
+    seed: int | None,
+    controlnet_input: Any,
+) -> dict[str, Any]:
+    """Build the DiffSynth Qwen call shape without importing the ML runtime."""
+    input_image = _resize_to_target(image)
+    width, height = input_image.size
+    return {
+        "prompt": _GLOBAL_PROMPT,
+        "negative_prompt": _GLOBAL_NEGATIVE,
+        "cfg_scale": GLOBAL_CFG,
+        "input_image": input_image,
+        "denoising_strength": float(strength),
+        "height": height,
+        "width": width,
+        "seed": seed,
+        "rand_device": "cpu",
+        "num_inference_steps": GLOBAL_STEPS,
+        # The source graph applies ModelSamplingAuraFlow with shift=3. DiffSynth
+        # expresses the same rational sigma shift as exp(mu), so mu=log(3).
+        "exponential_shift_mu": math.log(3.0),
+        "blockwise_controlnet_inputs": [controlnet_input],
+    }
+
+
+def build_face_kwargs(crop: Image.Image, *, strength: float, seed: int | None) -> dict[str, Any]:
+    """Build the DiffSynth Z-Image face-detail call shape."""
+    input_image = _resize_to_target(crop)
+    width, height = input_image.size
+    return {
+        "prompt": _FACE_PROMPT,
+        "negative_prompt": _FACE_NEGATIVE,
+        "cfg_scale": FACE_CFG,
+        "input_image": input_image,
+        "denoising_strength": float(strength),
+        "height": height,
+        "width": width,
+        "seed": seed,
+        "rand_device": "cpu",
+        "num_inference_steps": FACE_STEPS,
+    }
+
+
+def composite_face(base: np.ndarray, detail: np.ndarray, mask: np.ndarray, *, feather: int = 10) -> np.ndarray:
+    """Feather ``detail`` into ``base`` while preserving every zero-mask pixel."""
+    import cv2
+
+    if base.shape != detail.shape:
+        raise ValueError("base and detail must have identical shapes")
+    if mask.shape != base.shape[:2]:
+        raise ValueError("mask must match the image height and width")
+
+    alpha = mask.astype(np.float32) / 255.0
+    if feather > 0:
+        sigma = max(0.1, float(feather) / 3.0)
+        alpha = cv2.GaussianBlur(alpha, (0, 0), sigmaX=sigma, sigmaY=sigma)
+        # Blurring may introduce tiny values far outside the intended mask. Keep an
+        # explicit support dilation so pixels beyond the feather radius stay exact.
+        support = cv2.dilate((mask > 0).astype(np.uint8), np.ones((2 * feather + 1, 2 * feather + 1), np.uint8))
+        alpha *= support
+    alpha = np.clip(alpha, 0.0, 1.0)[:, :, None]
+    merged = base.astype(np.float32) * (1.0 - alpha) + detail.astype(np.float32) * alpha
+    return np.clip(np.rint(merged), 0, 255).astype(np.uint8)
+
+
+def _expanded_box(
+    box: tuple[int, int, int, int],
+    image_size: tuple[int, int],
+    *,
+    factor: float = 2.5,
+) -> tuple[int, int, int, int]:
+    """Expand a face box around its center, matching the reference crop factor."""
+    x1, y1, x2, y2 = box
+    image_width, image_height = image_size
+    center_x = (x1 + x2) / 2.0
+    center_y = (y1 + y2) / 2.0
+    width = max(1.0, (x2 - x1) * factor)
+    height = max(1.0, (y2 - y1) * factor)
+    return (
+        max(0, round(center_x - width / 2.0)),
+        max(0, round(center_y - height / 2.0)),
+        min(image_width, round(center_x + width / 2.0)),
+        min(image_height, round(center_y + height / 2.0)),
+    )
+
+
+def _model_cache_dir() -> Path:
+    root = os.environ.get("XDG_CACHE_HOME")
+    base = Path(root) if root else Path.home() / ".cache"
+    return base / "remove-ai-watermarks"
+
+
+def _yunet_model_path() -> Path:
+    """Download the small MIT-licensed YuNet ONNX model on first use."""
+    model_path = _model_cache_dir() / YUNET_MODEL_NAME
+    if model_path.exists() and hashlib.sha256(model_path.read_bytes()).hexdigest() == YUNET_MODEL_SHA256:
+        return model_path
+
+    model_path.parent.mkdir(parents=True, exist_ok=True)
+    log.info("Downloading YuNet face detector: %s", YUNET_MODEL_URL)
+    request = urllib.request.Request(
+        YUNET_MODEL_URL,
+        headers={"User-Agent": "remove-ai-watermarks"},
+    )
+    with urllib.request.urlopen(request, timeout=60) as response:  # noqa: S310 - fixed HTTPS source
+        payload = response.read()
+        log.info(
+            "YuNet download response: status=%s content_length=%s",
+            getattr(response, "status", None),
+            len(payload),
+        )
+    digest = hashlib.sha256(payload).hexdigest()
+    if digest != YUNET_MODEL_SHA256:
+        raise OSError(
+            "YuNet download failed integrity verification: "
+            f"expected {YUNET_MODEL_SHA256}, got {digest} ({len(payload)} bytes)"
+        )
+    with tempfile.NamedTemporaryFile(dir=model_path.parent, suffix=".onnx", delete=False) as handle:
+        handle.write(payload)
+        temporary = Path(handle.name)
+    temporary.replace(model_path)
+    return model_path
+
+
+def _nms_boxes(
+    boxes: list[tuple[int, int, int, int]],
+    scores: list[float],
+    *,
+    threshold: float = 0.3,
+) -> list[tuple[int, int, int, int]]:
+    """Apply OpenCV NMS to detections collected at multiple image scales."""
+    if not boxes:
+        return []
+    import cv2
+
+    xywh = [[x1, y1, x2 - x1, y2 - y1] for x1, y1, x2, y2 in boxes]
+    indices = cv2.dnn.NMSBoxes(xywh, scores, score_threshold=0.2, nms_threshold=threshold)
+    if len(indices) == 0:
+        return []
+    return [boxes[int(index)] for index in np.asarray(indices).reshape(-1)]
+
+
+def detect_faces(image: Image.Image) -> list[tuple[int, int, int, int]]:
+    """Detect face boxes with YuNet at two scales for large and small faces."""
+    import cv2
+
+    rgb = np.asarray(image.convert("RGB"))
+    original_height, original_width = rgb.shape[:2]
+    detections: list[tuple[int, int, int, int]] = []
+    scores: list[float] = []
+    model_path = _yunet_model_path()
+
+    for long_side in (640, 1280):
+        scale = min(1.0, long_side / max(original_width, original_height))
+        width = max(1, round(original_width * scale))
+        height = max(1, round(original_height * scale))
+        resized = cv2.resize(rgb, (width, height), interpolation=cv2.INTER_AREA) if scale < 1.0 else rgb
+        bgr = cv2.cvtColor(resized, cv2.COLOR_RGB2BGR)
+        detector = cv2.FaceDetectorYN.create(
+            str(model_path),
+            "",
+            (width, height),
+            YUNET_SCORE_THRESHOLD,
+            0.3,
+            5000,
+        )
+        _, rows = detector.detect(bgr)
+        if rows is None:
+            continue
+        inverse = 1.0 / scale
+        for row in rows:
+            x, y, box_width, box_height = (float(value) for value in row[:4])
+            x1 = max(0, round(x * inverse))
+            y1 = max(0, round(y * inverse))
+            x2 = min(original_width, round((x + box_width) * inverse))
+            y2 = min(original_height, round((y + box_height) * inverse))
+            if x2 <= x1 or y2 <= y1:
+                continue
+            detections.append((x1, y1, x2, y2))
+            scores.append(float(row[-1]))
+        if scale == 1.0:
+            break
+    return _nms_boxes(detections, scores)
+
+
+def _ellipse_masks(
+    boxes: list[tuple[int, int, int, int]],
+    image_size: tuple[int, int],
+) -> list[np.ndarray]:
+    """Safe fallback masks when SAM is unavailable."""
+    import cv2
+
+    width, height = image_size
+    masks: list[np.ndarray] = []
+    for x1, y1, x2, y2 in boxes:
+        mask = np.zeros((height, width), dtype=np.uint8)
+        center = ((x1 + x2) // 2, (y1 + y2) // 2)
+        axes = (max(1, int((x2 - x1) * 0.55)), max(1, int((y2 - y1) * 0.62)))
+        cv2.ellipse(mask, center, axes, 0, 0, 360, 255, -1)
+        masks.append(mask)
+    return masks
+
+
+def _prepare_sam_inputs(inputs: Any, device: str, dtype: Any) -> Any:
+    """Move SAM inputs to the target device without casting geometric prompts."""
+    prepared = inputs.to(device)
+    if "pixel_values" in prepared:
+        prepared["pixel_values"] = prepared["pixel_values"].to(dtype=dtype)
+    return prepared
+
+
+def _sam_point_prompts(
+    boxes: list[tuple[int, int, int, int]],
+) -> tuple[list[list[list[list[float]]]], list[list[list[int]]]]:
+    """Build Impact Pack's center-1 positive prompt for every face box."""
+    points = [[[[(x1 + x2) / 2.0, (y1 + y2) / 2.0]] for x1, y1, x2, y2 in boxes]]
+    labels = [[[1] for _box in boxes]]
+    return points, labels
+
+
+def _clip_sam_masks_to_boxes(
+    masks: list[np.ndarray],
+    boxes: list[tuple[int, int, int, int]],
+    image_size: tuple[int, int],
+) -> list[np.ndarray]:
+    """Match Impact Pack by intersecting each SAM mask with its detector box."""
+    width, height = image_size
+    clipped: list[np.ndarray] = []
+    for mask, (x1, y1, x2, y2) in zip(masks, boxes, strict=True):
+        box_mask = np.zeros((height, width), dtype=np.uint8)
+        box_mask[max(0, y1) : min(height, y2), max(0, x1) : min(width, x2)] = 255
+        clipped.append(np.bitwise_and(mask.astype(np.uint8), box_mask))
+    return clipped
+
+
+def _select_sam_masks(
+    masks: np.ndarray,
+    scores: np.ndarray,
+    *,
+    threshold: float = 0.93,
+) -> list[np.ndarray]:
+    """Select and combine SAM proposals like Impact Pack's ``sub_threshold``."""
+    mask_array = np.asarray(masks)
+    score_array = np.asarray(scores)
+    if mask_array.ndim == 5 and mask_array.shape[0] == 1:
+        mask_array = mask_array[0]
+    if score_array.ndim == 3 and score_array.shape[0] == 1:
+        score_array = score_array[0]
+    if mask_array.ndim == 3:
+        mask_array = mask_array[:, None, :, :]
+    if score_array.ndim == 1:
+        score_array = score_array[:, None]
+    if mask_array.ndim != 4 or score_array.ndim != 2:
+        raise ValueError("SAM masks and scores have unexpected dimensions")
+    if mask_array.shape[:2] != score_array.shape:
+        raise ValueError("SAM mask proposals and IoU scores do not align")
+
+    selected_masks: list[np.ndarray] = []
+    for candidates, candidate_scores in zip(mask_array, score_array, strict=True):
+        selected = np.flatnonzero(candidate_scores >= threshold)
+        if selected.size == 0:
+            selected = np.asarray([int(np.argmax(candidate_scores))])
+        combined = np.any(candidates[selected] > 0, axis=0)
+        selected_masks.append(combined.astype(np.uint8) * 255)
+    return selected_masks
+
+
+def _sam_outputs_to_numpy(masks: Any, scores: Any) -> tuple[np.ndarray, np.ndarray]:
+    """Convert SAM tensors through float32 because NumPy rejects bfloat16."""
+    mask_array = masks.detach().float().cpu().numpy()
+    score_array = scores.detach().float().cpu().numpy()
+    return mask_array, score_array
+
+
+@dataclass
+class QwenZImagePipeline:
+    """Lazy runtime for the two-stage Qwen/Z-Image profile."""
+
+    device: str
+    torch_dtype: Any
+    hf_token: str | None = None
+    progress_callback: Callable[[str], None] | None = None
+    controlnet_conditioning_scale: float = GLOBAL_CONTROLNET_SCALE
+    keep_face_models_on_device: bool | None = None
+    cache_prompt_embeddings: bool = True
+
+    def __post_init__(self) -> None:
+        self._qwen_pipe: Any = None
+        self._zimage_pipe: Any = None
+        self._sam_model: Any = None
+        self._sam_processor: Any = None
+
+    def _progress(self, message: str) -> None:
+        if self.progress_callback is not None:
+            with contextlib.suppress(Exception):
+                self.progress_callback(message)
+
+    def _require_cuda(self) -> None:
+        if self.device != "cuda":
+            raise RuntimeError(
+                "The qwen-zimage pipeline is CUDA-only. Its Qwen-Image-2512 and "
+                "Z-Image models do not fit the supported MPS path."
+            )
+
+    def _vram_limit(self) -> float | None:
+        import torch
+
+        with contextlib.suppress(Exception):
+            return max(1.0, torch.cuda.mem_get_info("cuda")[1] / (1024**3) - 0.5)
+        return None
+
+    def _keep_face_models_resident(self) -> bool:
+        import torch
+
+        total_memory_gib = 0.0
+        with contextlib.suppress(Exception):
+            total_memory_gib = torch.cuda.get_device_properties("cuda").total_memory / (1024**3)
+        return resolve_face_model_residency(
+            self.keep_face_models_on_device,
+            total_memory_gib=total_memory_gib,
+        )
+
+    @staticmethod
+    def _qwen_vram_config() -> dict[str, Any]:
+        import torch
+
+        return {
+            "offload_dtype": "disk",
+            "offload_device": "disk",
+            "onload_dtype": torch.float8_e4m3fn,
+            "onload_device": "cpu",
+            "preparing_dtype": torch.float8_e4m3fn,
+            "preparing_device": "cuda",
+            "computation_dtype": torch.bfloat16,
+            "computation_device": "cuda",
+        }
+
+    @staticmethod
+    def _zimage_vram_config() -> dict[str, Any]:
+        import torch
+
+        return {
+            "offload_dtype": torch.bfloat16,
+            "offload_device": "cpu",
+            "onload_dtype": torch.bfloat16,
+            "onload_device": "cpu",
+            "preparing_dtype": torch.bfloat16,
+            "preparing_device": "cuda",
+            "computation_dtype": torch.bfloat16,
+            "computation_device": "cuda",
+        }
+
+    def _load_qwen(self) -> Any:
+        if self._qwen_pipe is not None:
+            return self._qwen_pipe
+        self._require_cuda()
+        os.environ.setdefault("DIFFSYNTH_DOWNLOAD_SOURCE", "huggingface")
+        if self.hf_token:
+            os.environ.setdefault("HF_TOKEN", self.hf_token)
+        try:
+            from diffsynth.pipelines.qwen_image import ControlNetInput, ModelConfig, QwenImagePipeline
+        except ImportError as exc:
+            raise ImportError(
+                "The qwen-zimage pipeline needs the optional dependency group. "
+                "Install: pip install 'remove-ai-watermarks[qwen-zimage]'"
+            ) from exc
+
+        self._progress("Loading Qwen-Image-2512, Lightning LoRA, and Canny ControlNet...")
+        config = self._qwen_vram_config()
+        model_configs = [
+            ModelConfig(
+                model_id=QWEN_IMAGE_2512_MODEL_ID,
+                origin_file_pattern="transformer/diffusion_pytorch_model*.safetensors",
+                **config,
+            ),
+            ModelConfig(
+                model_id=QWEN_IMAGE_2512_MODEL_ID,
+                origin_file_pattern="text_encoder/model*.safetensors",
+                **config,
+            ),
+            ModelConfig(
+                model_id=QWEN_IMAGE_2512_MODEL_ID,
+                origin_file_pattern="vae/diffusion_pytorch_model.safetensors",
+                **config,
+            ),
+            ModelConfig(
+                model_id=QWEN_CANNY_CONTROLNET_MODEL_ID,
+                origin_file_pattern="model.safetensors",
+                **config,
+            ),
+        ]
+        pipe = QwenImagePipeline.from_pretrained(
+            torch_dtype=self.torch_dtype,
+            device=self.device,
+            model_configs=model_configs,
+            tokenizer_config=ModelConfig(
+                model_id=QWEN_IMAGE_2512_MODEL_ID,
+                origin_file_pattern="tokenizer/",
+            ),
+            vram_limit=self._vram_limit(),
+        )
+        from diffsynth.diffusion import FlowMatchScheduler
+
+        # Avoid the base Qwen scheduler's terminal rescale for the distilled LoRA.
+        # With exponential_shift_mu=log(3), this is the closest DiffSynth equivalent
+        # of the source graph's four-step sgm_uniform + AuraFlow shift 3.
+        pipe.scheduler = FlowMatchScheduler("Qwen-Image-Lightning")
+        lightning = ModelConfig(
+            model_id=QWEN_LIGHTNING_MODEL_ID,
+            origin_file_pattern=QWEN_LIGHTNING_PATTERN,
+        )
+        pipe.load_lora(pipe.dit, lightning, alpha=0.8)
+        if self.cache_prompt_embeddings:
+            _cache_static_prompt_embeddings(
+                pipe,
+                ("prompt_emb", "prompt_emb_mask"),
+            )
+        self._qwen_pipe = (pipe, ControlNetInput)
+        return self._qwen_pipe
+
+    def _load_zimage(self) -> Any:
+        if self._zimage_pipe is not None:
+            return self._zimage_pipe
+        self._require_cuda()
+        os.environ.setdefault("DIFFSYNTH_DOWNLOAD_SOURCE", "huggingface")
+        if self.hf_token:
+            os.environ.setdefault("HF_TOKEN", self.hf_token)
+        try:
+            from diffsynth.pipelines.z_image import ModelConfig, ZImagePipeline
+        except ImportError as exc:
+            raise ImportError(
+                "The qwen-zimage pipeline needs the optional dependency group. "
+                "Install: pip install 'remove-ai-watermarks[qwen-zimage]'"
+            ) from exc
+
+        self._progress("Loading Z-Image Turbo face-detail model...")
+        keep_on_device = self._keep_face_models_resident()
+        config = self._zimage_vram_config()
+        pipe = ZImagePipeline.from_pretrained(
+            torch_dtype=self.torch_dtype,
+            device=self.device,
+            model_configs=[
+                ModelConfig(
+                    model_id=ZIMAGE_TURBO_MODEL_ID,
+                    origin_file_pattern="transformer/*.safetensors",
+                    **config,
+                ),
+                ModelConfig(
+                    model_id=ZIMAGE_TURBO_MODEL_ID,
+                    origin_file_pattern="text_encoder/*.safetensors",
+                    **config,
+                ),
+                ModelConfig(
+                    model_id=ZIMAGE_TURBO_MODEL_ID,
+                    origin_file_pattern="vae/diffusion_pytorch_model.safetensors",
+                    **config,
+                ),
+            ],
+            tokenizer_config=ModelConfig(
+                model_id=ZIMAGE_TURBO_MODEL_ID,
+                origin_file_pattern="tokenizer/",
+            ),
+            vram_limit=self._vram_limit(),
+        )
+        if keep_on_device:
+            _pin_vram_managed_models(pipe)
+        if self.cache_prompt_embeddings:
+            _cache_static_prompt_embeddings(pipe, ("prompt_embeds",))
+        self._zimage_pipe = pipe
+        return pipe
+
+    def _load_sam(self) -> tuple[Any, Any]:
+        if self._sam_model is not None and self._sam_processor is not None:
+            return self._sam_model, self._sam_processor
+        self._progress("Loading SAM face-mask model...")
+        try:
+            from transformers import AutoModelForMaskGeneration, AutoProcessor
+        except ImportError as exc:
+            raise ImportError("SAM needs transformers and torchvision from the qwen-zimage extra.") from exc
+        kwargs: dict[str, Any] = {}
+        if self.hf_token:
+            kwargs["token"] = self.hf_token
+        processor = AutoProcessor.from_pretrained(SAM_MODEL_ID, **kwargs)
+        model = AutoModelForMaskGeneration.from_pretrained(
+            SAM_MODEL_ID,
+            torch_dtype=self.torch_dtype,
+            **kwargs,
+        ).to(self.device)
+        model.eval()
+        self._sam_model = model
+        self._sam_processor = processor
+        return model, processor
+
+    def _sam_masks(
+        self,
+        image: Image.Image,
+        boxes: list[tuple[int, int, int, int]],
+    ) -> list[np.ndarray]:
+        if not boxes:
+            return []
+        import torch
+
+        try:
+            model, processor = self._load_sam()
+            input_points, input_labels = _sam_point_prompts(boxes)
+            inputs = processor(
+                images=image,
+                input_boxes=[[list(box) for box in boxes]],
+                input_points=input_points,
+                input_labels=input_labels,
+                return_tensors="pt",
+            )
+            original_sizes = inputs["original_sizes"].clone()
+            reshaped_sizes = inputs["reshaped_input_sizes"].clone()
+            inputs = _prepare_sam_inputs(inputs, self.device, self.torch_dtype)
+            with torch.inference_mode():
+                outputs = model(**inputs, multimask_output=True)
+            processed = processor.post_process_masks(
+                outputs.pred_masks.detach().cpu(),
+                original_sizes,
+                reshaped_sizes,
+            )[0]
+            mask_array, score_array = _sam_outputs_to_numpy(processed, outputs.iou_scores)
+            binary_masks = _select_sam_masks(
+                mask_array,
+                score_array,
+            )
+            return _clip_sam_masks_to_boxes(binary_masks, boxes, image.size)
+        except Exception as exc:
+            log.warning("SAM face-mask refinement failed (%s); using box-derived ellipse masks", exc)
+            return _ellipse_masks(boxes, image.size)
+
+    def preload(self) -> None:
+        """Eagerly load both diffusion stages and the face segmentation model."""
+        self._load_qwen()
+        self._load_zimage()
+        self._load_sam()
+        _yunet_model_path()
+
+    def _run_global(self, image: Image.Image, strength: float, seed: int | None) -> Image.Image:
+        pipe, controlnet_input_cls = self._load_qwen()
+        input_image = _resize_to_target(image)
+        control = build_canny_control_image(input_image)
+        control_input = controlnet_input_cls(
+            image=control,
+            scale=float(self.controlnet_conditioning_scale),
+        )
+        self._progress(f"Running Qwen-Image-2512 Canny pass: strength={strength:.4f}, steps={GLOBAL_STEPS}...")
+        result = pipe(
+            **build_global_kwargs(
+                input_image,
+                strength=strength,
+                seed=seed,
+                controlnet_input=control_input,
+            )
+        )
+        if result.size != image.size:
+            result = result.resize(image.size, Image.Resampling.LANCZOS)
+        return result.convert("RGB")
+
+    @staticmethod
+    def _detail_size(
+        crop_size: tuple[int, int],
+        face_size: tuple[int, int],
+    ) -> tuple[int, int]:
+        """Scale a crop toward a 768px face guide while capping it at 1024px."""
+        crop_width, crop_height = crop_size
+        face_width, face_height = face_size
+        scale_for_face = 768.0 / max(1, max(face_width, face_height))
+        scale_for_crop = 1024.0 / max(1, max(crop_width, crop_height))
+        scale = min(scale_for_face, scale_for_crop)
+        # Never shrink below the crop's current size unless the 1024 cap requires it.
+        if max(crop_width, crop_height) <= 1024:
+            scale = max(1.0, scale)
+        width = max(16, round(crop_width * scale / 16.0) * 16)
+        height = max(16, round(crop_height * scale / 16.0) * 16)
+        return width, height
+
+    def _run_faces(
+        self,
+        original: Image.Image,
+        global_result: Image.Image,
+        boxes: list[tuple[int, int, int, int]],
+        masks: list[np.ndarray],
+        *,
+        strength: float,
+        seed: int | None,
+    ) -> Image.Image:
+        if not boxes:
+            return global_result
+        pipe = self._load_zimage()
+        base = np.asarray(global_result.convert("RGB")).copy()
+        source = np.asarray(original.convert("RGB"))
+        detail_seed = None if seed is None else seed + 1
+
+        for index, (box, mask) in enumerate(zip(boxes, masks, strict=True), start=1):
+            crop_box = _expanded_box(box, original.size)
+            cx1, cy1, cx2, cy2 = crop_box
+            crop_source = source[cy1:cy2, cx1:cx2]
+            crop_mask = mask[cy1:cy2, cx1:cx2]
+            if crop_source.size == 0 or not np.any(crop_mask):
+                continue
+
+            face_width = box[2] - box[0]
+            face_height = box[3] - box[1]
+            process_size = self._detail_size((cx2 - cx1, cy2 - cy1), (face_width, face_height))
+            crop_image = Image.fromarray(crop_source).resize(process_size, Image.Resampling.LANCZOS)
+            self._progress(
+                f"Regenerating face {index}/{len(boxes)} with Z-Image: strength={strength:.4f}, steps={FACE_STEPS}..."
+            )
+            detailed = pipe(**build_face_kwargs(crop_image, strength=strength, seed=detail_seed))
+            detailed = detailed.convert("RGB").resize((cx2 - cx1, cy2 - cy1), Image.Resampling.LANCZOS)
+
+            base_crop = base[cy1:cy2, cx1:cx2]
+            base[cy1:cy2, cx1:cx2] = composite_face(
+                base_crop,
+                np.asarray(detailed),
+                crop_mask,
+                feather=10,
+            )
+        return Image.fromarray(base)
+
+    def run(
+        self,
+        image: Image.Image,
+        *,
+        strength: float | None,
+        seed: int | None,
+        tile: bool = False,
+        tile_size: int = 1024,
+        tile_overlap: int = 128,
+    ) -> Image.Image:
+        """Execute global regeneration and masked face repair."""
+        self._require_cuda()
+        seed = resolve_seed(seed, "qwen-zimage")
+        global_strength = (
+            resolution_adaptive_denoise(image.width, image.height) if strength is None else float(strength)
+        )
+        if tile and max(image.size) > tile_size:
+            from remove_ai_watermarks.noai.tiling import run_tiled
+
+            global_result = run_tiled(
+                lambda tile_image: self._run_global(tile_image, global_strength, seed),
+                image,
+                tile_size,
+                tile_overlap,
+                self._progress,
+            )
+        else:
+            global_result = self._run_global(image, global_strength, seed)
+
+        self._progress("Detecting faces on the original image...")
+        boxes = detect_faces(image)
+        if not boxes:
+            self._progress("No faces detected; keeping the Qwen global result.")
+            return global_result
+        masks = self._sam_masks(image, boxes)
+        face_strength = largest_face_denoise(boxes, image.size)
+        return self._run_faces(
+            image,
+            global_result,
+            boxes,
+            masks,
+            strength=face_strength,
+            seed=seed,
+        )

src/remove_ai_watermarks/noai/watermark_profiles.py

@@ -31,7 +31,11 @@ QWEN_MODEL_ID = "Qwen/Qwen-Image"
 # profile is ``sdxl``; ``default`` is kept as an accepted alias (it was the profile's
 # name before ``controlnet`` became the default-selected pipeline, 2026-06-09).
 SDXL_PROFILE = "sdxl"
-_PROFILE_ALIASES = {"default": SDXL_PROFILE}
+QWEN_ZIMAGE_PROFILE = "qwen-zimage"
+_PROFILE_ALIASES = {
+    "default": SDXL_PROFILE,
+    "qwen_zimage": QWEN_ZIMAGE_PROFILE,
+}
 
 
 def normalize_profile(profile: str) -> str:
@@ -40,6 +44,24 @@ def normalize_profile(profile: str) -> str:
     return _PROFILE_ALIASES.get(normalized, normalized)
 
 
+def resolve_steps(num_inference_steps: int | None, pipeline: str) -> int:
+    """Resolve a profile-specific step default while preserving explicit values.
+
+    The Lightning LoRA in ``qwen-zimage`` is distilled for four steps. Existing
+    SDXL and Qwen profiles keep the long-standing 50-step CLI default.
+    """
+    if num_inference_steps is not None:
+        return num_inference_steps
+    return 4 if normalize_profile(pipeline) == QWEN_ZIMAGE_PROFILE else 50
+
+
+def resolve_seed(seed: int | None, pipeline: str) -> int | None:
+    """Keep the oracle-verified qwen-zimage profile deterministic by default."""
+    if seed is not None:
+        return seed
+    return 0 if normalize_profile(pipeline) == QWEN_ZIMAGE_PROFILE else None
+
+
 # The SDXL-native canny ControlNet used by the ``controlnet`` pipeline. The
 # ControlNet is an add-on to the SDXL base checkpoint (DEFAULT_MODEL_ID), not a
 # separate base model, so both the ``sdxl`` and ``controlnet`` profiles load the
@@ -113,7 +135,8 @@ def strength_default_help() -> str:
     """
     return (
         f"vendor-adaptive (OpenAI {OPENAI_STRENGTH} / Google {GEMINI_STRENGTH} / "
-        f"unknown {UNKNOWN_STRENGTH}, from the C2PA issuer; same ladder for both pipelines)"
+        f"unknown {UNKNOWN_STRENGTH}, from the C2PA issuer; qwen-zimage instead uses "
+        "resolution-adaptive denoise)"
     )
 
 

src/remove_ai_watermarks/noai/watermark_remover.py

@@ -1,8 +1,11 @@
 """Watermark removal using diffusion model regeneration attack.
 
-Three pipelines (selected by the explicit ``pipeline`` ctor arg):
+Four pipelines (selected by the explicit ``pipeline`` ctor arg):
 
-0. ``qwen`` -- Qwen-Image (20B MMDiT, Apache-2.0) img2img. The scrub still comes from
+0. ``qwen-zimage`` -- Qwen-Image-2512 Lightning + DiffSynth Canny regenerates the
+   frame, then SAM-masked Z-Image Turbo regenerates original face crops and feathers
+   them into the global result. CUDA-only and installed through its own optional extra.
+1. ``qwen`` -- Qwen-Image (20B MMDiT, Apache-2.0) img2img. The scrub still comes from
    the img2img ``strength``; Qwen preserves text (incl. CJK) and structure markedly
    better than SDXL at the scrub floor, so it over-regenerates real photos far less.
    CUDA/cloud-class (does not fit MPS). See ``watermark_profiles`` for the certified
@@ -11,14 +14,14 @@ Three pipelines (selected by the explicit ``pipeline`` ctor arg):
    is the release gate, not sweeping seeds.)
 
 Two SDXL pipelines:
-1. ``controlnet`` (DEFAULT) -- SDXL img2img with a canny ControlNet. The watermark
+2. ``controlnet`` (DEFAULT) -- SDXL img2img with a canny ControlNet. The watermark
    REMOVAL still comes from the img2img regeneration (``strength``); the ControlNet
    only PRESERVES structure (text/faces) by conditioning on the edge map. No original
    pixels are ever copied or frozen. Because the edge map keeps the regeneration
    closer to the original, it needs a higher ``strength`` floor than ``default`` to
    destroy SynthID (the certified controlnet ladder; see ``watermark_profiles``).
    ``controlnet_conditioning_scale`` is the preservation knob.
-2. ``default`` -- plain SDXL img2img. Partial-noise regeneration scrubs the
+3. ``default`` -- plain SDXL img2img. Partial-noise regeneration scrubs the
    invisible watermark; ``strength`` controls how much is regenerated. Lighter (no
    ControlNet weights), but at the low default strength it leaves SynthID on
    flat-graphic content -- use it for inputs without text/faces.
@@ -47,7 +50,10 @@ from remove_ai_watermarks.noai.watermark_profiles import (
     DEFAULT_MODEL_ID,
     DEFAULT_STRENGTH,
     QWEN_MODEL_ID,
+    QWEN_ZIMAGE_PROFILE,
     normalize_profile,
+    resolve_seed,
+    resolve_steps,
     resolve_strength,
     viable_steps,
 )
@@ -388,6 +394,13 @@ class WatermarkRemover:
         # the legacy "default" alias resolves to "sdxl".
         self.model_profile = normalize_profile(pipeline)
         self.controlnet_conditioning_scale = controlnet_conditioning_scale
+        if self.model_profile == QWEN_ZIMAGE_PROFILE and self.model_id != self.DEFAULT_MODEL_ID:
+            raise ValueError(
+                "The qwen-zimage pipeline uses a fixed Qwen-Image-2512 + Z-Image model stack; "
+                "--model is not supported for this profile."
+            )
+        if self.model_profile == QWEN_ZIMAGE_PROFILE:
+            self.model_id = "Qwen/Qwen-Image-2512 + Tongyi-MAI/Z-Image-Turbo"
 
         if not is_watermark_removal_available():
             _ensure_watermark_deps()
@@ -399,8 +412,8 @@ class WatermarkRemover:
         if torch_dtype is None:
             if self.device == "cpu" or self.device == "mps":
                 self.torch_dtype = torch.float32  # type: ignore
-            elif self.model_profile == "qwen":
-                # Qwen-Image is published in bf16; fp16 risks overflow on the 20B MMDiT.
+            elif self.model_profile in {"qwen", QWEN_ZIMAGE_PROFILE}:
+                # Qwen-Image and Z-Image are published in bf16; fp16 risks overflow.
                 # cuda/xpu-only by construction: the cpu/mps guard above already forced
                 # fp32, and the 20B model does not fit MPS anyway.
                 self.torch_dtype = torch.bfloat16  # type: ignore
@@ -412,6 +425,7 @@ class WatermarkRemover:
         self._pipeline: AutoImg2ImgPipeline | None = None
         self._controlnet_pipeline: Any = None
         self._qwen_pipeline: Any = None
+        self._qwen_zimage_pipeline: Any = None
         self._progress_callback = progress_callback
         self.hf_token: str | None = hf_token or os.environ.get("HF_TOKEN")
 
@@ -426,7 +440,9 @@ class WatermarkRemover:
 
     def preload(self) -> None:
         """Eagerly load the pipeline so download progress bars are visible."""
-        if self.model_profile == "qwen":
+        if self.model_profile == QWEN_ZIMAGE_PROFILE:
+            self._load_qwen_zimage_pipeline().preload()
+        elif self.model_profile == "qwen":
             self._load_qwen_pipeline()
         elif self.model_profile == "controlnet":
             self._load_controlnet_pipeline()
@@ -611,6 +627,20 @@ class WatermarkRemover:
 
         return self._qwen_pipeline
 
+    def _load_qwen_zimage_pipeline(self) -> Any:
+        """Load the two-stage Qwen-Image-2512 + Z-Image runtime lazily."""
+        if self._qwen_zimage_pipeline is None:
+            from remove_ai_watermarks.noai.qwen_zimage_pipeline import QwenZImagePipeline
+
+            self._qwen_zimage_pipeline = QwenZImagePipeline(
+                device=self.device,
+                torch_dtype=self.torch_dtype,
+                hf_token=self.hf_token,
+                progress_callback=self._progress_callback,
+                controlnet_conditioning_scale=self.controlnet_conditioning_scale,
+            )
+        return self._qwen_zimage_pipeline
+
     # ── Core removal ─────────────────────────────────────────────────
 
     def remove_watermark(
@@ -618,7 +648,7 @@ class WatermarkRemover:
         image_path: Path,
         output_path: Path | None = None,
         strength: float | None = None,
-        num_inference_steps: int = 50,
+        num_inference_steps: int | None = None,
         guidance_scale: float | None = None,
         seed: int | None = None,
         vendor: str | None = None,
@@ -637,7 +667,8 @@ class WatermarkRemover:
                 default (see ``vendor``).
             num_inference_steps: Number of denoising steps.
             guidance_scale: Classifier-free guidance scale.
-            seed: Random seed for reproducibility.
+            seed: Random seed for reproducibility. None resolves to 0 for
+                qwen-zimage and stays random for the other profiles.
             vendor: SynthID vendor (``"openai"`` / ``"google"`` / None) used to pick the
                 default strength when ``strength`` is None. Detect it from the ORIGINAL
                 input with ``watermark_profiles.vendor_for_strength`` before processing
@@ -647,7 +678,7 @@ class WatermarkRemover:
                 The lossless alternative to a ``--max-resolution`` downscale for large
                 inputs that OOM on MPS/GPU (issue #10). Only engages when the long side
                 exceeds ``tile_size``; smaller images run a single pass unchanged.
-            tile_size: Tile dimension in px (default 1024, SDXL's training size).
+            tile_size: Tile dimension in px (default 1024).
             tile_overlap: Overlap between adjacent tiles in px (default 128), feather-
                 blended so there is no visible seam.
             region: Restrict the regeneration to the AI-composited box ``(x, y, w, h)``
@@ -673,19 +704,27 @@ class WatermarkRemover:
         if output_path is None:
             output_path = image_path
 
-        strength = resolve_strength(strength, vendor, self.model_profile)
-
-        if not 0.0 <= strength <= 1.0:
-            raise ValueError(f"Strength must be between 0.0 and 1.0, got {strength}")
-
-        if guidance_scale is None:
-            guidance_scale = 7.5
-
         self._set_progress("Loading and preprocessing input image...")
         init_image = Image.open(image_path).convert("RGB")
         w, h = init_image.size
         self._set_progress(f"Image loaded: {w}x{h}px | Model: {self.model_id}")
 
+        if self.model_profile == QWEN_ZIMAGE_PROFILE:
+            from remove_ai_watermarks.noai.qwen_zimage_pipeline import resolution_adaptive_denoise
+
+            strength = strength if strength is not None else resolution_adaptive_denoise(w, h)
+        else:
+            strength = resolve_strength(strength, vendor, self.model_profile)
+        seed = resolve_seed(seed, self.model_profile)
+        if not 0.0 <= strength <= 1.0:
+            raise ValueError(f"Strength must be between 0.0 and 1.0, got {strength}")
+
+        num_inference_steps = resolve_steps(num_inference_steps, self.model_profile)
+        if guidance_scale is None:
+            guidance_scale = 1.0 if self.model_profile == QWEN_ZIMAGE_PROFILE else 7.5
+        elif self.model_profile == QWEN_ZIMAGE_PROFILE and guidance_scale != 1.0:
+            raise ValueError("The qwen-zimage profile fixes both diffusion stages at CFG 1.0.")
+
         generator = None
         if seed is not None and _HAS_TORCH:
             self._set_progress(f"Setting reproducible seed: {seed}")
@@ -695,17 +734,26 @@ class WatermarkRemover:
         # inside attention with an opaque reshape error, so raise it to the minimum that
         # denoises. Must be applied to the value HANDED TO THE PIPELINE -- the old
         # max(1, ...) below only clamped the number in the log line.
-        adjusted = viable_steps(num_inference_steps, strength)
-        if adjusted != num_inference_steps:
-            logger.warning(
-                "steps=%s at strength=%s denoises 0 steps and would crash; using steps=%s (1 effective)",
-                num_inference_steps,
-                strength,
-                adjusted,
-            )
-            num_inference_steps = adjusted
+        if self.model_profile != QWEN_ZIMAGE_PROFILE:
+            adjusted = viable_steps(num_inference_steps, strength)
+            if adjusted != num_inference_steps:
+                logger.warning(
+                    "steps=%s at strength=%s denoises 0 steps and would crash; using steps=%s (1 effective)",
+                    num_inference_steps,
+                    strength,
+                    adjusted,
+                )
+                num_inference_steps = adjusted
+        elif num_inference_steps != 4:
+            raise ValueError("The qwen-zimage profile uses the 4-step Lightning LoRA, so --steps must be 4.")
 
-        effective_steps = max(1, int(num_inference_steps * strength))
+        # DiffSynth keeps all timesteps and compresses their sigma range for a low
+        # denoise value. Diffusers instead truncates the schedule by strength.
+        effective_steps = (
+            num_inference_steps
+            if self.model_profile == QWEN_ZIMAGE_PROFILE
+            else max(1, int(num_inference_steps * strength))
+        )
         self._set_progress(
             f"Config: strength={strength}, steps={num_inference_steps} "
             f"(~{effective_steps} effective), guidance={guidance_scale}, device={self.device}"
@@ -721,6 +769,17 @@ class WatermarkRemover:
             return self._run_img2img(img, strength, num_inference_steps, guidance_scale, generator)
 
         def _generate() -> Image.Image:
+            # qwen-zimage owns its global-only tiling because its face stage must run
+            # once after the tiles are blended. Other profiles tile their whole pass.
+            if self.model_profile == QWEN_ZIMAGE_PROFILE:
+                return self._run_qwen_zimage(
+                    init_image,
+                    strength,
+                    seed,
+                    tile=tile,
+                    tile_size=tile_size,
+                    tile_overlap=tile_overlap,
+                )
             # Tile only when asked AND the image is larger than one tile; otherwise a
             # single full-image pass (tiling a sub-tile image is pure overhead).
             if tile and max(init_image.size) > tile_size:
@@ -742,6 +801,7 @@ class WatermarkRemover:
             self.torch_dtype = torch.float32
             self._pipeline = None
             self._controlnet_pipeline = None
+            self._qwen_zimage_pipeline = None
             cleaned_image = _generate()
 
         # Region-targeted regeneration for AI-enhanced composites: keep the real photo
@@ -941,6 +1001,27 @@ class WatermarkRemover:
         result = pipeline(**kwargs)
         return result.images[0]
 
+    def _run_qwen_zimage(
+        self,
+        init_image: Image.Image,
+        strength: float,
+        seed: int | None,
+        *,
+        tile: bool = False,
+        tile_size: int = 1024,
+        tile_overlap: int = 128,
+    ) -> Image.Image:
+        """Run the Qwen 2512 Canny pass and masked Z-Image face repair."""
+        pipeline = self._load_qwen_zimage_pipeline()
+        return pipeline.run(
+            init_image,
+            strength=strength,
+            seed=seed,
+            tile=tile,
+            tile_size=tile_size,
+            tile_overlap=tile_overlap,
+        )
+
     # ── Batch ────────────────────────────────────────────────────────
 
     def remove_watermark_batch(
@@ -948,7 +1029,7 @@ class WatermarkRemover:
         input_dir: Path,
         output_dir: Path,
         strength: float | None = None,
-        num_inference_steps: int = 50,
+        num_inference_steps: int | None = None,
         extensions: tuple[str, ...] = (".png", ".jpg", ".jpeg", ".webp"),
     ) -> list[Path]:
         """Remove watermarks from all images in a directory."""