Downscaled a Gemini original to 0.57 and 1.40 MP and ran the deployed worker on
both, so the profile applied its own low-end strengths of 0.0896 and 0.1066. Both
come back clean in the Gemini app.
That rules out the failure mode the ladder raised: 0.08 failed at 4.33 MP, and the
curve sends sub-1 MP images to 0.084-0.094, which looked like it might mean small
Gemini uploads were under-processed in production. They are not, at these sizes.
Written as validation of the shipped curve rather than as evidence that the
boundary moves with resolution. 0.0896 sits inside the untested gap at 4.33 MP,
where only 0.08 and 0.10 were probed, so it may clear at both sizes; the direction
of any resolution dependence stays unproven. These are also downscales rather than
natively small Gemini outputs, which remain untested.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The document asserted it in two places and recommended capping Gemini at 1536
with 0.30, or native-calibrating to ~0.35+. Nothing measured that, and the one
relevant measurement points the other way: the 2026-06-14 deployed-worker re-test
cleared Gemini at 0.15 on two NATIVE 2816x1536 images, the same rung as capped
1536. The document already recorded that as contradicting the "native >= 0.30"
guess, then kept the guess anyway in the historical-certification paragraph and
restated it as fact in the strength floors.
Replaced with what was measured, plus an explicit statement that the direction is
unproven and the low-resolution end has never been through the Gemini oracle on
any pipeline.
Also removes the same appeal from the qwen-zimage denoise-boundary note added
earlier in this branch, which had used the unproven trend as reassurance that the
bottom of the adaptive curve is safe. It is not reassurance; it is an open
question, and it is now written as one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A ladder on one native 2816x1536 Gemini original at seed 0, verified through the
Gemini app, puts the boundary between 0.08 and 0.10: 0.154, 0.12 and 0.10 read
clean, 0.08 reads SynthID FOUND. Fidelity rises monotonically all the way down,
so 0.10 buys +1.54 dB whole-image and +0.98 dB inside the face boxes over the
0.154 the profile ships for that size.
Recorded with the two constraints that stop it being acted on directly. It
brackets rather than calibrates - one image, one seed, and shipping the lowest
clean rung means shipping at the measured cliff edge. And the untested end is the
bottom, not the top: every Gemini oracle fixture is 2816x1536, so the Google-side
certification only ever covered 0.154, while the curve sends sub-1 MP images to
0.084-0.094. The resolution trend already recorded in this document says lower
processing resolution needs less strength, which is the shape the curve has, but
that is an inference and no small Gemini original has ever been through the
oracle. Downscaling is valid test material since SynthID survives it by design.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both qwen-zimage stages prompt with module constants, and at CFG 1.0 DiffSynth's
PipelineUnitRunner reuses the positive embedding for the negative side rather
than encoding it, so exactly one embedding per stage is ever computed. Persist it
and neither text encoder has to be loaded at all.
Measured on an H100 volume: this drops 15.45 GiB (Qwen2.5-VL) and 7.49 GiB
(Z-Image) of an 87.6 GiB per-request read, worth a median 11.76 s and 4.10 s of
load time paired within five containers. A nine-face fixture returned
sha256 c8567e11077de32a both with and without the cache, so the output is
byte-identical and the provider-oracle clearance is untouched.
The cache key carries the cache version, model id, pipeline output params and the
exact prompt, so a model bump or a prompt edit recomputes instead of reading a
stale embedding. The write is atomic because a torn file must never read back as
a hit, and a miss after the text encoder was already dropped raises rather than
calling a model that is not loaded.
_model_cache_dir now prefers HF_HOME: on a scale-to-zero runner that is the only
persistently mounted path, so anything below it is re-derived every request. The
YuNet download follows the same root.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Joins fix/qwen-vram-residency, which carried the global-stack residency
change on the v0.20.1 line that raiw-app pins. The change itself is already
on main via port/qwen-vram-residency-main, reapplied there because the
package layout moved under _internal/ in between, so this merge is history
only and its tree is identical to the commit before it.
Conflicts resolved in favour of main throughout, including dropping the
noai/watermark_remover.py the branch resurrected; that module now lives at
_internal/watermark_remover.py.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Port of the same change made on the v0.20.1 line, reapplied here because the
package layout moved under _internal/ in the meantime.
The mandatory Qwen stack was configured to offload to disk unconditionally.
DiffSynth implements that by dropping the weights to the meta device and
re-reading every parameter through its DiskMap on the next onload, and the
pipeline moves between text encoder, transformer and VAE on every pass, so
each generation paid a full model reload. That is the right trade on a
consumer card, where it is what makes a 20B model runnable at all, and pure
waste on a card that can simply hold the stack.
Residency is now resolved from total VRAM, mirroring how the optional
Z-Image face stack is already gated. Above the floor the config passes no
"disk" value anywhere, which is what actually disables the behavior:
DiffSynth latches disk_offload once from offload_dtype, so pointing every
device at CUDA while leaving the sentinel would keep both the meta-drop and
the re-read.
Measured on an H100: a warm global pass went from 37.3s at 0.8 GiB resident
to 2.2s at 28.7 GiB, with both stacks resident peaking at 48.0 GiB of 79.2.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The mandatory Qwen stack was configured to offload to disk unconditionally.
DiffSynth implements that by dropping the weights to the meta device and
re-reading every parameter through its DiskMap on the next onload, and the
pipeline moves between text encoder, transformer and VAE on every pass, so
each generation paid a full model reload. That is the right trade on a
consumer card, where it is what makes a 20B model runnable at all, and pure
waste on a card that can simply hold the stack.
Residency is now resolved from total VRAM, mirroring how the optional
Z-Image face stack is already gated. Above the floor the config passes no
"disk" value anywhere, which is what actually disables the behavior:
DiffSynth latches disk_offload once from offload_dtype, so pointing every
device at CUDA while leaving the sentinel would keep both the meta-drop and
the re-read.
The floor is set equal to the face floor rather than lower because that is
the configuration measured with both stacks resident; a tighter gate is
plausible but unvalidated. cpu_offload now forces both stacks to stream, so
a caller asking for low VRAM no longer gets the larger stack pinned anyway.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
