fix(qwen-image): hold the dynamic shift at the end of its fit

[Pfannkuchensack]

Summary

Qwen-Image sets its noise-schedule shift (mu) from the image's token count. It draws a line through two points from the scheduler config: 256 tokens give 0.5, and 8192 tokens (about 1448×1448 of area, 2.1 MP) give 0.9. Above 8192 tokens the line kept climbing.

Image	mu before	mu after	40-step img2img at strength 0.5 starts at sigma
1024²	0.69	0.69	unchanged
2048²	1.31	0.90	0.77 → 0.70
4096²	3.79	0.90	0.95 → 0.70

qwen_image_denoise now holds mu at the end of the fit, as the FLUX.1, Z-Image and LTX-2 fits already do (Krea-2 Raw: #9887).

• The end of the fit comes from the checkpoint's own scheduler/scheduler_config.json.
• An explicit shift (Lightning LoRAs) still wins.
• Edit models share the node. Their reference tokens never counted toward mu, and still don't.

Why now: at 4096², an image-to-image pass at strength 0.5 starts from almost pure noise (sigma 0.95), so it replaces the image instead of refining it. The Upscale widget's Qwen-Image entry, which comes next, needs mid-range starts at 2048²–4096².

How other implementations handle this:

• ComfyUI samples Qwen-Image at a constant mu of 1.15 (supported_models.py, QwenImage.sampling_settings).
• diffusers QwenImagePipeline extrapolates like the old code.

Other changes:

• The fit values are read as attributes of scheduler.config instead of .get(key, default). The scheduler registers every init argument, so those defaults never applied, and the max_shift default (0.9) did not match the scheduler's real one (1.15). The comment now names the real defaults.
• The node goes from 1.0.0 to 1.0.1; its inputs are unchanged. The shift description now says where the hold starts.

Related Issues / Discussions

Same fix as #9887 (Krea-2 Raw) and the Z-Image shift hold. It is a prerequisite for Qwen-Image in the Upscale widget (follow-up to #9825 and #9889).

Qwen-Image-2.1 (#9784) has its own calculate_mu with the same fit, and it still extrapolates. Not touched here.

QA Instructions

Automated

• pytest tests/app/invocations/test_qwen_image_denoise.py tests/app/invocations/test_qwen_image_packed_models.py: 34 passed.
  • The new tests run _run_diffusion with the real FlowMatchEulerDiscreteScheduler and read mu where it reaches set_timesteps. Expected values are hand-computed.
  • Stock config: 1024², 2048×1008 (8064 tokens, below the hold), 2048×1024, 2048² and 4096².
  • A checkpoint's own scheduler_config.json holds at its own end.
  • An explicit shift gives mu = ln(shift) at 4096².
  • Each of these breaks at least one test:
    • removing the hold;
    • hard-coding the hold at 8192 or at 0.9;
    • an off-by-one at the boundary;
    • counting tokens from the width only;
    • skipping the scheduler_config.json path.
• ruff@0.11.2 check / format --check: pass. openapi.json and schema.ts are regenerated; only the field description and the version change.

Text-to-image A/B (RTX 4090, Qwen Image 2512 Q4_K_M GGUF, Qwen2.5-VL fp8 encoder, 2048², 40 steps, CFG 4, model cache emptied before each run, held and old mu interleaved per seed; the old mu is passed through shift = e^1.3129 = 3.717):

Seed	Result	Time per image
1234	Same composition, lettering, lighthouse and house; the waves differ in detail	413 s held / 409 s old
99	Same lighthouse and lettering; the keeper's house changes shape and position	411 s / 409 s
7	Near-identical	411 s / 411 s

No image shows doubled subjects, broken lettering or a damaged composition.

Review

No material findings. Resolved: the docs now say that Shift on this node is e^mu, not mu as on Krea-2, and which value gives back the old schedule.

Remaining: the low end still extrapolates below 256 tokens, as the FLUX.1, Z-Image and Krea-2 fits do.

Compatibility / Rollout

• Same seed, different image: Qwen-Image and Qwen-Image-Edit images above about 1448×1448 of area (8192 tokens) no longer reproduce with the same seed. Image metadata does not record mu.
  • To get the old schedule back, set Shift on Denoise - Qwen Image in a workflow to e^(old mu): about 3.72 at 2048×2048, about 44 at 4096×4096.
  • Generate and Canvas don't send Shift.
• Unchanged: images at or below that area, and any explicit shift.
• Saved workflows: the 1.0.1 node has the same fields as 1.0.0, so older workflows still load.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}