[Pfannkuchensack]

Summary

Qwen-Image-2.1's base model set its noise-schedule shift (mu) from the image's token count along a line through its scheduler config: 256 tokens give 0.5, and 8192 tokens give 0.9 (about 1448×1448 of area, 2.1 MP). 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	0.66, unchanged
2048²	1.31	0.90	0.77 → 0.70
4096²	3.79	0.90	0.95 → 0.70

calculate_mu now holds mu at the end of its fit. Qwen-Image (#9892), Krea-2 Raw (#9887), FLUX.1 and Z-Image already do the same.

• Turbo samples its own fixed table, so it is unaffected.
• The qwen_image_2_1_denoise node goes from 1.0.0 to 1.0.1; its inputs are unchanged.
• The docstrings of calculate_mu and build_sigmas say the schedule matches the pipeline only up to the end of the fit.

Why: at 4096², an image-to-image pass at strength 0.5 started from almost pure noise, so it replaced the image instead of refining it.

How the reference implementations handle this:

• diffusers QwenImage21Pipeline extrapolates like the old code.
• ComfyUI samples Qwen-Image-2.1 at a constant mu of 0.69 (supported_models.py, QwenImage21.sampling_settings).

Related Issues / Discussions

Follow-up to #9784. The same fix for Qwen-Image is #9892, and for Krea-2 Raw it is #9887.

QA Instructions

Automated

• pytest tests/backend/qwen_image_2_1 tests/app/invocations/test_qwen_image_2_1_nodes.py: 57 passed.
  • Pipeline parity: the parity test against QwenImage21Pipeline recordings now keeps only cases within the fit. The 2048² recording was the extrapolated schedule, so it was removed.
  • Above the fit: new cases check that 8192, 16384 and 65536 tokens all give the 8192-token schedule. The expected head and tail sigmas were computed with diffusers' FlowMatchEulerDiscreteScheduler at mu 0.9, outside the code under test.
  • Below the end: 1440² (8100 tokens) still follows the line.
  • Mutation probes: each of these fails at least one test: removing the hold; holding at 8191, 8193, 4096 or 16384 tokens; a constant mu of 0.69.
• ruff@0.11.2 check / format --check: pass. openapi.json is regenerated; only the node version changes.

GPU A/B (RTX 4090, Qwen-Image-2.1 base Q4_K_M GGUF, Qwen3-VL 8B fp8 encoder, 40 steps, CFG 1, model cache emptied before each run). Main's code and this branch ran the same graphs:

Graph	Result	main / this PR
text-to-image 1024²	bit-identical	66 s / 53 s (first, cold run)
text-to-image 2048², seed 1234	Same composition, lettering, lighthouse and house; waves and rocks differ in detail	183 s / 170 s
text-to-image 2048², seed 99	Same scene; the sun sits on the horizon and the sky is crisper with the hold	203 s / 174 s
image-to-image 2048², strength 0.5	Both keep the source's composition and lettering; the held run changes a little less	104 s / 96 s

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

Review

No material findings. Resolved:

• The build_sigmas docstring still claimed exact parity with the pipeline; it now scopes that to within the fit.
• The docs now say that text-to-image above about 1448×1448 changes too.

Compatibility / Rollout

• Same seed, different image: base-model images above about 1448×1448 of area no longer reproduce with the same seed, and they differ from diffusers' QwenImage21Pipeline there.
  • The node has no Shift input, so the old schedule cannot be selected.
  • Image metadata does not record mu.
• Unchanged: Turbo, and base images at or below that area.
• Saved workflows: the 1.0.1 node has the same fields as 1.0.0. The editor offers to update 1.0.0 nodes, and running them is not blocked.

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.}