Fix DyPE high-resolution noise by simplifying NTK scaling#8847
Open
Pfannkuchensack wants to merge 16 commits intoinvoke-ai:mainfrom
Open
Fix DyPE high-resolution noise by simplifying NTK scaling#8847Pfannkuchensack wants to merge 16 commits intoinvoke-ai:mainfrom
Pfannkuchensack wants to merge 16 commits intoinvoke-ai:mainfrom
Conversation
- Replace complex YaRN 3-band frequency blending with direct NTK interpolation - Update mscale formula: 1.0 + 0.1 * log(s) / sqrt(s) - Add k_t calculation: dype_scale * (sigma ^ dype_exponent) - Interpolate ntk_factor from 1.0 (late steps) to base_ntk (early steps) - Make mscale timestep-dependent for smoother transitions - Add helper functions: find_correction_factor, find_correction_range, linear_ramp_mask - Update tests for new implementation
Pfannkuchensack
requested review from
JPPhoto,
blessedcoolant,
dunkeroni and
lstein
as code owners
github-actions
bot
added
python
JPPhoto
requested changes
Feb 6, 2026
Collaborator
JPPhoto
left a comment
JPPhoto
left a comment
There was a problem hiding this comment.
I don't think there's anything wrong with the code, and results are definitely sharper overall. But I'm still seeing "chroma noise" (for lack of a better word) and wider scenes are now left-biased as a result of these changes. Any insight?
Aligns InvokeAI's DyPE implementation with ComfyUI-DyPE to fix chroma
noise and left-biased composition on wide images.
Key changes:
- Skip DyPE scaling on axis 0 (time/channel), only scale spatial axes
- Implement YaRN 3-band frequency blending (base/linear/NTK) with
DyPE-modulated beta/gamma correction masks
- Use per-axis linear_scale with global ntk_scale for non-square images
- Fix FLUX_BASE_PE_LEN (256 -> 64) to match actual spatial positions
- Add timestep-dependent mscale matching ComfyUI's _get_mscale
- Use floor/ceil in find_correction_range matching reference impl
…sack/InvokeAI into fix/dype-ntk-scaling
The previous implementation used a simplified NTK-only approach that destroyed positional information at high resolutions. This restores proper YaRN 3-band blending (base/linear/NTK) with correct mask inversion and timestep-dependent mscale, matching ComfyUI-DyPE's get_1d_dype_yarn_pos_embed implementation.
Collaborator
|
@lstein Recommend retargeting for 6.13.x since this needs more of @Pfannkuchensack's time. |
Collaborator
Done. |
- Cap FLUX noise schedule shift (mu) at base resolution when DyPE is active. Previously mu extrapolated linearly with resolution (e.g. mu=3.39 at 2496x1776 vs correct 1.15), causing the model to spend nearly all denoising at high noise levels, producing "crumpled paper" texture artifacts.
Collaborator
|
I recommend we abandon this and switch to #8980 to re-align DyPE implementation more closely with the original paper and accept the loss of high-frequency detail. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fix DyPE high-resolution noise by simplifying NTK scaling
Summary
1.0 + 0.1 * log(s) / sqrt(s)dype_scale * (sigma ^ dype_exponent)find_correction_factor,find_correction_range,linear_ramp_maskRelated Issues / Discussions
Fixes noise/grain artifacts in high-resolution (2K+) image generation with DyPE enabled.
QA Instructions
pytest tests/backend/flux/dype/test_dype.py -vMerge Plan
No special merge considerations. Changes are isolated to the DyPE module.
Checklist
What's Newcopy (if doing a release after this PR)