Fix LTX2 dynamic timestep shift ignoring the sequence length#14244
Open
dxqb wants to merge 1 commit into
Open
Conversation
`calculate_shift` was called with `max_image_seq_len` as its `image_seq_len` argument, so `mu` reduced to `max_shift` for every call regardless of the resolution or number of frames. Pass the actual packed video sequence length (`latent_num_frames * latent_height * latent_width`) instead, matching the Flux/SD3 pipelines and the previously commented-out intent, so `use_dynamic_shifting` varies with the sample size again. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
github-actions
Bot
added
fixes-issue
pipelines
size/S
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #14243.
LTX2Pipeline.__call__computes the dynamic timestep-shiftmuby passingmax_image_seq_lenas theimage_seq_lenargument ofcalculate_shift. Becausecalculate_shiftalso receivesmax_image_seq_lenas itsmax_seq_len, the two cancel andmucollapses tomax_shiftfor every call, regardless of resolution or frame count — so the resolution/length-dependent shift that the scheduler config requests (use_dynamic_shifting=True, distinctbase_shift/max_shift) never takes effect.This passes the actual packed sequence length (
latent_num_frames * latent_height * latent_width) instead, which is the value the commented-outvideo_sequence_lengthline already spelled out, and what the Flux/SD3 pipelines and the Lightricks/ComfyUI reference implementations use (math.prod(latent.shape[2:])).See #14243 for the derivation, the match against the reference implementation and the scheduler config, and the
git blameevidence that this appears to be an unfinished port rather than an intentional constant.Drafted by Claude.