Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

LTX2Pipeline: dynamic timestep shift `mu` is constant, ignores the actual sequence length

Open Beginner friendly
#14,243 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
2/5
Estimated time
1-3 hours
Newbie friendliness
68/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Quiet
Tech stack
python
Domain
machine-learning

Research direction

Start in src/diffusers/pipelines/ltx2/pipeline_ltx2.py, inside LTX2Pipeline.call, and read the nearby latent dimension calculations and calculate_shift call. Confirm how the packed sequence length is derived and compare the analogous Flux/SD3 pipeline logic. Done means dynamic shifting uses the actual latent sequence length and varies with resolution or frame count.

Written by the indexing model from the issue text.

Description

bug pipelines

In LTX2Pipeline.__call__ (src/diffusers/pipelines/ltx2/pipeline_ltx2.py), mu is computed with max_image_seq_len passed as the first argument of calculate_shift — the image_seq_len parameter:

mu = calculate_shift(
    self.scheduler.config.get("max_image_seq_len", 4096),   # <-- this is image_seq_len; should be the sample's real sequence length
    self.scheduler.config.get("base_image_seq_len", 1024),
    self.scheduler.config.get("max_image_seq_len", 4096),
    self.scheduler.config.get("base_shift", 0.95),
    self.scheduler.config.get("max_shift", 2.05),
)

calculate_shift(image_seq_len, base, max, base_shift, max_shift) returns image_seq_len * m + b. With image_seq_len == max_seq_len, that reduces to max_shift, so mu is constant regardless of resolution or frame count — use_dynamic_shifting=True has no effect.

image_seq_len should be the sample's packed sequence length, latent_num_frames * latent_height * latent_width — which is already computed a few lines above (and even present as the commented line # video_sequence_length = latent_num_frames * latent_height * latent_width). This matches the Flux/SD3 pipelines and the LTX reference (math.prod(latent.shape[2:])).

Fix: pass latent_num_frames * latent_height * latent_width as the first argument.

Dominant language
Python
Stars
34.6k
Forks
7.4k
Avg merge
3d 7h
Merged PRs (30d)
73

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from huggingface/diffusers

All issues in huggingface/diffusers

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.