Fix CogVideoX scheduler prev_timestep for non-leading spacing #13125
+40
−2
Conversation
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
When timestep_spacing is set to 'linspace' or 'trailing', the formula `timestep - num_train_timesteps // num_inference_steps` does not produce the correct previous timestep because timesteps are not uniformly spaced by that stride. This causes wrong alpha_prod_t_prev lookups and produces incorrect denoising results. Replace the hardcoded arithmetic with a `previous_timestep()` method that looks up the actual next entry in `self.timesteps`, matching the approach already used in DDPMScheduler.
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Outdated suggestions cannot be applied.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.
Summary
Bug:
CogVideoXDDIMScheduler.step()andCogVideoXDPMScheduler.step()computeprev_timestepusingtimestep - num_train_timesteps // num_inference_steps, which only works correctly whentimestep_spacing="leading". For"linspace"or"trailing"spacing, timesteps are not uniformly spaced by that stride, so the formula produces incorrect previous timestep values and wrongalpha_prod_t_prevlookups, leading to degraded or incorrect denoising.Fix: Replace the hardcoded arithmetic with a
previous_timestep()method that looks up the actual next entry inself.timesteps, matching the approach already used byDDPMScheduler.previous_timestep().Files Changed
src/diffusers/schedulers/scheduling_ddim_cogvideox.py— useprevious_timestep()instep(), add methodsrc/diffusers/schedulers/scheduling_dpm_cogvideox.py— useprevious_timestep()instep(), add methodTest plan
CogVideoXDDIMSchedulerwithtimestep_spacing="trailing"and confirmprev_timestepvalues now match the actual timestep scheduleCogVideoXDPMSchedulerwithtimestep_spacing="linspace"and confirm correct behavior