Skip to content

Update LCMScheduler Inference Timesteps to be More Evenly Spaced #5836

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Nov 20, 2023
Merged

Update LCMScheduler Inference Timesteps to be More Evenly Spaced #5836

Changes from 3 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
ac6c537
Change LCMScheduler.set_timesteps to pick more evenly spaced inferenc…
dg845 Nov 17, 2023
322f049
Change inference_indices implementation to better match previous beha…
dg845 Nov 18, 2023
103a4d1
Merge branch 'main' into lcm-scheduler-fix-timestep-schedule
patil-suraj Nov 20, 2023
8e8ac32
Add num_inference_steps=26 test case to test_inference_steps.
dg845 Nov 20, 2023
b17824c
run CI
patil-suraj Nov 20, 2023
eedfa94
Merge branch 'lcm-scheduler-fix-timestep-schedule' of https://github.…
patil-suraj Nov 20, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 11 additions & 6 deletions src/diffusers/schedulers/scheduling_lcm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -373,10 +373,11 @@ def set_timesteps(
)

# LCM Timesteps Setting
# Currently, only linear spacing is supported.
c = self.config.num_train_timesteps // original_steps
# LCM Training Steps Schedule
lcm_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * c - 1
# The skipping step parameter k from the paper.
k = self.config.num_train_timesteps // original_steps
# LCM Training/Distillation Steps Schedule
# Currently, only a linearly-spaced schedule is supported (same as in the LCM distillation scripts).
lcm_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * k - 1
skipping_step = len(lcm_origin_timesteps) // num_inference_steps

if skipping_step < 1:
Expand All @@ -385,9 +386,13 @@ def set_timesteps(
)

# LCM Inference Steps Schedule
timesteps = lcm_origin_timesteps[::-skipping_step][:num_inference_steps]
lcm_origin_timesteps = lcm_origin_timesteps[::-1].copy()
# Select (approximately) evenly spaced indices from lcm_origin_timesteps.
inference_indices = np.linspace(0, len(lcm_origin_timesteps), num=num_inference_steps, endpoint=False)
inference_indices = np.floor(inference_indices).astype(np.int64)
timesteps = lcm_origin_timesteps[inference_indices]

self.timesteps = torch.from_numpy(timesteps.copy()).to(device=device, dtype=torch.long)
self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.long)

self._step_index = None

Expand Down