Skip to content

Fix custom sigmas for supported schedulers #1510

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

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix custom sigmas for supported schedulers #1510

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
425035d
Fix custom sigmas for supported schedulers
cmeka Oct 21, 2025
f653544
Merge branch 'kijai:main' into custom-sigmas
cmeka Oct 21, 2025
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
55 changes: 38 additions & 17 deletions wanvideo/schedulers/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,23 +31,29 @@
"rcm"
]

def _apply_custom_sigmas(sample_scheduler, sigmas, device):
sample_scheduler.sigmas = sigmas.to(device)
sample_scheduler.timesteps = (sample_scheduler.sigmas[:-1] * 1000).to(torch.int64).to(device)
sample_scheduler.num_inference_steps = len(sample_scheduler.timesteps)

def get_scheduler(scheduler, steps, start_step, end_step, shift, device, transformer_dim=5120, flowedit_args=None, denoise_strength=1.0, sigmas=None, log_timesteps=False, **kwargs):
timesteps = None
if 'unipc' in scheduler:
sample_scheduler = FlowUniPCMultistepScheduler(shift=shift)
if sigmas is None:
sample_scheduler.set_timesteps(steps, device=device, shift=shift, use_beta_sigmas=('beta' in scheduler))
else:
sample_scheduler.sigmas = sigmas.to(device)
sample_scheduler.timesteps = (sample_scheduler.sigmas[:-1] * 1000).to(torch.int64).to(device)
sample_scheduler.num_inference_steps = len(sample_scheduler.timesteps)
_apply_custom_sigmas(sample_scheduler, sigmas, device)

elif scheduler in ['euler/beta', 'euler']:
sample_scheduler = FlowMatchEulerDiscreteScheduler(shift=shift, use_beta_sigmas=(scheduler == 'euler/beta'))
if flowedit_args: #seems to work better
timesteps, _ = retrieve_timesteps(sample_scheduler, device=device, sigmas=get_sampling_sigmas(steps, shift))
if sigmas is None:
if flowedit_args: #seems to work better
timesteps, _ = retrieve_timesteps(sample_scheduler, device=device, sigmas=get_sampling_sigmas(steps, shift))
else:
sample_scheduler.set_timesteps(steps, device=device)
else:
sample_scheduler.set_timesteps(steps, device=device, sigmas=sigmas[:-1].tolist() if sigmas is not None else None)
_apply_custom_sigmas(sample_scheduler, sigmas, device)
elif 'dpm' in scheduler:
if 'sde' in scheduler:
algorithm_type = "sde-dpmsolver++"
Expand All @@ -57,16 +63,20 @@ def get_scheduler(scheduler, steps, start_step, end_step, shift, device, transfo
if sigmas is None:
sample_scheduler.set_timesteps(steps, device=device, use_beta_sigmas=('beta' in scheduler))
else:
sample_scheduler.sigmas = sigmas.to(device)
sample_scheduler.timesteps = (sample_scheduler.sigmas[:-1] * 1000).to(torch.int64).to(device)
sample_scheduler.num_inference_steps = len(sample_scheduler.timesteps)
_apply_custom_sigmas(sample_scheduler, sigmas, device)
elif scheduler == 'deis':
sample_scheduler = DEISMultistepScheduler(use_flow_sigmas=True, prediction_type="flow_prediction", flow_shift=shift)
sample_scheduler.set_timesteps(steps, device=device)
sample_scheduler.sigmas[-1] = 1e-6
if sigmas is None:
sample_scheduler.set_timesteps(steps, device=device)
sample_scheduler.sigmas[-1] = 1e-6
else:
_apply_custom_sigmas(sample_scheduler, sigmas, device)
elif 'lcm' in scheduler:
sample_scheduler = FlowMatchLCMScheduler(shift=shift, use_beta_sigmas=(scheduler == 'lcm/beta'))
sample_scheduler.set_timesteps(steps, device=device, sigmas=sigmas[:-1].tolist() if sigmas is not None else None)
if sigmas is None:
sample_scheduler.set_timesteps(steps, device=device)
else:
_apply_custom_sigmas(sample_scheduler, sigmas, device)
elif 'flowmatch_causvid' in scheduler:
if sigmas is not None:
raise NotImplementedError("This scheduler does not support custom sigmas")
Expand Down Expand Up @@ -99,17 +109,28 @@ def get_scheduler(scheduler, steps, start_step, end_step, shift, device, transfo
sample_scheduler.sigmas = torch.cat([sample_scheduler.timesteps / 1000, torch.tensor([0.0], device=device)])
elif 'flowmatch_pusa' in scheduler:
sample_scheduler = FlowMatchSchedulerPusa(shift=shift, sigma_min=0.0, extra_one_step=True)
sample_scheduler.set_timesteps(steps+1, denoising_strength=denoise_strength, shift=shift,
sigmas=sigmas[:-1].tolist() if sigmas is not None else None)
if sigmas is None:
sample_scheduler.set_timesteps(steps+1, denoising_strength=denoise_strength, shift=shift)
else:
_apply_custom_sigmas(sample_scheduler, sigmas, device)
elif scheduler == 'res_multistep':
sample_scheduler = FlowMatchSchedulerResMultistep(shift=shift)
sample_scheduler.set_timesteps(steps, denoising_strength=denoise_strength, sigmas=sigmas[:-1].tolist() if sigmas is not None else None)
if sigmas is None:
sample_scheduler.set_timesteps(steps, denoising_strength=denoise_strength)
else:
_apply_custom_sigmas(sample_scheduler, sigmas, device)
elif "sa_ode_stable" in scheduler:
sample_scheduler = FlowMatchSAODEStableScheduler(shift=shift, **kwargs)
sample_scheduler.set_timesteps(steps, device=device, sigmas=sigmas[:-1].tolist() if sigmas is not None else None)
if sigmas is None:
sample_scheduler.set_timesteps(steps, device=device)
else:
_apply_custom_sigmas(sample_scheduler, sigmas, device)
elif 'rcm' in scheduler:
sample_scheduler = rCMFlowMatchScheduler()
sample_scheduler.set_timesteps(steps, sigma_max=120)
if sigmas is None:
sample_scheduler.set_timesteps(steps, sigma_max=120)
else:
_apply_custom_sigmas(sample_scheduler, sigmas, device)

if timesteps is None:
timesteps = sample_scheduler.timesteps
Expand Down