298977887/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-04-18 18:33:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

uni_pc now works with KSamplerAdvanced return_with_leftover_noise.

Browse Source
This commit is contained in:
comfyanonymous 2023-02-13 12:29:21 -05:00
parent 3fd6b7027c
commit e3451cea4f
1 changed files with 10 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
comfy/extra_samplers/uni_pc.py
View File

@ -821,7 +821,13 @@ def expand_dims(v, dims):
def sample_unipc(model, noise, image, sigmas, sampling_function, extra_args=None, callback=None, disable=None): def sample_unipc(model, noise, image, sigmas, sampling_function, extra_args=None, callback=None, disable=None):
timesteps = torch.nn.functional.interpolate(sigmas[None,None,:-1], size=(len(sigmas),), mode='linear')[0][0] to_zero = False
if sigmas[-1] == 0:
timesteps = torch.nn.functional.interpolate(sigmas[None,None,:-1], size=(len(sigmas),), mode='linear')[0][0]
to_zero = True
else:
timesteps = sigmas.clone()
for s in range(timesteps.shape[0]): for s in range(timesteps.shape[0]):
timesteps[s] = (model.sigma_to_t(timesteps[s]) / 1000) + (1 / len(model.sigmas)) timesteps[s] = (model.sigma_to_t(timesteps[s]) / 1000) + (1 / len(model.sigmas))
@ -832,7 +838,7 @@ def sample_unipc(model, noise, image, sigmas, sampling_function, extra_args=None
else: else:
img = noise img = noise
if sigmas[-1] == 0: if to_zero:
timesteps[-1] = (1 / len(model.sigmas)) timesteps[-1] = (1 / len(model.sigmas))
device = noise.device device = noise.device
@ -853,4 +859,6 @@ def sample_unipc(model, noise, image, sigmas, sampling_function, extra_args=None
uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False) uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False)
x = uni_pc.sample(img, timesteps=timesteps, skip_type="time_uniform", method="multistep", order=3, lower_order_final=True) x = uni_pc.sample(img, timesteps=timesteps, skip_type="time_uniform", method="multistep", order=3, lower_order_final=True)
if not to_zero:
x /= ns.marginal_alpha(timesteps[-1])
return x return x