Skip to content

Support res_multistep_ancestral on RF and fix its discretization #9444

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 3 commits into
base: master
Choose a base branch
from
+16 −8
Open

Support res_multistep_ancestral on RF and fix its discretization #9444

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4ce6a1d
fix res_multistep and its ancestral.
Balladie Aug 20, 2025
8c6c3e5
remove unused var
Balladie Aug 20, 2025
e82f886
restore first stochastic step.
Balladie Aug 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
24 changes: 16 additions & 8 deletions comfy/k_diffusion/sampling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1342,12 +1342,15 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
seed = extra_args.get("seed", None)
noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
t_fn = lambda sigma: sigma.log().neg()

model_sampling = model.inner_model.model_patcher.get_model_object('model_sampling')
lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)

phi1_fn = lambda t: torch.expm1(t) / t
phi2_fn = lambda t: (phi1_fn(t) - 1.0) / t

old_sigma_down = None
old_sigma_next = None
old_denoised = None
uncond_denoised = None
def post_cfg_function(args):
Expand Down Expand Up @@ -1375,29 +1378,34 @@ def post_cfg_function(args):
x = x + d * dt
else:
# Second order multistep method in https://arxiv.org/pdf/2308.02157
t, t_old, t_next, t_prev = t_fn(sigmas[i]), t_fn(old_sigma_down), t_fn(sigma_down), t_fn(sigmas[i - 1])
t, t_old, t_next, t_prev = lambda_fn(sigmas[i]), lambda_fn(old_sigma_next), lambda_fn(sigmas[i + 1]), lambda_fn(sigmas[i - 1])
h = t_next - t
h_eta = h * (eta + 1)
c2 = (t_prev - t_old) / h

phi1_val, phi2_val = phi1_fn(-h), phi2_fn(-h)
alpha_next = sigmas[i + 1] * t_next.exp()

phi1_val, phi2_val = phi1_fn(-h_eta), phi2_fn(-h_eta)
b1 = torch.nan_to_num(phi1_val - phi2_val / c2, nan=0.0)
b2 = torch.nan_to_num(phi2_val / c2, nan=0.0)

if cfg_pp:
x = x + (denoised - uncond_denoised)
x = sigma_fn(h) * x + h * (b1 * uncond_denoised + b2 * old_denoised)
x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_next * h_eta * (b1 * uncond_denoised + b2 * old_denoised)
else:
x = sigma_fn(h) * x + h * (b1 * denoised + b2 * old_denoised)
x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_next * h_eta * (b1 * denoised + b2 * old_denoised)

# Noise addition
if sigmas[i + 1] > 0:
if old_denoised is not None:
sigma_up = sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt()
x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up

if cfg_pp:
old_denoised = uncond_denoised
else:
old_denoised = denoised
old_sigma_down = sigma_down
old_sigma_next = sigmas[i + 1]
return x

@torch.no_grad()
Expand Down
Loading