The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Only append zero to noise schedule if last sigma isn't zero.
comfyanonymous95fa9545f167
diff --git a/comfy/samplers.py b/comfy/samplers.pyindex 0ab08a4cd..3f7633814 100644--- a/comfy/samplers.py+++ b/comfy/samplers.py@@ -313,13 +313,18 @@ def simple_scheduler(model_sampling, steps):def ddim_scheduler(model_sampling, steps):s = model_samplingsigs = []- ss = max(len(s.sigmas) // steps, 1)x = 1+ if math.isclose(float(s.sigmas[x]), 0, abs_tol=0.00001):+ steps += 1+ sigs = []+ else:+ sigs = [0.0]++ ss = max(len(s.sigmas) // steps, 1)while x < len(s.sigmas):sigs += [float(s.sigmas[x])]x += sssigs = sigs[::-1]- sigs += [0.0]return torch.FloatTensor(sigs)def normal_scheduler(model_sampling, steps, sgm=False, floor=False):@@ -327,16 +332,23 @@ def normal_scheduler(model_sampling, steps, sgm=False, floor=False):start = s.timestep(s.sigma_max)end = s.timestep(s.sigma_min)+ append_zero = Trueif sgm:timesteps = torch.linspace(start, end, steps + 1)[:-1]else:+ if math.isclose(float(s.sigma(end)), 0, abs_tol=0.00001):+ steps += 1+ append_zero = Falsetimesteps = torch.linspace(start, end, steps)sigs = []for x in range(len(timesteps)):ts = timesteps[x]- sigs.append(s.sigma(ts))- sigs += [0.0]+ sigs.append(float(s.sigma(ts)))++ if append_zero:+ sigs += [0.0]+return torch.FloatTensor(sigs)# Implemented based on: https://arxiv.org/abs/2407.12173