Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Set the seed in the SDE samplers to make them more reproducible.

comfyanonymous4eab00e14bc0
diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py
index 26930428f..65d061997 100644
--- a/comfy/k_diffusion/sampling.py
+++ b/comfy/k_diffusion/sampling.py
@@ -77,7 +77,7 @@ class BatchedBrownianTree:
except TypeError:
seed = [seed]
self.batched = False
- self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
+ self.trees = [torchsde.BrownianTree(t0.cpu(), w0.cpu(), t1.cpu(), entropy=s, **kwargs) for s in seed]
@staticmethod
def sort(a, b):
@@ -85,7 +85,7 @@ class BatchedBrownianTree:
def __call__(self, t0, t1):
t0, t1, sign = self.sort(t0, t1)
- w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
+ w = torch.stack([tree(t0.cpu().float(), t1.cpu().float()).to(t0.dtype).to(t0.device) for tree in self.trees]) * (self.sign * sign)
return w if self.batched else w[0]
@@ -543,7 +543,8 @@ def sample_dpmpp_2s_ancestral(model, x, sigmas, extra_args=None, callback=None,
def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
"""DPM-Solver++ (stochastic)."""
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
- noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max) if noise_sampler is None else noise_sampler
+ seed = extra_args.get("seed", None)
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed) if noise_sampler is None else noise_sampler
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
sigma_fn = lambda t: t.neg().exp()
@@ -613,8 +614,9 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
if solver_type not in {'heun', 'midpoint'}:
raise ValueError('solver_type must be \'heun\' or \'midpoint\'')
+ seed = extra_args.get("seed", None)
sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
- noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max) if noise_sampler is None else noise_sampler
+ noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed) if noise_sampler is None else noise_sampler
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
diff --git a/comfy/sample.py b/comfy/sample.py
index 284efca61..dde5e42f8 100644
--- a/comfy/sample.py
+++ b/comfy/sample.py
@@ -65,7 +65,7 @@ def cleanup_additional_models(models):
for m in models:
m.cleanup()
-def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False, noise_mask=None, sigmas=None, callback=None, disable_pbar=False):
+def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False, noise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None):
device = comfy.model_management.get_torch_device()
if noise_mask is not None:
@@ -85,7 +85,7 @@ def sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative
sampler = comfy.samplers.KSampler(real_model, steps=steps, device=device, sampler=sampler_name, scheduler=scheduler, denoise=denoise, model_options=model.model_options)
- samples = sampler.sample(noise, positive_copy, negative_copy, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar)
+ samples = sampler.sample(noise, positive_copy, negative_copy, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
samples = samples.cpu()
cleanup_additional_models(models)
diff --git a/comfy/samplers.py b/comfy/samplers.py
index d6a8f609a..3aaf8ac4e 100644
--- a/comfy/samplers.py
+++ b/comfy/samplers.py
@@ -13,7 +13,7 @@ def lcm(a, b): #TODO: eventually replace by math.lcm (added in python3.9)
#The main sampling function shared by all the samplers
#Returns predicted noise
-def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, cond_concat=None, model_options={}):
+def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, cond_concat=None, model_options={}, seed=None):
def get_area_and_mult(cond, x_in, cond_concat_in, timestep_in):
area = (x_in.shape[2], x_in.shape[3], 0, 0)
strength = 1.0
@@ -292,8 +292,8 @@ class CFGNoisePredictor(torch.nn.Module):
super().__init__()
self.inner_model = model
self.alphas_cumprod = model.alphas_cumprod
- def apply_model(self, x, timestep, cond, uncond, cond_scale, cond_concat=None, model_options={}):
- out = sampling_function(self.inner_model.apply_model, x, timestep, uncond, cond, cond_scale, cond_concat, model_options=model_options)
+ def apply_model(self, x, timestep, cond, uncond, cond_scale, cond_concat=None, model_options={}, seed=None):
+ out = sampling_function(self.inner_model.apply_model, x, timestep, uncond, cond, cond_scale, cond_concat, model_options=model_options, seed=seed)
return out
@@ -301,11 +301,11 @@ class KSamplerX0Inpaint(torch.nn.Module):
def __init__(self, model):
super().__init__()
self.inner_model = model
- def forward(self, x, sigma, uncond, cond, cond_scale, denoise_mask, cond_concat=None, model_options={}):
+ def forward(self, x, sigma, uncond, cond, cond_scale, denoise_mask, cond_concat=None, model_options={}, seed=None):
if denoise_mask is not None:
latent_mask = 1. - denoise_mask
x = x * denoise_mask + (self.latent_image + self.noise * sigma.reshape([sigma.shape[0]] + [1] * (len(self.noise.shape) - 1))) * latent_mask
- out = self.inner_model(x, sigma, cond=cond, uncond=uncond, cond_scale=cond_scale, cond_concat=cond_concat, model_options=model_options)
+ out = self.inner_model(x, sigma, cond=cond, uncond=uncond, cond_scale=cond_scale, cond_concat=cond_concat, model_options=model_options, seed=seed)
if denoise_mask is not None:
out *= denoise_mask
@@ -542,7 +542,7 @@ class KSampler:
sigmas = self.calculate_sigmas(new_steps).to(self.device)
self.sigmas = sigmas[-(steps + 1):]
- def sample(self, noise, positive, negative, cfg, latent_image=None, start_step=None, last_step=None, force_full_denoise=False, denoise_mask=None, sigmas=None, callback=None, disable_pbar=False):
+ def sample(self, noise, positive, negative, cfg, latent_image=None, start_step=None, last_step=None, force_full_denoise=False, denoise_mask=None, sigmas=None, callback=None, disable_pbar=False, seed=None):
if sigmas is None:
sigmas = self.sigmas
sigma_min = self.sigma_min
@@ -589,7 +589,7 @@ class KSampler:
if latent_image is not None:
latent_image = self.model.process_latent_in(latent_image)
- extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": self.model_options}
+ extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": self.model_options, "seed":seed}
cond_concat = None
if hasattr(self.model, 'concat_keys'): #inpaint
diff --git a/nodes.py b/nodes.py
index 456805c17..7280d7880 100644
--- a/nodes.py
+++ b/nodes.py
@@ -965,7 +965,7 @@ def common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive,
samples = comfy.sample.sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image,
denoise=denoise, disable_noise=disable_noise, start_step=start_step, last_step=last_step,
- force_full_denoise=force_full_denoise, noise_mask=noise_mask, callback=callback)
+ force_full_denoise=force_full_denoise, noise_mask=noise_mask, callback=callback, seed=seed)
out = latent.copy()
out["samples"] = samples
return (out, )