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.

Separate denoised and noise estimation in Euler CFG++ (#9008)

chaObserve729a5cc1157
diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py
index 2ed415b1f..a2bc492fd 100644
--- a/comfy/k_diffusion/sampling.py
+++ b/comfy/k_diffusion/sampling.py
@@ -1210,39 +1210,21 @@ def sample_deis(model, x, sigmas, extra_args=None, callback=None, disable=None,
return x_next
-def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
- extra_args = {} if extra_args is None else extra_args
-
- temp = [0]
- def post_cfg_function(args):
- temp[0] = args["uncond_denoised"]
- return args["denoised"]
-
- model_options = extra_args.get("model_options", {}).copy()
- extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
-
- s_in = x.new_ones([x.shape[0]])
- for i in trange(len(sigmas) - 1, disable=disable):
- sigma_hat = sigmas[i]
- denoised = model(x, sigma_hat * s_in, **extra_args)
- d = to_d(x, sigma_hat, temp[0])
- if callback is not None:
- callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised})
- # Euler method
- x = denoised + d * sigmas[i + 1]
- return x
-
@torch.no_grad()
def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
- """Ancestral sampling with Euler method steps."""
+ """Ancestral sampling with Euler method steps (CFG++)."""
extra_args = {} if extra_args is None else extra_args
seed = extra_args.get("seed", None)
noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
- temp = [0]
+ model_sampling = model.inner_model.model_patcher.get_model_object("model_sampling")
+ lambda_fn = partial(sigma_to_half_log_snr, model_sampling=model_sampling)
+
+ uncond_denoised = None
+
def post_cfg_function(args):
- temp[0] = args["uncond_denoised"]
+ nonlocal uncond_denoised
+ uncond_denoised = args["uncond_denoised"]
return args["denoised"]
model_options = extra_args.get("model_options", {}).copy()
@@ -1251,15 +1233,33 @@ def sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=No
s_in = x.new_ones([x.shape[0]])
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
- sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
- d = to_d(x, sigmas[i], temp[0])
- # Euler method
- x = denoised + d * sigma_down
- if sigmas[i + 1] > 0:
- x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
+ if sigmas[i + 1] == 0:
+ # Denoising step
+ x = denoised
+ else:
+ alpha_s = sigmas[i] * lambda_fn(sigmas[i]).exp()
+ alpha_t = sigmas[i + 1] * lambda_fn(sigmas[i + 1]).exp()
+ d = to_d(x, sigmas[i], alpha_s * uncond_denoised) # to noise
+
+ # DDIM stochastic sampling
+ sigma_down, sigma_up = get_ancestral_step(sigmas[i] / alpha_s, sigmas[i + 1] / alpha_t, eta=eta)
+ sigma_down = alpha_t * sigma_down
+
+ # Euler method
+ x = alpha_t * denoised + sigma_down * d
+ if eta > 0 and s_noise > 0:
+ x = x + alpha_t * noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
return x
+
+
+def sample_euler_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
+ """Euler method steps (CFG++)."""
+ return sample_euler_ancestral_cfg_pp(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=0.0, s_noise=0.0, noise_sampler=None)
+
+
@torch.no_grad()
def sample_dpmpp_2s_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
"""Ancestral sampling with DPM-Solver++(2S) second-order steps."""