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.

CFG++ for gradient estimation sampler (#7809)

chaObservc15909bb621b
diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py
index 6388d3faf..77ef748e8 100644
--- a/comfy/k_diffusion/sampling.py
+++ b/comfy/k_diffusion/sampling.py
@@ -1345,28 +1345,52 @@ def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, cal
return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True)
@torch.no_grad()
-def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):
+def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False):
"""Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK"""
extra_args = {} if extra_args is None else extra_args
s_in = x.new_ones([x.shape[0]])
old_d = None
+ uncond_denoised = None
+ def post_cfg_function(args):
+ nonlocal uncond_denoised
+ uncond_denoised = args["uncond_denoised"]
+ return args["denoised"]
+
+ if cfg_pp:
+ 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)
+
for i in trange(len(sigmas) - 1, disable=disable):
denoised = model(x, sigmas[i] * s_in, **extra_args)
- d = to_d(x, sigmas[i], denoised)
+ if cfg_pp:
+ d = to_d(x, sigmas[i], uncond_denoised)
+ else:
+ d = to_d(x, sigmas[i], denoised)
if callback is not None:
callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
dt = sigmas[i + 1] - sigmas[i]
if i == 0:
# Euler method
- x = x + d * dt
+ if cfg_pp:
+ x = denoised + d * sigmas[i + 1]
+ else:
+ x = x + d * dt
else:
# Gradient estimation
- d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
- x = x + d_bar * dt
+ if cfg_pp:
+ d_bar = (ge_gamma - 1) * (d - old_d)
+ x = denoised + d * sigmas[i + 1] + d_bar * dt
+ else:
+ d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
+ x = x + d_bar * dt
old_d = d
return x
+def sample_gradient_estimation_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):
+ return sample_gradient_estimation(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, ge_gamma=ge_gamma, cfg_pp=True)
+
@torch.no_grad()
def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3):
"""
diff --git a/comfy/samplers.py b/comfy/samplers.py
index 27dfce45a..67ae09a25 100644
--- a/comfy/samplers.py
+++ b/comfy/samplers.py
@@ -710,7 +710,7 @@ KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_c
"lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
"dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
"ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
- "gradient_estimation", "er_sde", "seeds_2", "seeds_3"]
+ "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"]
class KSAMPLER(Sampler):
def __init__(self, sampler_function, extra_options={}, inpaint_options={}):