The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Multistep DPM++ SDE samplers for RF (#8541)
chaObserv8e81c507d212
diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.pyindex a8fd98493..8030048fc 100644--- a/comfy/k_diffusion/sampling.py+++ b/comfy/k_diffusion/sampling.py@@ -781,6 +781,7 @@ def sample_dpmpp_2m(model, x, sigmas, extra_args=None, callback=None, disable=Noold_denoised = denoisedreturn x+@torch.no_grad()def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):"""DPM-Solver++(2M) SDE."""@@ -796,9 +797,12 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disablnoise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_samplers_in = x.new_ones([x.shape[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)+ sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)+old_denoised = None- h_last = None- h = None+ h, h_last = None, Nonefor i in trange(len(sigmas) - 1, disable=disable):denoised = model(x, sigmas[i] * s_in, **extra_args)@@ -809,26 +813,29 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disablx = denoisedelse:# DPM-Solver++(2M) SDE- t, s = -sigmas[i].log(), -sigmas[i + 1].log()- h = s - t- eta_h = eta * h+ lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])+ h = lambda_t - lambda_s+ h_eta = h * (eta + 1)++ alpha_t = sigmas[i + 1] * lambda_t.exp()- x = sigmas[i + 1] / sigmas[i] * (-eta_h).exp() * x + (-h - eta_h).expm1().neg() * denoised+ x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_t * (-h_eta).expm1().neg() * denoisedif old_denoised is not None:r = h_last / hif solver_type == 'heun':- x = x + ((-h - eta_h).expm1().neg() / (-h - eta_h) + 1) * (1 / r) * (denoised - old_denoised)+ x = x + alpha_t * ((-h_eta).expm1().neg() / (-h_eta) + 1) * (1 / r) * (denoised - old_denoised)elif solver_type == 'midpoint':- x = x + 0.5 * (-h - eta_h).expm1().neg() * (1 / r) * (denoised - old_denoised)+ x = x + 0.5 * alpha_t * (-h_eta).expm1().neg() * (1 / r) * (denoised - old_denoised)- if eta:- x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * eta_h).expm1().neg().sqrt() * s_noise+ if eta > 0 and s_noise > 0:+ x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noiseold_denoised = denoisedh_last = hreturn x+@torch.no_grad()def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):"""DPM-Solver++(3M) SDE."""@@ -842,6 +849,10 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disablnoise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_samplers_in = x.new_ones([x.shape[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)+ sigmas = offset_first_sigma_for_snr(sigmas, model_sampling)+denoised_1, denoised_2 = None, Noneh, h_1, h_2 = None, None, None@@ -853,13 +864,16 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl# Denoising stepx = denoisedelse:- t, s = -sigmas[i].log(), -sigmas[i + 1].log()- h = s - t+ lambda_s, lambda_t = lambda_fn(sigmas[i]), lambda_fn(sigmas[i + 1])+ h = lambda_t - lambda_sh_eta = h * (eta + 1)- x = torch.exp(-h_eta) * x + (-h_eta).expm1().neg() * denoised+ alpha_t = sigmas[i + 1] * lambda_t.exp()++ x = sigmas[i + 1] / sigmas[i] * (-h * eta).exp() * x + alpha_t * (-h_eta).expm1().neg() * denoisedif h_2 is not None:+ # DPM-Solver++(3M) SDEr0 = h_1 / hr1 = h_2 / hd1_0 = (denoised - denoised_1) / r0@@ -868,20 +882,22 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabld2 = (d1_0 - d1_1) / (r0 + r1)phi_2 = h_eta.neg().expm1() / h_eta + 1phi_3 = phi_2 / h_eta - 0.5- x = x + phi_2 * d1 - phi_3 * d2+ x = x + (alpha_t * phi_2) * d1 - (alpha_t * phi_3) * d2elif h_1 is not None:+ # DPM-Solver++(2M) SDEr = h_1 / hd = (denoised - denoised_1) / rphi_2 = h_eta.neg().expm1() / h_eta + 1- x = x + phi_2 * d+ x = x + (alpha_t * phi_2) * d- if eta:+ if eta > 0 and s_noise > 0:x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * sigmas[i + 1] * (-2 * h * eta).expm1().neg().sqrt() * s_noisedenoised_1, denoised_2 = denoised, denoised_1h_1, h_2 = h, h_1return x+@torch.no_grad()def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):if len(sigmas) <= 1:@@ -891,6 +907,7 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, dinoise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_samplerreturn sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)+@torch.no_grad()def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):if len(sigmas) <= 1:@@ -900,6 +917,7 @@ def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, dinoise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_samplerreturn sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)+@torch.no_grad()def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):if len(sigmas) <= 1: