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.

Add ipndm_v sampler, works best with the exponential scheduler.

comfyanonymousd5efde89b76c
diff --git a/comfy/k_diffusion/sampling.py b/comfy/k_diffusion/sampling.py
index 214199266..78896b8b5 100644
--- a/comfy/k_diffusion/sampling.py
+++ b/comfy/k_diffusion/sampling.py
@@ -882,3 +882,66 @@ def sample_ipndm(model, x, sigmas, extra_args=None, callback=None, disable=None,
buffer_model.append(d_cur)
return x_next
+
+#From https://github.com/zju-pi/diff-sampler/blob/main/diff-solvers-main/solvers.py
+#under Apache 2 license
+def sample_ipndm_v(model, x, sigmas, extra_args=None, callback=None, disable=None, max_order=4):
+ extra_args = {} if extra_args is None else extra_args
+ s_in = x.new_ones([x.shape[0]])
+
+ x_next = x
+ t_steps = sigmas
+
+ buffer_model = []
+ for i in trange(len(sigmas) - 1, disable=disable):
+ t_cur = sigmas[i]
+ t_next = sigmas[i + 1]
+
+ x_cur = x_next
+
+ denoised = model(x_cur, t_cur * s_in, **extra_args)
+ if callback is not None:
+ callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+
+ d_cur = (x_cur - denoised) / t_cur
+
+ order = min(max_order, i+1)
+ if order == 1: # First Euler step.
+ x_next = x_cur + (t_next - t_cur) * d_cur
+ elif order == 2: # Use one history point.
+ h_n = (t_next - t_cur)
+ h_n_1 = (t_cur - t_steps[i-1])
+ coeff1 = (2 + (h_n / h_n_1)) / 2
+ coeff2 = -(h_n / h_n_1) / 2
+ x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1])
+ elif order == 3: # Use two history points.
+ h_n = (t_next - t_cur)
+ h_n_1 = (t_cur - t_steps[i-1])
+ h_n_2 = (t_steps[i-1] - t_steps[i-2])
+ temp = (1 - h_n / (3 * (h_n + h_n_1)) * (h_n * (h_n + h_n_1)) / (h_n_1 * (h_n_1 + h_n_2))) / 2
+ coeff1 = (2 + (h_n / h_n_1)) / 2 + temp
+ coeff2 = -(h_n / h_n_1) / 2 - (1 + h_n_1 / h_n_2) * temp
+ coeff3 = temp * h_n_1 / h_n_2
+ x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1] + coeff3 * buffer_model[-2])
+ elif order == 4: # Use three history points.
+ h_n = (t_next - t_cur)
+ h_n_1 = (t_cur - t_steps[i-1])
+ h_n_2 = (t_steps[i-1] - t_steps[i-2])
+ h_n_3 = (t_steps[i-2] - t_steps[i-3])
+ temp1 = (1 - h_n / (3 * (h_n + h_n_1)) * (h_n * (h_n + h_n_1)) / (h_n_1 * (h_n_1 + h_n_2))) / 2
+ temp2 = ((1 - h_n / (3 * (h_n + h_n_1))) / 2 + (1 - h_n / (2 * (h_n + h_n_1))) * h_n / (6 * (h_n + h_n_1 + h_n_2))) \
+ * (h_n * (h_n + h_n_1) * (h_n + h_n_1 + h_n_2)) / (h_n_1 * (h_n_1 + h_n_2) * (h_n_1 + h_n_2 + h_n_3))
+ coeff1 = (2 + (h_n / h_n_1)) / 2 + temp1 + temp2
+ coeff2 = -(h_n / h_n_1) / 2 - (1 + h_n_1 / h_n_2) * temp1 - (1 + (h_n_1 / h_n_2) + (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3)))) * temp2
+ coeff3 = temp1 * h_n_1 / h_n_2 + ((h_n_1 / h_n_2) + (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3))) * (1 + h_n_2 / h_n_3)) * temp2
+ coeff4 = -temp2 * (h_n_1 * (h_n_1 + h_n_2) / (h_n_2 * (h_n_2 + h_n_3))) * h_n_1 / h_n_2
+ x_next = x_cur + (t_next - t_cur) * (coeff1 * d_cur + coeff2 * buffer_model[-1] + coeff3 * buffer_model[-2] + coeff4 * buffer_model[-3])
+
+ if len(buffer_model) == max_order - 1:
+ for k in range(max_order - 2):
+ buffer_model[k] = buffer_model[k+1]
+ buffer_model[-1] = d_cur.detach()
+ else:
+ buffer_model.append(d_cur.detach())
+
+ return x_next
diff --git a/comfy/samplers.py b/comfy/samplers.py
index 4de377cc0..55b1486b4 100644
--- a/comfy/samplers.py
+++ b/comfy/samplers.py
@@ -540,7 +540,7 @@ class Sampler:
KSAMPLER_NAMES = ["euler", "euler_ancestral", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
"lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_sde", "dpmpp_sde_gpu",
"dpmpp_2m", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
- "ipndm"]
+ "ipndm", "ipndm_v"]
class KSAMPLER(Sampler):
def __init__(self, sampler_function, extra_options={}, inpaint_options={}):