nodes/Clybius-ComfyUI-Extra-Samplers
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Attempt at fixing reversible_heun_1s prev_denoised, and slightly dampen it more by default.

Browse Source
This commit is contained in:
Clybius
2024-04-05 12:16:25 -05:00
parent e08968efe0
commit 0c89dee207
1 changed files with 4 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
extra_samplers.py
+4 -7
View File
@@ -1146,7 +1146,6 @@ def sampler_supreme(model, x, sigmas, extra_args=None, callback=None, disable=No
case "reversible_heun": # 2 model calls
sigma_i, sigma_i_plus_1 = sigmas[i], sigma_down
dt = sigma_i_plus_1 - sigma_i
dt_nonancestral = sigmas[i + 1] - sigma_i
# Calculate the derivative using the model
d_i = to_d(z_k, sigma_i, denoised)
@@ -1161,15 +1160,14 @@ def sampler_supreme(model, x, sigmas, extra_args=None, callback=None, disable=No
d_i_plus_1 = to_d(x_pred, sigma_i_plus_1, denoised_i_plus_1)
# Update the sample using the Reversible Heun formula
z_k = z_k + dt * (d_i + d_i_plus_1) / 2 - dt_nonancestral**2 * (d_i_plus_1 - d_i) / (4 * reversible_dampen)
z_k = z_k + dt * (d_i + d_i_plus_1) / 2 - dt**2 * (d_i_plus_1 - d_i) / (4 * reversible_dampen)
case "reversible_heun_1s": # Experimental 1 model call variant, utilizing previous denoised variables to speed up diffusion.
# Reversible Heun-inspired update (first-order)
sigma_i, sigma_i_plus_1 = sigmas[i], sigma_down
dt = sigma_i_plus_1 - sigma_i
dt_nonancestral = sigmas[i + 1] - sigma_i
# Calculate the derivative using the model
d_i_old = to_d(prev_x, sigma_i, prev_denoised)
d_i_old = to_d(prev_x, sigma_i, prev_denoised) if prev_denoised is not None else to_d(prev_x, sigma_i, model(prev_x, sigma_i * s_in, **extra_args))
# Predict the sample at the next sigma using Euler step
x_pred = prev_x + d_i_old * dt
@@ -1178,7 +1176,7 @@ def sampler_supreme(model, x, sigmas, extra_args=None, callback=None, disable=No
d_i_plus_1 = to_d(x_pred, sigma_i_plus_1, denoised)
# Update the sample using the Reversible Heun formula
z_k = z_k + dt * (d_i_old + d_i_plus_1) / 2 - dt_nonancestral**2 * (d_i_plus_1 - d_i_old) / (1 * reversible_dampen)
z_k = z_k + dt * (d_i_old + d_i_plus_1) / 2 - dt**2 * (d_i_plus_1 - d_i_old) / (2 * reversible_dampen)
case "rkf45": # 6 model calls (expensive)
sigma_i, sigma_i_plus_1 = sigmas[i], sigma_down
dt = sigma_i_plus_1 - sigma_i
@@ -1226,7 +1224,6 @@ def sampler_supreme(model, x, sigmas, extra_args=None, callback=None, disable=No
case "reversible_bogacki_shampine":
sigma_i, sigma_i_plus_1 = sigmas[i], sigma_down
dt = sigma_i_plus_1 - sigma_i
dt_nonancestral = sigmas[i + 1] - sigma_i
# Calculate the derivative using the model
d_i = to_d(z_k, sigma_i, denoised)
@@ -1237,7 +1234,7 @@ def sampler_supreme(model, x, sigmas, extra_args=None, callback=None, disable=No
k3 = to_d(z_k + 3 * k1 / 4 + k2 / 4, sigma_i + 3 * dt / 4, model(z_k + 3 * k1 / 4 + k2 / 4, (sigma_i + 3 * dt / 4) * s_in, **extra_args)) * dt
# Reversible correction term (inspired by Reversible Heun)
correction = dt_nonancestral**2 * (4 * k3 / 9 - k2 / 3) / (6 * reversible_dampen)
correction = dt**2 * (4 * k3 / 9 - k2 / 3) / (6 * reversible_dampen)
# Update the sample
z_k = z_k + 2 * k1 / 9 + k2 / 3 + 4 * k3 / 9 - correction