fixed noise

Author: 67372a

train_network.py

@@ -2622,7 +2622,7 @@ def loss_fn(noise_pred: torch.Tensor, target: torch.Tensor, scale: float = 1.0):
                             # Sample noise, sample a random timestep for each image, and add noise to the latents,
                             # with noise offset and/or multires noise if specified
                             noise, noisy_guidance, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents_for_guidance)
-                            _, noisy_train, _ = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents_for_train, timesteps)
+                            _, noisy_train, _ = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents_for_train, timesteps, fixed_noise=noise)
 
                             noisy_guidance = noise_scheduler.add_noise(latents_for_guidance, noise, timesteps)
                             noisy_train = noise_scheduler.add_noise(latents_for_train, noise, timesteps)