Adjust the x0 of some DFEs

This commit is contained in:
Jaret Burkett committed 2026-07-22 10:38:40 -06:00
1 parent e8573dad34
commit d5612dd35c
1 file changed
+16 -1
+16 -1
View File
@@ -537,7 +537,18 @@ class DiffusionFeatureExtractor4(nn.Module):
if model is not None and hasattr(model, 'get_stepped_pred'): if model is not None and hasattr(model, 'get_stepped_pred'):
stepped_latents = model.get_stepped_pred(noise_pred, noise) stepped_latents = model.get_stepped_pred(noise_pred, noise)
else: else:
stepped_latents = self.step_latents(noise, noise_pred, noisy_latents, timesteps, scheduler) with torch.no_grad():
tv = timesteps.to(noise_pred.device).to(noise_pred.dtype) / 1000.0
# expand shape to match noise_pred
while len(tv.shape) < len(noise_pred.shape):
tv = tv.unsqueeze(-1)
# min 0.001
tv = torch.clamp(tv, min=0.001)
# step latent
x0 = noisy_latents - tv * noise_pred
stepped_latents = x0
latents = stepped_latents.to(self.vae.device, dtype=self.vae.dtype) latents = stepped_latents.to(self.vae.device, dtype=self.vae.dtype)
@@ -684,6 +695,10 @@ class DiffusionFeatureExtractor6(nn.Module):
if not torch.is_floating_point(x): if not torch.is_floating_point(x):
x = x.float() x = x.float()
# VAE decode can overshoot [0, 1] slightly; clamp so the rescale
# heuristic below never mistakes an overshoot for 0..255 input
x = torch.clamp(x, 0.0, 1.0)
# Resize # Resize
# if not divisible by 16 or total pixels > max_res*max_res, resize to fit within 16 patches # if not divisible by 16 or total pixels > max_res*max_res, resize to fit within 16 patches
max_res = 512 max_res = 512