github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Scale dfe 7 with velocity_equiv_weight

Browse Source
This commit is contained in:
Jaret Burkett
2026-05-13 09:06:59 -06:00
parent fc4d6ebf39
commit 4bfe944792
1 changed files with 14 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
toolkit/models/diffusion_feature_extraction.py
View File

@@ -950,6 +950,8 @@ class DiffusionFeatureExtractor7(nn.Module):
device = self.model.device device = self.model.device
dtype = self.model.dtype dtype = self.model.dtype
velocity_equiv_weight = (1.0 / torch.clamp(tv, min=0.1) ** 2)
with torch.no_grad(): with torch.no_grad():
target = self.get_pred(target_0_1) target = self.get_pred(target_0_1)
@@ -957,20 +959,24 @@ class DiffusionFeatureExtractor7(nn.Module):
pred = self.get_pred(pred_images) pred = self.get_pred(pred_images)
head_loss = torch.nn.functional.mse_loss( head_loss = torch.nn.functional.mse_loss(
pred.head.float(), target.head.float() pred.head.float(), target.head.float(), reduction='none'
) ) * velocity_equiv_weight
head_loss = head_loss.mean()
depth_loss = torch.nn.functional.l1_loss( depth_loss = torch.nn.functional.l1_loss(
pred.depth.float(), target.depth.float() pred.depth.float(), target.depth.float(), reduction='none'
) ) * velocity_equiv_weight
depth_loss = depth_loss.mean()
normals_loss = torch.nn.functional.l1_loss( normals_loss = torch.nn.functional.l1_loss(
pred.normals.float(), target.normals.float() pred.normals.float(), target.normals.float(), reduction='none'
) ) * velocity_equiv_weight
normals_loss = normals_loss.mean()
segmentation_loss = torch.nn.functional.l1_loss( segmentation_loss = torch.nn.functional.l1_loss(
pred.segmentation.float(), target.segmentation.float() pred.segmentation.float(), target.segmentation.float(), reduction='none'
) ) * velocity_equiv_weight
segmentation_loss = segmentation_loss.mean()
total_loss = (head_loss + depth_loss + normals_loss + segmentation_loss) / 4.0 total_loss = (head_loss + depth_loss + normals_loss + segmentation_loss) / 4.0