fix bugs in i2v training dataset (#185)

2025-04-29 19:08:59 +08:00
parent 7da0e755eb
commit 7d91e1369c
1 changed files with 1 additions and 1 deletions
@@ -759,7 +759,7 @@ class ImageVideoControlDataset(Dataset):

        if self.enable_inpaint and not self.enable_bucket:
            mask = get_random_mask(pixel_values.size())
-            mask_pixel_values = pixel_values * (1 - mask) + torch.ones_like(pixel_values) * -1 * mask
+            mask_pixel_values = pixel_values * (1 - mask) + torch.zeros_like(pixel_values) * mask
            sample["mask_pixel_values"] = mask_pixel_values
            sample["mask"] = mask