Todo fix the issue with the conditioning we do not re adjust the values into the model.

2026-10-29 09:07:10 +00:00 · 2025-02-13 18:03:26 +00:00
parent d65958b181
commit 6c2ba374ae
4 changed files with 378 additions and 88 deletions
@@ -813,7 +813,7 @@ class Trainer:
                            noisy_latents = noisy_latents.to(target_video_latents.dtype)

                            # add pose information to both channels
-                            pose_noisy_latents = noisy_latents + pose_video_latents["latents"]
+                            pose_noisy_latents = noisy_latents
                            pose_img_ref_latents = img_refs_latents["latents"] + pose_video_latents["latents"]
                            
                            # expand channel information # B x 2C  latent will be projected to adapter to scale it back to 128d using adapter.