Make a fix to the video tensors.

finetrainers/conditioning/conditioned_pipeline.py

@@ -198,7 +198,7 @@ class LTXConditionedPipeline(LTXPipeline):
 
         img_ref_latent = condition_latents_prepare.prepare_latents_for_conditioning(
             vae=self.vae,
-            image_or_video=pose_video,
+            image_or_video=img_ref_video,
             patch_size=self.transformer.config.patch_size,
             patch_size_t=self.transformer.config.patch_size_t,
             device=device,

finetrainers/trainer.py

@@ -1359,6 +1359,8 @@ class Trainer:
         frames = frames[: max_num_frames].float()
         frames = frames.permute(0, 3, 1, 2).contiguous()
         frames = torch.stack([frame for frame in frames], dim=0)
+        # add singleton dimension for downstream inference since no preprocessing took place
+        frames = frames.unsqueeze(0)
         return frames
     
     @staticmethod
@@ -1381,4 +1383,6 @@ class Trainer:
         # nearest_res = self._find_nearest_resolution(frames.shape[2], frames.shape[3])
         # frames_resized = torch.stack([frame for frame in frames], dim=0)
         frames = torch.stack([frame for frame in frames], dim=0)
+        # adding singleton dimension for downstream inference.
+        frames = frames.unsqueeze(0)
         return frames