From 4ab3b25c13d4db9b236d4328addc25fd11fd44e1 Mon Sep 17 00:00:00 2001 From: CrossProduct Date: Tue, 28 Jan 2025 23:03:04 +0000 Subject: [PATCH] Make a fix to the video tensors. --- finetrainers/conditioning/conditioned_pipeline.py | 2 +- finetrainers/trainer.py | 4 ++++ 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/finetrainers/conditioning/conditioned_pipeline.py b/finetrainers/conditioning/conditioned_pipeline.py index 3f6a8dc..4355359 100644 --- a/finetrainers/conditioning/conditioned_pipeline.py +++ b/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, diff --git a/finetrainers/trainer.py b/finetrainers/trainer.py index 1fba463..35015a1 100644 --- a/finetrainers/trainer.py +++ b/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