Revert VAE encoding to original sequential loop

gbuffer_utils.py

@@ -21,15 +21,15 @@ class WanVideoUnit_GBufferEncoder(PipelineUnit):
         if gbuffer_videos is None:
             return {}
         pipe.load_models_to_device(self.onload_model_names)
-        # Batch all modalities into a single vae.encode() call
-        video_tensors = [pipe.preprocess_video(gv) for gv in gbuffer_videos]
-        # vae.encode expects a list of (C,T,H,W) tensors; preprocess_video returns (1,C,T,H,W)
-        video_tensors = [vt.squeeze(0) for vt in video_tensors]
-        all_latents = pipe.vae.encode(
-            video_tensors, device=pipe.device,
-            tiled=tiled, tile_size=tile_size, tile_stride=tile_stride
-        ).to(dtype=pipe.torch_dtype, device=pipe.device)  # [N, 16, T, H, W]
-        gbuffer_latents = all_latents.reshape(1, -1, *all_latents.shape[2:])  # [1, N*16, T, H, W]
+        all_latents = []
+        for gbuffer_video in gbuffer_videos:
+            video_tensor = pipe.preprocess_video(gbuffer_video)
+            latent = pipe.vae.encode(
+                video_tensor, device=pipe.device,
+                tiled=tiled, tile_size=tile_size, tile_stride=tile_stride
+            ).to(dtype=pipe.torch_dtype, device=pipe.device)
+            all_latents.append(latent)
+        gbuffer_latents = torch.cat(all_latents, dim=1)  # [1, N*16, T, H, W]
         if y is not None:
             gbuffer_latents = torch.cat([y, gbuffer_latents], dim=1)
         return {"y": gbuffer_latents}