Simplify IMAGES decode: full latent in one pass, then apply transitions

Removed per-scene slicing (_decode_video/_decode_tensor) and the aggressive
del-before-decode GC block. Now: whole sampling loop runs to completion →
full latent decoded in one pass → transition pixel effects applied after.
LATENT output is unchanged. Per-scene decode was causing seam cross-contamination
anyway and contradicted the post-process-only contract for transitions.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

samplers.py

@@ -1193,45 +1193,6 @@ class FunPackLTXAVSceneChainSampler:
                 if sigma > 0:
                     frames[i] = self._blur_pixel_frame(frames[i], sigma)
 
-    def _decode_tensor(self, video_tensor, vae, tile_size=0):
-        """Decode a single video tensor to [N, H, W, C] pixel frames."""
-        if tile_size > 0:
-            try:
-                decoded = vae.decode_tiled(video_tensor, tile_x=tile_size // 8, tile_y=tile_size // 8)
-            except Exception:
-                decoded = vae.decode(video_tensor)
-        else:
-            decoded = vae.decode(video_tensor)
-        if decoded.dim() == 5:
-            b, t, h, w, c = decoded.shape
-            decoded = decoded.reshape(b * t, h, w, c)
-        return decoded
-
-    def _decode_video(self, output, vae, tile_size=0, scene_boundaries=None):
-        """Decode video latent to [N, H, W, C] pixel frames.
-
-        When scene_boundaries is provided (list of latent frame positions where
-        each new scene starts), each scene slice is decoded independently.
-        Decoding the full concatenated latent in one pass lets the VAE's temporal
-        attention spread seam artifacts forward across scenes — per-scene decode
-        eliminates that cross-contamination.
-        """
-        video_tensor = self._latent_tensors(output)[0]
-        if not scene_boundaries:
-            return self._decode_tensor(video_tensor, vae, tile_size)
-        # Build scene segments: [0, b0, b1, ..., total]
-        total = self._tensor_frames(video_tensor)
-        edges = [0] + list(scene_boundaries) + [total]
-        parts = []
-        for start, end in zip(edges, edges[1:]):
-            if start >= end:
-                continue
-            sliced = self._time_slice(video_tensor, start, end)
-            parts.append(self._decode_tensor(sliced, vae, tile_size))
-        if not parts:
-            return self._decode_tensor(video_tensor, vae, tile_size)
-        return torch.cat(parts, dim=0)
-
     def _apply_transitions_pixel(self, decoded, boundary_entries, transition_duration):
         """Apply visual transitions on a full decoded [N, H, W, C] frame tensor."""
         active = [e for e in boundary_entries if e.get("effect") and e["effect"] != "none"]
@@ -1341,31 +1302,27 @@ class FunPackLTXAVSceneChainSampler:
             report_lines.append(f"Scene {scene_index + 1}: seed={scene_seed}, text={self._scene_text(scene_cond, scene_index)}")
 
         # RETURN_TYPES slot indices: 0=latent, 1=images
+        # Sampling is fully complete. Latent is untouched and returned as-is.
+        # IMAGES: decode the whole latent in one pass, then apply transition effects.
         want_image = self._output_connected(prompt, unique_id, 1)
 
-        # Release all generation-side references so ComfyUI can unload the diffusion
-        # model and text encoder before VAE decode runs. scene_conditionings is a
-        # slice of positive; scene_positive/negative are loop-scoped conditioning
-        # copies — all keep the LTXAVTEModel alive through their embedding tensors.
-        del model, positive, negative, scene_conditionings, scene_cond, scene_positive, scene_negative, chunk, sampled
-
-        # Scene boundaries for per-scene decode: each entry's boundary_latent is
-        # where the next scene's new content starts in the concatenated latent.
-        scene_boundaries = [int(e["boundary_latent"]) for e in boundary_entries] if boundary_entries else None
-
         images = None
         if want_image:
-            # Decode each scene slice independently so the VAE's temporal attention
-            # never crosses a seam — full-video decode spreads seam artifacts forward.
-            # Transitions are pure post-process on the concatenated pixel frames.
-            decoded = self._decode_video(output, vae, decode_tile_size, scene_boundaries)
-            if boundary_entries:
-                images = self._apply_transitions_pixel(decoded, boundary_entries, transition_duration)
+            video_tensor = self._latent_tensors(output)[0]
+            if decode_tile_size > 0:
+                try:
+                    decoded = vae.decode_tiled(video_tensor, tile_x=decode_tile_size // 8, tile_y=decode_tile_size // 8)
+                except Exception:
+                    decoded = vae.decode(video_tensor)
             else:
-                images = decoded
+                decoded = vae.decode(video_tensor)
+            if decoded.dim() == 5:
+                b, t, h, w, c = decoded.shape
+                decoded = decoded.reshape(b * t, h, w, c)
+            images = self._apply_transitions_pixel(decoded, boundary_entries, transition_duration)
 
         if images is None:
-            images = torch.zeros(1, 8, 8, 3)  # placeholder when IMAGE not connected
+            images = torch.zeros(1, 8, 8, 3)
 
         import json as _json
         final_frames = self._tensor_frames(self._latent_tensors(output)[0])