Normalize MSR slideshow frame sizes

ltx_director.py

@@ -1661,6 +1661,12 @@ def _build_reference_mode_outputs(
         if vae is None:
             raise ValueError("Licon MSR (Prefix) ref option requires connecting the VAE to LTX Director!")
 
+        def _normalize_msr_slide(tensor: torch.Tensor) -> torch.Tensor:
+            if tensor.shape[1] != latent_h or tensor.shape[2] != latent_w:
+                # MSR slideshow frames must all match exactly before stacking.
+                tensor = _resize_image(tensor, latent_w, latent_h, "stretch to fit", divisible_by)
+            return tensor
+
         prompt_text = (global_prompt or "") + " " + (local_prompts or "")
         tag_groups = _build_character_tag_groups(characters)
         referenced_slots = [i for i, tags in enumerate(tag_groups) if any(t in prompt_text for t in tags)]
@@ -1672,12 +1678,10 @@ def _build_reference_mode_outputs(
             selected = char_images
         log.info("[LTXDirector] MSR slideshow subjects: referenced slots=%s, images=%d", referenced_slots, len(selected))
 
-        identity_images = [prepare_tensor_image(c, latent_w, latent_h) for c in selected]
+        identity_images = [_normalize_msr_slide(prepare_tensor_image(c, latent_w, latent_h)) for c in selected]
         scene_images = list(guide_data["images"])
         if scene_images:
-            bg_slide = scene_images[0]
-            if bg_slide.shape[1] != latent_h or bg_slide.shape[2] != latent_w:
-                bg_slide = _resize_image(bg_slide, latent_w, latent_h, resize_method, divisible_by)
+            bg_slide = _normalize_msr_slide(scene_images[0])
         else:
             bg_slide = torch.zeros((1, latent_h, latent_w, 3), dtype=torch.float32)