From cecc49da31ea62b8244391dd1ccd8b87fd9b79bc Mon Sep 17 00:00:00 2001
From: Enrico <erosdiffusionai@gmail.com>
Date: Mon, 19 Jan 2026 23:58:52 +0100
Subject: [PATCH] tensor fix

---
 audio_blender.py | 15 +++++++++++----
 test_parsing.py  | 30 ++++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+), 4 deletions(-)
 create mode 100644 test_parsing.py

diff --git a/audio_blender.py b/audio_blender.py
index 2489729..b0005c9 100644
--- a/audio_blender.py
+++ b/audio_blender.py
@@ -111,11 +111,18 @@ class AudioOverlapBlender:
         )
         
         # Reshape for broadcasting
-        # Audio latent is typically [batch, channels, frames, features]
-        while alpha.dim() < prev_tail.dim():
-            alpha = alpha.unsqueeze(0)
+        # Audio latent is [batch, channels, frames, freq_bins]
+        # Alpha needs to be [1, 1, actual_overlap, 1] to broadcast
         if prev_tail.dim() == 4:
-            alpha = alpha.unsqueeze(-1)  # [1, 1, overlap, 1]
+            # Audio: [B, C, T, F] -> alpha: [1, 1, overlap, 1]
+            alpha = alpha.view(1, 1, actual_overlap, 1)
+        elif prev_tail.dim() == 5:
+            # Video: [B, C, T, H, W] -> alpha: [1, 1, overlap, 1, 1]
+            alpha = alpha.view(1, 1, actual_overlap, 1, 1)
+        else:
+            # Fallback: add dims at front
+            while alpha.dim() < prev_tail.dim():
+                alpha = alpha.unsqueeze(0)
         
         alpha = alpha.expand_as(prev_tail)
         
diff --git a/test_parsing.py b/test_parsing.py
new file mode 100644
index 0000000..fe22197
--- /dev/null
+++ b/test_parsing.py
@@ -0,0 +1,30 @@
+import sys
+sys.path.insert(0, r'D:\ComfyUI7\ComfyUI\custom_nodes\ComfyUI-Erosdiffusion-LTX2')
+
+from script_parser import parse_scene_script
+
+# Test script 1: User's format
+script1 = """[00:00.00-00:04.04] the woman turns from left to right, static camera | audio:techno music | first:ComfyUI-zimage-diffusers-wrapper_00213_.png | end:ComfyUI-zimage-diffusers-wrapper_00213_.png
+[00:04.04-00:08.00] the woman turns from right to left, static camera  | first:ComfyUI-zimage-diffusers-wrapper_00213_.png | audio:techno music  | end:ComfyUI-zimage-diffusers-wrapper_00212_.png"""
+
+# Test script 2: Missing parts
+script2 = """[00:00-00:04] no prompt no guides
+[00:04-00:08] just a prompt | first:image.png
+[00:08-00:12] | audio:silence"""
+
+print("=== Test 1: User's script ===")
+chunks = parse_scene_script(script1)
+print(f"Parsed {len(chunks)} chunks:")
+for i, c in enumerate(chunks):
+    print(f"  Chunk {i}: {c.start_sec:.2f}s-{c.end_sec:.2f}s")
+    print(f"    prompt: '{c.prompt[:50]}...'")
+    print(f"    audio: '{c.audio_spec}'")
+    print(f"    guides: {len(c.guides)}")
+    for g in c.guides:
+        print(f"      - {g.position}: {g.image_ref}")
+
+print("\n=== Test 2: Missing parts ===")
+chunks2 = parse_scene_script(script2)
+print(f"Parsed {len(chunks2)} chunks:")
+for i, c in enumerate(chunks2):
+    print(f"  Chunk {i}: prompt='{c.prompt}', audio='{c.audio_spec}', guides={len(c.guides)}")