Fix: Restore natural look for 7b model (v2.5.6)

- Replace split-stack-mean with unflatten in unconcat_coalesce
- Corrects computation order to eliminate plastic/high-specular artifacts
- Maintains torch.compile compatibility (no .item() graph breaks)
- Applied to both dit_3b and dit_7b models

README.md

@@ -36,7 +36,9 @@ We're actively working on improvements and new features. To stay informed:
 
 ## 🚀 Updates
 
-**2025.11.09 - Version 2.5.5**
+**2025.11.09 - Version 2.5.6**
+
+- 🎨 **Fix: Restored natural look for 7b model** - Corrected torch.compile optimization that was causing overly plastic/ high-specular appearance in upscaled videos with 7b model.
 
 - 💾 **Memory: Fixed RAM leak for long videos** - On-demand reconstruction with lightweight batch indices instead of storing full transformed videos, fixed release_tensor_memory to handle CPU/CUDA/MPS consistently, and refactored batch processing helpers
 

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "seedvr2_videoupscaler"
 description = "SeedVR2 official ComfyUI integration: ByteDance-Seed's one-step diffusion-based video/image upscaling with memory-efficient inference"
-version = "2.5.5"
+version = "2.5.6"
 authors = [
     {name = "numz"},
     {name = "adrientoupet"}

src/models/dit_3b/na.py

@@ -400,22 +400,19 @@ def repeat_concat_idx(
         The text features appear multiple times (once per window) and need
         to be averaged to produce a single set of text features.
         
-        COMPILE OPTIMIZATION: Uses split-stack-mean instead of reshape to avoid .item()
+        COMPILE OPTIMIZATION: Uses unflatten with tensor dims (compile-friendly)
         """
         vid_out, txt_out = all[src_idx].split([len(vid_idx), txt_idx_len])
         
-        # Coalesce repeated text using pure tensor operations (no .item())
+        # Coalesce repeated text using unflatten and mean
         txt_splits = _tensor_split(txt_out, repeat_txt_len, dim=0)
         txt_out_coalesced = []
         
-        for txt, base_len in zip(txt_splits, txt_len):
+        for txt in txt_splits:
             # txt has shape (base_len * num_repeats, *other_dims)
-            # Split into num_repeats equal pieces using tensor operations
-            split_lens = base_len.unsqueeze(0).repeat_interleave(num_repeats_tensor.squeeze())
-            pieces = _tensor_split(txt, split_lens, dim=0)
-            # Stack and average across repetitions
-            txt_avg = torch.stack(pieces, dim=0).mean(0)
-            txt_out_coalesced.append(txt_avg)
+            # unflatten to (base_len, num_repeats, *other_dims) then average dim 1
+            txt = txt.unflatten(0, (-1, num_repeats_tensor.squeeze())).mean(1)
+            txt_out_coalesced.append(txt)
         
         return vid_out, torch.cat(txt_out_coalesced)
 

src/models/dit_7b/na.py

@@ -400,22 +400,19 @@ def repeat_concat_idx(
         The text features appear multiple times (once per window) and need
         to be averaged to produce a single set of text features.
         
-        COMPILE OPTIMIZATION: Uses split-stack-mean instead of reshape to avoid .item()
+        COMPILE OPTIMIZATION: Uses unflatten with tensor dims (compile-friendly)
         """
         vid_out, txt_out = all[src_idx].split([len(vid_idx), txt_idx_len])
         
-        # Coalesce repeated text using pure tensor operations (no .item())
+        # Coalesce repeated text using unflatten and mean
         txt_splits = _tensor_split(txt_out, repeat_txt_len, dim=0)
         txt_out_coalesced = []
         
-        for txt, base_len in zip(txt_splits, txt_len):
+        for txt in txt_splits:
             # txt has shape (base_len * num_repeats, *other_dims)
-            # Split into num_repeats equal pieces using tensor operations
-            split_lens = base_len.unsqueeze(0).repeat_interleave(num_repeats_tensor.squeeze())
-            pieces = _tensor_split(txt, split_lens, dim=0)
-            # Stack and average across repetitions
-            txt_avg = torch.stack(pieces, dim=0).mean(0)
-            txt_out_coalesced.append(txt_avg)
+            # unflatten to (base_len, num_repeats, *other_dims) then average dim 1
+            txt = txt.unflatten(0, (-1, num_repeats_tensor.squeeze())).mean(1)
+            txt_out_coalesced.append(txt)
         
         return vid_out, torch.cat(txt_out_coalesced)