From e6f97e233e028b9cbbd32cc45a8e41b35f22e155 Mon Sep 17 00:00:00 2001
From: Hawk Lee <l.y@live.cn>
Date: Thu, 22 Jan 2026 18:13:00 +0800
Subject: [PATCH] =?UTF-8?q?perf:=20Optimize=20Ditto=20frame=E2=86=92tensor?=
 =?UTF-8?q?=20conversion=20to=20avoid=20memory=20doubling?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 aiia_ditto_nodes.py | 22 ++++++++++++++++++++--
 pyproject.toml      |  2 +-
 2 files changed, 21 insertions(+), 3 deletions(-)

diff --git a/aiia_ditto_nodes.py b/aiia_ditto_nodes.py
index b6a6603..6cb9278 100644
--- a/aiia_ditto_nodes.py
+++ b/aiia_ditto_nodes.py
@@ -771,9 +771,27 @@ class AIIA_DittoSampler:
             raise RuntimeError("Ditto generated 0 frames.")
             
         # Convert List[np.array (H,W,C)] -> Batch Tensor (B,H,W,C)
-        # Note: generated frames are RGB (from writer_queue which usually gets RGB).
+        # Memory-optimized: pre-allocate tensor and release frames progressively
         import torch
-        video_tensor = torch.from_numpy(np.array(generated)).float() / 255.0
+        import gc
+        
+        num_frames = len(generated)
+        h, w, c = generated[0].shape
+        
+        # Pre-allocate output tensor
+        video_tensor = torch.zeros((num_frames, h, w, c), dtype=torch.float32)
+        
+        # Copy frames one by one and release original
+        for i in range(num_frames):
+            video_tensor[i] = torch.from_numpy(generated[i].astype(np.float32) / 255.0)
+            generated[i] = None  # Release original frame
+            if i > 0 and i % 100 == 0:
+                gc.collect()
+        
+        # Final cleanup
+        del generated
+        master_sdk.generated_frames = []
+        gc.collect()
         
         return (video_tensor, audio)
 
diff --git a/pyproject.toml b/pyproject.toml
index beba489..b3500e8 100755
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,7 +1,7 @@
 [project]
 name = "aiia"
 description = "The Ultimate AI Audio/Video toolkit for ComfyUI. Features an enhanced Ditto (with optimizations that outperform official demos and other SOTA talking head models in lip-sync accuracy and natural motion), EchoMimic V3 & FLOAT, VibeVoice & CosyVoice 3.0 (Zero-Shot Voice Cloning), Multi-Role Podcast Generation, and a powerful Media Browser."
-version = "1.9.22"
+version = "1.9.23"
 license = {file = "LICENSE"}
 readme = "README.md"
 authors = [