Update app.py

app.py

@@ -43,10 +43,25 @@ STYLE_PRESETS = {
     "Artistic": "artistic style, creative composition, unique visual style, expressive animation, stylized rendering"
 }
 
-# 固定模型配置 - 使用您的完整私人仓库（添加配置文件后）
-PRIVATE_MODEL = "alexander00001/NSFW_Wan_14b"
-# 备用官方模型
-FALLBACK_MODEL = "THUDM/CogVideoX-5b"
+# 固定模型配置 - 使用CogVideoX + LoRA架构
+BASE_MODEL = "THUDM/CogVideoX-5b"  # 稳定的官方base model
+# 实际可用的LoRA适配器列表
+LORA_CONFIGS = [
+    {
+        "repo_id": "hashu786/CogVideoX-LoRA-CineCam",
+        "filename": "pytorch_lora_weights.safetensors", 
+        "adapter_name": "cinematic_camera",
+        "scale": 0.6
+    },
+    {
+        "repo_id": "alibaba-pai/CogVideoX-Fun-V1.1-Reward-LoRAs",
+        "filename": "pytorch_lora_weights.safetensors",
+        "adapter_name": "quality_reward", 
+        "scale": 0.8
+    }
+    # 注意：由于是NSFW内容，暂时使用增强质量的LoRA
+    # 您可以later添加专门的NSFW LoRA
+]
 
 # 质量增强提示词 - 适配视频
 QUALITY_ENHANCERS = [
@@ -116,69 +131,57 @@ def initialize_model():
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         print(f"🖥️ Using device: {device}")
         
-        print(f"Loading Private NSFW Model: {PRIVATE_MODEL}")
-        print(f"Fallback Model: {FALLBACK_MODEL}")
+        print(f"Loading CogVideoX base model: {BASE_MODEL}")
+        print(f"LoRA configurations: {len(LORA_CONFIGS)} adapters")
         
-        # 首先尝试加载您的私人NSFW模型（完整仓库结构）
+        # 加载基础CogVideoX模型
         try:
-            # 使用WanPipeline而不是CogVideoXPipeline
-            from diffusers import WanPipeline
-            pipeline = WanPipeline.from_pretrained(
-                PRIVATE_MODEL,
+            from diffusers import CogVideoXPipeline
+            pipeline = CogVideoXPipeline.from_pretrained(
+                BASE_MODEL,
                 torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-                use_safetensors=True,
-                trust_remote_code=True,
-                # 更激进的内存优化
-                text_encoder_dtype=torch.float32,
-                device_map="balanced",
-                load_in_8bit=True,  # 8bit量化
-                low_cpu_mem_usage=True  # 低CPU内存使用
+                use_safetensors=True
             )
-            print("Successfully loaded private NSFW Wan model with memory optimization!")
+            print("Successfully loaded CogVideoX base model!")
+            
+            # 加载LoRA适配器
+            for lora_config in LORA_CONFIGS:
+                try:
+                    pipeline.load_lora_weights(
+                        lora_config["repo_id"],
+                        weight_name=lora_config["filename"],
+                        adapter_name=lora_config["adapter_name"]
+                    )
+                    print(f"✓ Loaded LoRA: {lora_config['adapter_name']}")
+                except Exception as lora_error:
+                    print(f"⚠ LoRA loading failed ({lora_config['adapter_name']}): {lora_error}")
             
-        except Exception as private_error:
-            print(f"Private Wan model loading failed: {private_error}")
-            print(f"Falling back to official model: {FALLBACK_MODEL}")
+            # 设置LoRA权重
+            adapter_names = [config["adapter_name"] for config in LORA_CONFIGS]
+            adapter_weights = [config["scale"] for config in LORA_CONFIGS]
+            if adapter_names:
+                pipeline.set_adapters(adapter_names, adapter_weights)
+                print(f"✓ Applied LoRA adapters with weights: {adapter_weights}")
             
-            # 备选：尝试官方Wan2.2-Diffusers版本
-            try:
-                pipeline = WanPipeline.from_pretrained(
-                    "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
-                    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-                    use_safetensors=True,
-                    text_encoder_dtype=torch.float32,
-                    device_map="balanced"
-                )
-                print("Loaded official Wan2.2-Diffusers model")
-            except Exception as wan_error:
-                print(f"Official Wan loading failed: {wan_error}")
-                # 最后备选：CogVideoX
-                from diffusers import CogVideoXPipeline
-                pipeline = CogVideoXPipeline.from_pretrained(
-                    FALLBACK_MODEL,
-                    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-                    use_safetensors=True
-                )
-                print("Loaded CogVideoX as final fallback")
+        except Exception as base_error:
+            print(f"Base model loading failed: {base_error}")
+            print("This should not happen with official CogVideoX model")
+            return False
         
         pipeline = pipeline.to(device)
         
-        # GPU优化 - Wan模型专用内存管理
+        # GPU优化 - CogVideoX优化（更简单可靠）
         if torch.cuda.is_available():
             try:
-                # Wan模型特有的优化方法
                 if hasattr(pipeline, 'enable_model_cpu_offload'):
-                    pipeline.enable_model_cpu_offload()  # 将部分组件移至CPU
+                    pipeline.enable_model_cpu_offload()
                 if hasattr(pipeline, 'enable_vae_tiling'):
-                    pipeline.enable_vae_tiling()  # VAE分块处理
-                if hasattr(pipeline, 'enable_sequential_cpu_offload'):
-                    pipeline.enable_sequential_cpu_offload()  # 顺序CPU卸载
-                # 通用内存优化
+                    pipeline.enable_vae_tiling()
                 try:
                     pipeline.enable_xformers_memory_efficient_attention()
                 except:
                     pass
-                print("Wan model memory optimizations applied")
+                print("CogVideoX memory optimizations applied")
             except Exception as mem_error:
                 print(f"Memory optimization warning: {mem_error}")