Update app.py

app.py

@@ -23,80 +23,83 @@ dl_cn_config(cn_dir)
 dl_tagger_model(tagger_dir)
 dl_lora_model(lora_dir)
 
+# グローバル変数でpipeを管理
+pipe = None
+current_lora_model = None
+
+def load_model(lora_model):
+    global pipe, current_lora_model
+    # 既にロードされたpipeがあり、同じLoRAモデルの場合は再利用
+    if pipe is not None and current_lora_model == lora_model:
+        return pipe  # キャッシュされたpipeを返す
+
+    # 新しいpipeの生成
+    dtype = torch.float16
+    vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=dtype)
+    controlnet = ControlNetModel.from_pretrained(cn_dir, torch_dtype=dtype, use_safetensors=True)
+
+    pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
+        "cagliostrolab/animagine-xl-3.1", controlnet=controlnet, vae=vae, torch_dtype=dtype
+    )
+    pipe.enable_model_cpu_offload()
+
+    # LoRAモデルの設定
+    if lora_model == "とりにく風":
+        pipe.load_lora_weights(lora_dir, weight_name="tori29umai_line.safetensors")             
+    elif lora_model == "少女漫画風":
+        pipe.load_lora_weights(lora_dir, weight_name="syoujomannga_line.safetensors")        
+    elif lora_model == "劇画調風":
+        pipe.load_lora_weights(lora_dir, weight_name="gekiga_line.safetensors")
+    elif lora_model == "プレーン":
+        pass  # プレーンの場合はLoRAを読み込まない
+
+    # 現在のLoRAモデルを保存
+    current_lora_model = lora_model
+    return pipe
+
+@spaces.GPU(duration=120)
+def predict(lora_model, input_image_path, prompt, negative_prompt, controlnet_scale):
+    # pipeをグローバル変数から取得
+    pipe = load_model(lora_model)
+    
+    # 画像読み込みとリサイズ
+    input_image = Image.open(input_image_path)
+    base_image = base_generation(input_image.size, (255, 255, 255, 255)).convert("RGB")
+    resize_image = resize_image_aspect_ratio(input_image)
+    resize_base_image = resize_image_aspect_ratio(base_image)
+    generator = torch.manual_seed(0)
+    last_time = time.time()
+    
+    # プロンプト生成
+    prompt = "masterpiece, best quality, monochrome, greyscale, lineart, white background, star-shaped pupils, " + prompt
+    execute_tags = ["realistic", "nose", "asian"]
+    prompt = execute_prompt(execute_tags, prompt)
+    prompt = remove_duplicates(prompt)        
+    prompt = remove_color(prompt)
+    print(prompt)
+
+    # 画像生成
+    output_image = pipe(
+        image=resize_base_image,
+        control_image=resize_image,
+        strength=1.0,
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        controlnet_conditioning_scale=float(controlnet_scale),
+        generator=generator,
+        num_inference_steps=30,
+        eta=1.0,
+    ).images[0]
+    print(f"Time taken: {time.time() - last_time}")
+    output_image = output_image.resize(input_image.size, Image.LANCZOS)
+    return output_image
+
 class Img2Img:
     def __init__(self):
         self.demo = self.layout()
         self.tagger_model = None
         self.input_image_path = None
         self.bg_removed_image = None
-        self.pipe = None
-        self.current_lora_model = None
-
-    def load_model(self, lora_model):
-        # 既にロードされたpipeがあり、同じLoRAモデルの場合は再利用
-        if self.pipe and self.current_lora_model == lora_model:
-            return self.pipe  # キャッシュされたpipeを返す
-
-        # 新しいpipeの生成
-        dtype = torch.float16
-        vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=dtype)
-        controlnet = ControlNetModel.from_pretrained(cn_dir, torch_dtype=dtype, use_safetensors=True)
-
-        self.pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
-            "cagliostrolab/animagine-xl-3.1", controlnet=controlnet, vae=vae, torch_dtype=dtype
-        )
-        self.pipe.enable_model_cpu_offload()
-
-        # LoRAモデルの設定
-        if lora_model == "とりにく風":
-            self.pipe.load_lora_weights(lora_dir, weight_name="tori29umai_line.safetensors")             
-        elif lora_model == "少女漫画風":
-            self.pipe.load_lora_weights(lora_dir, weight_name="syoujomannga_line.safetensors")        
-        elif lora_model == "劇画調風":
-            self.pipe.load_lora_weights(lora_dir, weight_name="gekiga_line.safetensors")
-        elif lora_model == "プレーン":
-            pass  # プレーンの場合はLoRAを読み込まない
-
-        # 現在のLoRAモデルを保存
-        self.current_lora_model = lora_model
-        return self.pipe
-
-    @spaces.GPU(duration=120)
-    def predict(self, lora_model, input_image_path, prompt, negative_prompt, controlnet_scale):
-        # ここで新たなpipeを作成するのではなく、キャッシュしたpipeを取得
-        pipe = self.load_model(lora_model)
-        
-        # 画像読み込みとリサイズ
-        input_image = Image.open(input_image_path)
-        base_image = base_generation(input_image.size, (255, 255, 255, 255)).convert("RGB")
-        resize_image = resize_image_aspect_ratio(input_image)
-        resize_base_image = resize_image_aspect_ratio(base_image)
-        generator = torch.manual_seed(0)
-        last_time = time.time()
-        
-        # プロンプト生成
-        prompt = "masterpiece, best quality, monochrome, greyscale, lineart, white background, star-shaped pupils, " + prompt
-        execute_tags = ["realistic", "nose", "asian"]
-        prompt = execute_prompt(execute_tags, prompt)
-        prompt = remove_duplicates(prompt)        
-        prompt = remove_color(prompt)
-        print(prompt)
-
-        # 画像生成
-        output_image = pipe(
-            image=resize_base_image,
-            control_image=resize_image,
-            strength=1.0,
-            prompt=prompt,
-            negative_prompt=negative_prompt,
-            controlnet_conditioning_scale=float(controlnet_scale),
-            generator=generator,
-            num_inference_steps=30,
-            eta=1.0,
-        ).images[0]
-        print(f"Time taken: {time.time() - last_time}")
-        output_image = output_image.resize(input_image.size, Image.LANCZOS)
-        return output_image
 
     def process_prompt_analysis(self, input_image_path):
         if self.tagger_model is None:
@@ -147,7 +150,7 @@ class Img2Img:
             )
 
             generate_button.click(
-                fn=self.predict,
+                fn=predict,
                 inputs=[self.lora_model, self.bg_removed_image_path, self.prompt, self.negative_prompt, self.controlnet_scale],
                 outputs=self.output_image
             )