Update app.py

app.py

@@ -34,26 +34,39 @@ from torchvision.transforms.functional import to_pil_image
 
 app = Flask(__name__)
 
-# Base paths for models
+# Chemins de base pour les modèles
 base_path = 'yisol/IDM-VTON'
 
-# Load models
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-unet = UNet2DConditionModel.from_pretrained(base_path, subfolder="unet", torch_dtype=torch.float16).to(device)
-tokenizer_one = AutoTokenizer.from_pretrained(base_path, subfolder="tokenizer", use_fast=False)
-tokenizer_two = AutoTokenizer.from_pretrained(base_path, subfolder="tokenizer_2", use_fast=False)
+# Chargement des modèles
+unet = UNet2DConditionModel.from_pretrained(
+    base_path,
+    subfolder="unet",
+    torch_dtype=torch.float16,
+    force_download=False
+)
+tokenizer_one = AutoTokenizer.from_pretrained(
+    base_path,
+    subfolder="tokenizer",
+    use_fast=False,
+    force_download=False
+)
+tokenizer_two = AutoTokenizer.from_pretrained(
+    base_path,
+    subfolder="tokenizer_2",
+    use_fast=False,
+    force_download=False
+)
 noise_scheduler = DDPMScheduler.from_pretrained(base_path, subfolder="scheduler")
-text_encoder_one = CLIPTextModel.from_pretrained(base_path, subfolder="text_encoder", torch_dtype=torch.float16).to(device)
-text_encoder_two = CLIPTextModelWithProjection.from_pretrained(base_path, subfolder="text_encoder_2", torch_dtype=torch.float16).to(device)
-image_encoder = CLIPVisionModelWithProjection.from_pretrained(base_path, subfolder="image_encoder", torch_dtype=torch.float16).to(device)
-vae = AutoencoderKL.from_pretrained(base_path, subfolder="vae", torch_dtype=torch.float16).to(device)
-UNet_Encoder = UNet2DConditionModel_ref.from_pretrained(base_path, subfolder="unet_encoder", torch_dtype=torch.float16).to(device)
+text_encoder_one = CLIPTextModel.from_pretrained(base_path, subfolder="text_encoder", torch_dtype=torch.float16)
+text_encoder_two = CLIPTextModelWithProjection.from_pretrained(base_path, subfolder="text_encoder_2", torch_dtype=torch.float16)
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(base_path, subfolder="image_encoder", torch_dtype=torch.float16)
+vae = AutoencoderKL.from_pretrained(base_path, subfolder="vae", torch_dtype=torch.float16)
+UNet_Encoder = UNet2DConditionModel_ref.from_pretrained(base_path, subfolder="unet_encoder", torch_dtype=torch.float16)
 
-parsing_model = Parsing(0).to(device)
-openpose_model = OpenPose(0).to(device)
+parsing_model = Parsing(0)
+openpose_model = OpenPose(0)
 
-# Prepare Tryon pipeline
+# Préparation du pipeline Tryon
 pipe = TryonPipeline.from_pretrained(
     base_path,
     unet=unet,
@@ -66,11 +79,12 @@ pipe = TryonPipeline.from_pretrained(
     scheduler=noise_scheduler,
     image_encoder=image_encoder,
     torch_dtype=torch.float16,
-).to(device)
+    force_download=False
+)
 pipe.unet_encoder = UNet_Encoder
 
-# Image transformation
-tensor_transform = transforms.Compose([
+# Utilisation des transformations d'images
+tensor_transfrom = transforms.Compose([
     transforms.ToTensor(),
     transforms.Normalize([0.5], [0.5]),
 ])
@@ -82,11 +96,13 @@ def pil_to_binary_mask(pil_image, threshold=0):
     mask = np.zeros(binary_mask.shape, dtype=np.uint8)
     mask[binary_mask] = 1
     return Image.fromarray((mask * 255).astype(np.uint8))
+    
+
 
 def get_image_from_url(url):
     try:
         response = requests.get(url)
-        response.raise_for_status()
+        response.raise_for_status()  # Vérifie les erreurs HTTP
         img = Image.open(BytesIO(response.content))
         return img
     except Exception as e:
@@ -117,7 +133,12 @@ def save_image(img):
     return unique_name
 
 @spaces.GPU
-def start_tryon(dict, garm_img, garment_des, is_checked, is_checked_crop, denoise_steps, seed, categorie='upper_body'):
+def start_tryon(dict, garm_img, garment_des, is_checked, is_checked_crop, denoise_steps, seed, categorie = 'upper_body'):
+    device = "cuda"
+    openpose_model.preprocessor.body_estimation.model.to(device)
+    pipe.to(device)
+    pipe.unet_encoder.to(device)
+
     garm_img = garm_img.convert("RGB").resize((768, 1024))
     human_img_orig = dict["background"].convert("RGB")
 
@@ -138,12 +159,11 @@ def start_tryon(dict, garm_img, garment_des, is_checked, is_checked_crop, denois
     if is_checked:
         keypoints = openpose_model(human_img.resize((384, 512)))
         model_parse, _ = parsing_model(human_img.resize((384, 512)))
-        mask, mask_gray = get_mask_location('hd', categorie, model_parse, keypoints)
+        mask, mask_gray = get_mask_location('hd', categorie , model_parse, keypoints)
         mask = mask.resize((768, 1024))
     else:
-        mask = dict['layers'][0].convert("RGB").resize((768, 1024))
-
-    mask_gray = (1 - transforms.ToTensor()(mask)) * tensor_transform(human_img)
+        mask = dict['layers'][0].convert("RGB").resize((768, 1024))#pil_to_binary_mask(dict['layers'][0].convert("RGB").resize((768, 1024)))
+    mask_gray = (1 - transforms.ToTensor()(mask)) * tensor_transfrom(human_img)
     mask_gray = to_pil_image((mask_gray + 1.0) / 2.0)
 
     human_img_arg = _apply_exif_orientation(human_img.resize((384, 512)))
@@ -158,58 +178,66 @@ def start_tryon(dict, garm_img, garment_des, is_checked, is_checked_crop, denois
         with torch.cuda.amp.autocast():
             prompt = "model is wearing " + garment_des
             negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
-            (
-                prompt_embeds,
-                negative_prompt_embeds,
-                pooled_prompt_embeds,
-                negative_pooled_prompt_embeds,
-            ) = pipe.encode_prompt(
-                prompt,
-                num_images_per_prompt=1,
-                do_classifier_free_guidance=True,
-                negative_prompt=negative_prompt,
-            )
-
-            prompt_c = "a photo of " + garment_des
-            negative_prompt_c = "monochrome, lowres, bad anatomy, worst quality, low quality, change color"
-            prompt = [prompt_c] if not isinstance(prompt_c, list) else prompt_c
-            negative_prompt = [negative_prompt_c] if not isinstance(negative_prompt_c, list) else negative_prompt_c
-            
-            (
-                prompt_embeds_c,
-                _,
-                _,
-                _,
-            ) = pipe.encode_prompt(
-                prompt,
-                num_images_per_prompt=1,
-                do_classifier_free_guidance=False,
-                negative_prompt=negative_prompt,
-            )
-
-            pose_img_tensor = tensor_transform(pose_img).unsqueeze(0).to(device, torch.float16)
-            garm_tensor = tensor_transform(garm_img).unsqueeze(0).to(device, torch.float16)
-            generator = torch.Generator(device).manual_seed(seed) if seed is not None else None
-            
-            images = pipe(
-                prompt_embeds=prompt_embeds.to(device),
-                negative_prompt_embeds=negative_prompt_embeds.to(device),
-                pooled_prompt_embeds=pooled_prompt_embeds.to(device),
-                negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.to(device),
-                num_inference_steps=denoise_steps,
-                generator=generator,
-                strength=1.5,
-                pose_img=pose_img_tensor.to(device),
-                text_embeds_cloth=prompt_embeds_c.to(device),
-                cloth=garm_tensor.to(device),
-                mask_image=mask,
-                image=human_img,
-                height=1024,
-                width=768,
-            )
-
-    final_image = images[0] if isinstance(images, list) else images
-    return encode_image_to_base64(final_image)
+            with torch.inference_mode():
+                (
+                    prompt_embeds,
+                    negative_prompt_embeds,
+                    pooled_prompt_embeds,
+                    negative_pooled_prompt_embeds,
+                ) = pipe.encode_prompt(
+                    prompt,
+                    num_images_per_prompt=1,
+                    do_classifier_free_guidance=True,
+                    negative_prompt=negative_prompt,
+                )
+
+                prompt = "a photo of " + garment_des
+                negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality , change color"
+                if not isinstance(prompt, list):
+                    prompt = [prompt] * 1
+                if not isinstance(negative_prompt, list):
+                    negative_prompt = [negative_prompt] * 1
+                with torch.inference_mode():
+                    (
+                        prompt_embeds_c,
+                        _,
+                        _,
+                        _,
+                    ) = pipe.encode_prompt(
+                        prompt,
+                        num_images_per_prompt=1,
+                        do_classifier_free_guidance=False,
+                        negative_prompt=negative_prompt,
+                    )
+
+                pose_img = tensor_transfrom(pose_img).unsqueeze(0).to(device, torch.float16)
+                garm_tensor = tensor_transfrom(garm_img).unsqueeze(0).to(device, torch.float16)
+                generator = torch.Generator(device).manual_seed(seed) if seed is not None else None
+                images = pipe(
+                    prompt_embeds=prompt_embeds.to(device, torch.float16),
+                    negative_prompt_embeds=negative_prompt_embeds.to(device, torch.float16),
+                    pooled_prompt_embeds=pooled_prompt_embeds.to(device, torch.float16),
+                    negative_pooled_prompt_embeds=negative_pooled_prompt_embeds.to(device, torch.float16),
+                    num_inference_steps=denoise_steps,
+                    generator=generator,
+                    strength=1.5,
+                    pose_img=pose_img.to(device, torch.float16),
+                    text_embeds_cloth=prompt_embeds_c.to(device, torch.float16),
+                    cloth=garm_tensor.to(device, torch.float16),
+                    mask_image=mask,
+                    image=human_img,
+                    height=1024,
+                    width=768,
+                    ip_adapter_image=garm_img.resize((768, 1024)),
+                    guidance_scale=1.5,
+                )[0]
+
+    if is_checked_crop:
+        out_img = images[0].resize(crop_size)
+        human_img_orig.paste(out_img, (int(left), int(top)))
+        return human_img_orig, mask_gray
+    else:
+        return images[0], mask_gray , mask
 
         
 @app.route('/tryon-v2', methods=['POST'])
@@ -286,42 +314,51 @@ def tryon():
         'mask_image': mask_base64
     })
 
-@spaces.GPU
-@app.route('/get_mask', methods=['POST'])
-def get_mask():
-    try:
-        # Récupérer l'image du corps à partir de la requête
-        data = request.json
-        img_file = process_image(data['image'])
-        img = img_file.convert("RGB").resize((384, 512))
-        categorie = request.form.get('categorie', 'upper_body')  # Paramètre avec valeur par défaut
-
-        # Appliquer la détection des points clés
-        keypoints = openpose_model(img)  # Utilise votre modèle
-        model_parse, _ = parsing_model(img)  # Utilise votre modèle
+@spaces.GPU 
+def generate_mask(human_img, categorie='upper_body'):
+    device = "cuda"
+    openpose_model.preprocessor.body_estimation.model.to(device)
+    pipe.to(device)
 
-        # Déplacer le modèle et les images sur le même dispositif
-        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        img_tensor = transforms.ToTensor()(img).unsqueeze(0).to(device)  # Convertir et déplacer l'image
-
-        # Assurez-vous que le modèle est sur le même dispositif
-        parsing_model.to(device)
-
-        # Obtenir le masque
-        mask, mask_gray = get_mask_location('hd', categorie, model_parse, keypoints)
+    try:
+        # Redimensionner l'image pour le modèle
+        human_img_resized = human_img.convert("RGB").resize((384, 512))
 
-        # Convertir le masque en image (si nécessaire)
-        mask_gray = (1 - transforms.ToTensor()(mask_gray)) * tensor_transform(img_tensor)
-        mask_gray = to_pil_image((mask_gray + 1.0) / 2.0)
+        # Générer les points clés et le masque
+        keypoints = openpose_model(human_img_resized)
+        model_parse, _ = parsing_model(human_img_resized)
+        mask, _ = get_mask_location('hd', categorie, model_parse, keypoints)
 
-        # Convertir l'image en base64 si besoin pour le retour
-        img_byte_arr = io.BytesIO()
-        mask_gray.save(img_byte_arr, format='PNG')
-        img_byte_arr.seek(0)
-        return jsonify({'mask': img_byte_arr.getvalue().decode('latin1')})
+        # Redimensionner le masque à la taille d'origine de l'image
+        mask_resized = mask.resize(human_img.size)
+        
+        return mask_resized
+    except Exception as e:
+        logging.error(f"Error generating mask: {e}")
+        raise e
 
+@app.route('/generate_mask', methods=['POST'])
+def generate_mask_api():
+    try:
+        # Récupérer les données de l'image à partir de la requête
+        data = request.json
+        base64_image = data.get('image')
+        categorie = data.get('categorie', 'upper_body')
+        
+        # Décodage de l'image à partir de base64
+        human_img = process_image(base64_image)
+        
+        # Appeler la fonction pour générer le masque
+        mask_resized = generate_mask(human_img, categorie)
+        
+        # Encodage du masque en base64 pour la réponse
+        mask_base64 = encode_image_to_base64(mask_resized)
+        
+        return jsonify({
+            'mask_image': mask_base64
+        }), 200
     except Exception as e:
-        print(e)
+        logging.error(f"Error generating mask: {e}")
         return jsonify({'error': str(e)}), 500
 
 # Route index