Update app.py

app.py

@@ -34,39 +34,26 @@ from torchvision.transforms.functional import to_pil_image
 
 app = Flask(__name__)
 
-# Chemins de base pour les modèles
+# Base paths for models
 base_path = 'yisol/IDM-VTON'
 
-# 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
-)
+# 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)
 noise_scheduler = DDPMScheduler.from_pretrained(base_path, subfolder="scheduler")
-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)
+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)
 
-parsing_model = Parsing(0)
-openpose_model = OpenPose(0)
+parsing_model = Parsing(0).to(device)
+openpose_model = OpenPose(0).to(device)
 
-# Préparation du pipeline Tryon
+# Prepare Tryon pipeline
 pipe = TryonPipeline.from_pretrained(
     base_path,
     unet=unet,
@@ -79,12 +66,11 @@ pipe = TryonPipeline.from_pretrained(
     scheduler=noise_scheduler,
     image_encoder=image_encoder,
     torch_dtype=torch.float16,
-    force_download=False
-)
+).to(device)
 pipe.unet_encoder = UNet_Encoder
 
-# Utilisation des transformations d'images
-tensor_transfrom = transforms.Compose([
+# Image transformation
+tensor_transform = transforms.Compose([
     transforms.ToTensor(),
     transforms.Normalize([0.5], [0.5]),
 ])
@@ -96,13 +82,11 @@ 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()  # Vérifie les erreurs HTTP
+        response.raise_for_status()
         img = Image.open(BytesIO(response.content))
         return img
     except Exception as e:
@@ -133,12 +117,7 @@ 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'):
-    device = "cuda"
-    openpose_model.preprocessor.body_estimation.model.to(device)
-    pipe.to(device)
-    pipe.unet_encoder.to(device)
-
+def start_tryon(dict, garm_img, garment_des, is_checked, is_checked_crop, denoise_steps, seed, categorie='upper_body'):
     garm_img = garm_img.convert("RGB").resize((768, 1024))
     human_img_orig = dict["background"].convert("RGB")
 
@@ -159,11 +138,12 @@ 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))#pil_to_binary_mask(dict['layers'][0].convert("RGB").resize((768, 1024)))
-    mask_gray = (1 - transforms.ToTensor()(mask)) * tensor_transfrom(human_img)
+        mask = dict['layers'][0].convert("RGB").resize((768, 1024))
+
+    mask_gray = (1 - transforms.ToTensor()(mask)) * tensor_transform(human_img)
     mask_gray = to_pil_image((mask_gray + 1.0) / 2.0)
 
     human_img_arg = _apply_exif_orientation(human_img.resize((384, 512)))
@@ -178,66 +158,58 @@ 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"
-            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
+            (
+                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)
 
         
 @app.route('/tryon-v2', methods=['POST'])