Update app.py

app.py

@@ -1,69 +1,177 @@
-import torch
-from diffusers import StableDiffusionImg2ImgPipeline
 import gradio as gr
-import cv2
 import numpy as np
+import random
+import torch
+import spaces
+import cv2
 import face_recognition
+from diffusers import DiffusionPipeline, StableDiffusionImg2ImgPipeline
+from PIL import Image
+from transformers import CLIPTextModel, CLIPTokenizer
+
+# Configuración del dispositivo y tipo de datos
+dtype = torch.bfloat16
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# Cargar los pipelines para las diferentes tareas
+pipe_diffusion = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=dtype).to(device)
+pipe_img2img = StableDiffusionImg2ImgPipeline.from_pretrained("runwayml/stable-diffusion-v1-5").to(device)
 
-# Cargar el pipeline con el modelo FLUX.1-schnell
-pipe = StableDiffusionImg2ImgPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell")
-pipe.to("cuda")  # Asegúrate de que esté en GPU para un mejor rendimiento
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 2048
+
+# Configurar tokenizer y modelo CLIP para truncar correctamente los prompts
+tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
+text_model = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
+
+def truncate_prompt(prompt, max_length=77):
+    inputs = tokenizer(prompt, return_tensors="pt")
+    input_ids = inputs["input_ids"][0][:max_length]
+    truncated_prompt = tokenizer.decode(input_ids, skip_special_tokens=True)
+    return truncated_prompt
 
-# Función para realizar el faceswap
 def swap_faces(source_image, target_image):
-    # Detectar las caras en ambas imágenes
     source_face = face_recognition.face_locations(source_image)[0]
-    source_encoding = face_recognition.face_encodings(source_image, [source_face])[0]
-    
     target_face = face_recognition.face_locations(target_image)[0]
-    target_encoding = face_recognition.face_encodings(target_image, [target_face])[0]
-    
-    # Obtener los puntos de la cara en la imagen original
-    source_points = face_recognition.face_landmarks(source_image, [source_face])[0]
-    target_points = face_recognition.face_landmarks(target_image, [target_face])[0]
-    
-    # Transformar la cara objetivo en la cara fuente
+
     target_face_image = target_image[target_face[0]:target_face[2], target_face[3]:target_face[1]]
     source_face_image = source_image[source_face[0]:source_face[2], source_face[3]:source_face[1]]
-    
-    # Redimensionar la cara fuente para que coincida con la cara objetivo
+
     target_face_resized = cv2.resize(target_face_image, (source_face_image.shape[1], source_face_image.shape[0]))
 
-    # Crear una máscara para la cara
     mask = np.zeros_like(source_face_image)
-    cv2.fillConvexPoly(mask, np.array(list(target_points.values())), (255, 255, 255))
+    cv2.fillConvexPoly(mask, np.array(list(face_recognition.face_landmarks(target_image, [target_face])[0].values())), (255, 255, 255))
 
-    # Intercambiar las caras
     swapped_face = cv2.seamlessClone(target_face_resized, source_image, mask, (source_face[3] + source_face_image.shape[1]//2, source_face[0] + source_face_image.shape[0]//2), cv2.NORMAL_CLONE)
     
     return swapped_face
 
-# Función de generación de imagen
-def generate_image(init_image, strength, prompt, reference_image=None):
-    with torch.cuda.amp.autocast():
-        generated_image = pipe(prompt=prompt, init_image=init_image, strength=strength).images[0]
+@spaces.GPU()
+def infer(prompt, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=4, init_image=None, reference_image=None, img2img_strength=0.75, progress=gr.Progress(track_tqdm=True)):
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    generator = torch.Generator().manual_seed(seed)
+    
+    # Truncar el prompt si es demasiado largo
+    prompt = truncate_prompt(prompt)
+    
+    # Convertir init_image a formato PIL si no lo está
+    if init_image and not isinstance(init_image, Image.Image):
+        init_image = Image.fromarray(np.array(init_image))
+
+    # Generación de la imagen
+    if init_image:
+        init_image = init_image.convert("RGB")
+        generated_image = pipe_img2img(
+            prompt=prompt, 
+            init_image=init_image,
+            strength=img2img_strength,
+            num_inference_steps=num_inference_steps,
+            generator=generator
+        ).images[0]
+    else:
+        generated_image = pipe_diffusion(
+            prompt=prompt, 
+            width=width,
+            height=height,
+            num_inference_steps=num_inference_steps, 
+            generator=generator,
+            guidance_scale=0.0
+        ).images[0]
     
-    # Si se proporciona una imagen de referencia, realizar el face swap
-    if reference_image is not None:
+    # Aplicación de Face Swap
+    if reference_image:
+        reference_image = np.array(reference_image)
         generated_image = np.array(generated_image)
-        generated_image = swap_faces(generated_image, np.array(reference_image))
+        generated_image = swap_faces(generated_image, reference_image)
         generated_image = Image.fromarray(generated_image)
     
-    return generated_image
-
-# Configurar la interfaz de Gradio
-interface = gr.Interface(
-    fn=generate_image,
-    inputs=[
-        gr.Image(source="upload", type="pil", label="Imagen Inicial"),
-        gr.Slider(0.0, 1.0, value=0.75, label="Strength"),
-        gr.Textbox(label="Prompt"),
-        gr.Image(source="upload", type="pil", label="Imagen de Referencia (opcional)"),
-    ],
-    outputs=gr.Image(label="Imagen Generada"),
-    title="Generador de Imágenes Estilo Img2Img con FaceSwap",
-    description="Genera imágenes utilizando el modelo FLUX.1-schnell a partir de una imagen inicial y un prompt, con la opción de intercambiar caras.",
-)
-
-# Ejecutar la aplicación
-interface.launch()
+    return generated_image, seed
+
+examples = [
+    "a tiny astronaut hatching from an egg on the moon",
+    "a cat holding a sign that says hello world",
+    "an anime illustration of a wiener schnitzel",
+]
+
+css = """
+#col-container {
+    margin: 0 auto;
+    max-width: 520px;
+}
+"""
+
+with gr.Blocks(css=css) as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown(f"""# FLUX.1 [schnell] + Stable Diffusion img2img + Face Swap
+Combinación de generación de imágenes, transformación de imágenes con img2img, y Face Swap.
+        """)
+        
+        with gr.Row():
+            prompt = gr.Text(
+                label="Prompt",
+                show_label=False,
+                max_lines=1,
+                placeholder="Enter your prompt",
+                container=False,
+            )
+            run_button = gr.Button("Run", scale=0)
+        
+        result = gr.Image(label="Result", show_label=False)
+        
+        with gr.Accordion("Advanced Settings", open=False):
+            seed = gr.Slider(
+                label="Seed",
+                minimum=0,
+                maximum=MAX_SEED,
+                step=1,
+                value=0,
+            )
+            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+            width = gr.Slider(
+                label="Width",
+                minimum=256,
+                maximum=MAX_IMAGE_SIZE,
+                step=32,
+                value=1024,
+            )
+            height = gr.Slider(
+                label="Height",
+                minimum=256,
+                maximum=MAX_IMAGE_SIZE,
+                step=32,
+                value=1024,
+            )
+            num_inference_steps = gr.Slider(
+                label="Number of inference steps",
+                minimum=1,
+                maximum=50,
+                step=1,
+                value=4,
+            )
+            init_image = gr.Image(type="pil", label="Imagen Inicial (opcional)")
+            img2img_strength = gr.Slider(
+                label="Img2Img Strength",
+                minimum=0.0,
+                maximum=1.0,
+                step=0.05,
+                value=0.75,
+            )
+            reference_image = gr.Image(type="pil", label="Imagen de Referencia (opcional)")
+
+        gr.Examples(
+            examples=examples,
+            fn=infer,
+            inputs=[prompt],
+            outputs=[result, seed],
+            cache_examples="lazy"
+        )
+
+        gr.on(
+            triggers=[run_button.click, prompt.submit],
+            fn=infer,
+            inputs=[prompt, seed, randomize_seed, width, height, num_inference_steps, init_image, reference_image, img2img_strength],
+            outputs=[result, seed]
+        )
+
+demo.launch()