Update app.py

app.py

@@ -1,19 +1,47 @@
-import torch
-import spaces
 import gradio as gr
-from diffusers import StableDiffusionXLPipeline
+import torch
+from diffusers import StableDiffusionPipeline, DDIMScheduler
+from ip_adapter.ip_adapter_faceid import IPAdapterFaceID
+from insightface.app import FaceAnalysis
+import cv2
+import numpy as np
+
+# Configuración inicial
+device = "cuda"
+model_id = "runwayml/stable-diffusion-v1-5"
+ip_ckpt = "ip-adapter-faceid_sd15.bin"
+
+# Cargar el pipeline
+pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to(device)
+app = FaceAnalysis(name="buffalo_l", providers=['CUDAExecutionProvider'])
+app.prepare(ctx_id=0, det_size=(640, 640))
+
+# Cargar IP-Adapter
+ip_model = IPAdapterFaceID(pipe, ip_ckpt, device)
 
-# Usamos un modelo SDXL base que es extremadamente confiable y tiene todos sus archivos
-model_id = "stabilityai/stable-diffusion-xl-base-1.0"
-pipe = StableDiffusionXLPipeline.from_pretrained(
-    model_id, torch_dtype=torch.float16, use_safetensors=True, variant="fp16"
-).to("cuda")
+def generate_image(input_image, prompt):
+    # Procesar cara
+    img_cv = cv2.cvtColor(np.array(input_image), cv2.COLOR_RGB2BGR)
+    faces = app.get(img_cv)
+    if not faces:
+        return None
+    
+    # Generar
+    image = ip_model.generate(
+        pil_image=input_image,
+        face_embed=faces[0].embedding,
+        prompt=prompt,
+        width=512,
+        height=512
+    )
+    return image[0]
 
-@spaces.GPU
-def generate(prompt):
-    # La clave es la consistencia, SDXL con un buen prompt es muy estable
-    image = pipe(prompt=prompt, num_inference_steps=25).images[0]
-    return image
+# Interfaz de Gradio
+demo = gr.Interface(
+    fn=generate_image,
+    inputs=[gr.Image(type="pil"), gr.Textbox(label="Prompt")],
+    outputs="image",
+    title="IP-Adapter FaceID Generator"
+)
 
-demo = gr.Interface(fn=generate, inputs="text", outputs="image")
 demo.launch()