Create app.py

app.py

@@ -0,0 +1,85 @@
+import cv2
+import torch
+import numpy as np
+import gradio as gr
+from diffusers import StableDiffusionXLPipeline, ControlNetModel
+from insightface.app import FaceAnalysis
+from accelerate import init_empty_weights, load_checkpoint_and_dispatch
+import os
+
+# Force offline mode to avoid runtime Hub connections
+os.environ["HF_HUB_OFFLINE"] = "1"
+
+# Set device to CPU (free tier has no GPU)
+device = "cpu"
+dtype = torch.float32
+
+# Load face encoder
+face_app = FaceAnalysis(providers=["CPUExecutionProvider"])
+face_app.prepare(ctx_id=0, det_size=(480, 480))
+
+# Define paths for preloaded weights
+controlnet_path = "./ControlNetModel"
+face_adapter_path = "./ip-adapter.bin"
+
+# Check if files exist
+if not os.path.exists(controlnet_path) or not os.path.exists(os.path.join(controlnet_path, "config.json")):
+    raise FileNotFoundError(f"ControlNetModel directory or config.json not found at {controlnet_path}")
+if not os.path.exists(face_adapter_path):
+    raise FileNotFoundError(f"ip-adapter.bin not found at {face_adapter_path}")
+
+# Initialize models with empty weights
+with init_empty_weights():
+    controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=dtype)
+    pipe = StableDiffusionXLPipeline.from_pretrained(
+        "stabilityai/stable-diffusion-xl-base-1.0",
+        controlnet=controlnet,
+        torch_dtype=dtype,
+        safety_checker=None,
+    )
+
+# Load and dispatch models with accelerate
+controlnet = load_checkpoint_and_dispatch(controlnet, controlnet_path, device_map="cpu", offload_folder=None)
+pipe = load_checkpoint_and_dispatch(pipe, "./", device_map="cpu", offload_folder=None)
+pipe.load_ip_adapter_instantid(face_adapter_path)
+
+def generate_image(uploaded_image, prompt):
+    # Convert Gradio image to OpenCV format
+    img = cv2.cvtColor(np.array(uploaded_image), cv2.COLOR_RGB2BGR)
+    faces = face_app.get(img)
+    if not faces:
+        return "No face detected!", None
+
+    face_info = faces[-1]  # Use largest face
+    face_emb = face_info["embedding"]
+
+    try:
+        image = pipe(
+            prompt=prompt,
+            image_embeds=face_emb,
+            num_inference_steps=20,
+            guidance_scale=7.5,
+            height=512,
+            width=512,
+            controlnet_conditioning_scale=1.0,
+        ).images[0]
+        return "Image generated successfully!", image
+    except Exception as e:
+        return f"Generation failed: {e}", None
+
+# Gradio interface
+interface = gr.Interface(
+    fn=generate_image,
+    inputs=[
+        gr.Image(type="pil", label="Upload Reference Image"),
+        gr.Textbox(label="Enter Prompt", placeholder="e.g., A photorealistic astronaut in space")
+    ],
+    outputs=[
+        gr.Textbox(label="Status"),
+        gr.Image(label="Generated Image")
+    ],
+    title="Face Reference Image Generator",
+    description="Upload an image with a face, enter a prompt, and generate a new image preserving the reference face."
+)
+
+interface.launch()