import gradio as gr
from PIL import Image
import base64
import io
import cv2
import numpy as np
import torch
from controlnet_aux import HEDdetector
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler

def predict(sketch, description):
    # Convert sketch to PIL image
    sketch_pil = Image.fromarray(sketch)

    hed = HEDdetector.from_pretrained('lllyasviel/Annotators')

    image = hed(sketch_pil, scribble=True)

    model_id = "runwayml/stable-diffusion-v1-5"
    controlnet_id = "lllyasviel/sd-controlnet-scribble"

    # Load ControlNet model
    controlnet = ControlNetModel.from_pretrained(controlnet_id, torch_dtype=torch.float16)

    # Create pipeline with ControlNet model
    pipe = StableDiffusionControlNetPipeline.from_pretrained(model_id, controlnet=controlnet, torch_dtype=torch.float16)

    # Use improved scheduler
    pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)

    # Enable smart CPU offloading and memory efficient attention
    # pipe.enable_model_cpu_offload()
    pipe.enable_xformers_memory_efficient_attention()


    result = pipe(description, image, num_inference_steps=20).images[0]
    
    return result
# Define sketchpad with custom size and stroke width
sketchpad = gr.Sketchpad(shape=(1024, 1024), brush_radius=5)

iface = gr.Interface(fn=predict, inputs=[sketchpad, "text"], outputs="image", live=False)
iface.launch(share=True)