import gradio as gr
from diffusers import StableDiffusionPipeline, ControlNetModel, StableDiffusionControlNetPipeline
from diffusers.utils import load_image
import torch
import cv2
import numpy as np
from PIL import Image

is_show_controlnet = True
prompts = ""
neg_prompt = "chinese letter"

controlnet_repo_id = "calihyper/trad-kor-controlnet"
repo_id = "calihyper/trad-kor-landscape-black"
controlnet = ControlNetModel.from_pretrained(controlnet_repo_id)
pipe = StableDiffusionControlNetPipeline.from_pretrained(repo_id, controlnet=controlnet).to("cuda")


def change_radio(input):
    return input

def output_radio(output):
    print(output)

def predict(canny, lt, ht, prompt, style_prompt, neg_prompt, ins, gs, seed):
    np_image = np.array(canny)
    low_threshold = lt
    high_threshold = ht
    np_image = cv2.Canny(np_image, low_threshold, high_threshold)
    np_image = np_image[:, :, None]
    np_image = np.concatenate([np_image, np_image, np_image], axis=2)
    canny_image = Image.fromarray(np_image)

    generator = torch.manual_seed(seed)

    global pipe

    output = pipe(
        prompt + style_prompt,
        canny_image,
        negative_prompt=neg_prompt,
        generator=generator,
        num_inference_steps=ins,
        guidance_scale=gs
    )
    return output.images[0]

with gr.Blocks() as demo:
    gr.Markdown("# Aiffelthon Choosa Project")

    with gr.Row():
        with gr.Column() as controlnet:
            
            canny_image = gr.Image(label="input_image", visible=is_show_controlnet , shape=(512,512), interactive=True)
        
            controlnet_radio = gr.Radio([True, False], label="Use ControlNet")
            lt = gr.Slider(50, 300, 120, step=1, label="Low threshold")
            ht = gr.Slider(50, 300, 120, step=1, label="High threshold")        

        with gr.Column():
            out_image = gr.Image()
            with gr.Column() as diff:
                prompt = gr.Textbox(placeholder="prompts", label="prompt")
                style_prompt = gr.Textbox(placeholder="style prompts", label="style prompt")
                examples = gr.Examples(examples=["<trad-kor-landscape-black>", "<trad-kor-landscape-ink-wash-painting>", "<trad-kor-landscape-thick-brush-strokes>", "<trad-kor-plants-black>", "<trad-kor-plants-color>"], 
                                       inputs=style_prompt,  label="style examples")

                neg_prompt = gr.Textbox(placeholder="negative prompts", value=neg_prompt, label="negative prompt")

                ins = gr.Slider(1, 60, 30, label="inference steps")
                gs = gr.Slider(1, 10, 2.5, step=1, label="guidance scale")

                seed = gr.Slider(0, 10, 2, step=1, label="seed")
    btn1 = gr.Button("실행")
    btn1.click(predict, [canny_image, lt, ht, prompt,style_prompt, neg_prompt, ins, gs, seed ], out_image)

if __name__ == "__main__":
    demo.launch()