app.py

import spaces
import os
import torch
import random
from huggingface_hub import snapshot_download
from kolors.pipelines.pipeline_stable_diffusion_xl_chatglm_256 import StableDiffusionXLPipeline
from kolors.models.modeling_chatglm import ChatGLMModel
from kolors.models.tokenization_chatglm import ChatGLMTokenizer
from diffusers import UNet2DConditionModel, AutoencoderKL
from diffusers import EulerDiscreteScheduler
import gradio as gr

# Download the model files
ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")

# Load the models
text_encoder = ChatGLMModel.from_pretrained(
    os.path.join(ckpt_dir, 'text_encoder'),
    torch_dtype=torch.float16).half()
tokenizer = ChatGLMTokenizer.from_pretrained(os.path.join(ckpt_dir, 'text_encoder'))
vae = AutoencoderKL.from_pretrained(os.path.join(ckpt_dir, "vae"), revision=None).half()
scheduler = EulerDiscreteScheduler.from_pretrained(os.path.join(ckpt_dir, "scheduler"))
unet = UNet2DConditionModel.from_pretrained(os.path.join(ckpt_dir, "unet"), revision=None).half()

pipe = StableDiffusionXLPipeline(
        vae=vae,
        text_encoder=text_encoder,
        tokenizer=tokenizer,
        unet=unet,
        scheduler=scheduler,
        force_zeros_for_empty_prompt=False)
pipe = pipe.to("cuda")

@spaces.GPU(duration=200)
def generate_image(prompt, negative_prompt, height, width, num_inference_steps, guidance_scale, num_images_per_prompt, use_random_seed, seed, progress=gr.Progress(track_tqdm=True)):
    if use_random_seed:
        seed = random.randint(0, 2**32 - 1)
    else:
        seed = int(seed)  # Ensure seed is an integer
    
    image = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        height=height,
        width=width,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
        num_images_per_prompt=num_images_per_prompt,
        generator=torch.Generator(pipe.device).manual_seed(seed)
    ).images
    return image, seed

description = """
<p align="center">Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis</p>
<p><center>
<a href="https://kolors.kuaishou.com/" target="_blank">[Official Website]</a>
<a href="https://github.com/Kwai-Kolors/Kolors/blob/master/imgs/Kolors_paper.pdf" target="_blank">[Tech Report]</a>
<a href="https://huggingface.co/Kwai-Kolors/Kolors" target="_blank">[Model Page]</a>
<a href="https://github.com/Kwai-Kolors/Kolors" target="_blank">[Github]</a>
</center></p>
"""

# Replace your current Gradio interface with this:
with gr.Blocks(theme='bethecloud/storj_theme', title="Kolors") as iface:
    gr.Markdown(description)
    
    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(label="Prompt")
            negative_prompt = gr.Textbox(label="Negative Prompt")
            
            with gr.Accordion("Advanced Settings", open=False):
                height = gr.Slider(512, 2048, 1024, step=64, label="Height")
                width = gr.Slider(512, 2048, 1024, step=64, label="Width")
                num_inference_steps = gr.Slider(20, 50, 20, step=1, label="Number of Inference Steps")
                guidance_scale = gr.Slider(1, 20, 5, step=0.5, label="Guidance Scale")
                num_images_per_prompt = gr.Slider(1, 4, 1, step=1, label="Number of images per prompt")
                use_random_seed = gr.Checkbox(label="Use Random Seed", value=True)
                seed = gr.Number(label="Seed", value=0, precision=0)
            
            generate_btn = gr.Button("Generate")
        
        with gr.Column():
            gallery = gr.Gallery(label="Result", show_label=False)
            seed_used = gr.Number(label="Seed Used")
    
    generate_btn.click(
        fn=generate_image,
        inputs=[prompt, negative_prompt, height, width, num_inference_steps, 
                guidance_scale, num_images_per_prompt, use_random_seed, seed],
        outputs=[gallery, seed_used]
    )

# Fix the launch call
iface.launch(debug=True, share=True)