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")
pipe.enable_model_cpu_offload()

@spaces.GPU
def generate_image(prompt, height, width, num_inference_steps, guidance_scale):
    seed = random.randint(0, 18446744073709551615)
    image = pipe(
        prompt=prompt,
        height=height,
        width=width,
        num_inference_steps=num_inference_steps,
        guidance_scale=guidance_scale,
        num_images_per_prompt=1,
        generator=torch.Generator(pipe.device).manual_seed(seed)
    ).images[0]
    return image, seed

# Gradio interface
iface = gr.Interface(
    fn=generate_image,
    inputs=[
        gr.Textbox(label="Prompt"),
        gr.Slider(512, 1344, 1024, step=64, label="Height"),
        gr.Slider(512, 1344, 1024, step=64, label="Width"),
        gr.Slider(20, 100, 20, step=1, label="Number of Inference Steps"),
        gr.Slider(1, 20, 5, step=0.5, label="Guidance Scale"),
    ],
    outputs=[
        gr.Image(label="Generated Image"),
        gr.Number(label="Seed")
    ],
    title="Kolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis",
    theme='bethecloud/storj_theme',
)

iface.launch(debug=True)