app.py

import gradio as gr
import torch
import spaces
from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, EulerDiscreteScheduler
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

assert torch.cuda.is_available()

device = "cuda"
dtype = torch.float16

base = "stabilityai/stable-diffusion-xl-base-1.0"
repo = "ByteDance/SDXL-Lightning"
opts = {
    "1 Step"  : ("sdxl_lightning_1step_unet_x0.safetensors", 1),
    "2 Steps" : ("sdxl_lightning_2step_unet.safetensors", 2),
    "4 Steps" : ("sdxl_lightning_4step_unet.safetensors", 4),
    "8 Steps" : ("sdxl_lightning_8step_unet.safetensors", 8),
}

step_loaded = None
# unet = UNet2DConditionModel.from_config(base, subfolder="unet").to(device, dtype)
# unet.load_state_dict(load_file(hf_hub_download(repo, opts["4 Steps"][0]), device=device))
pipe = StableDiffusionXLPipeline.from_pretrained(base, torch_dtype=dtype, variant="fp16").to(device)

@spaces.GPU(enable_queue=True)
def generate_image(prompt, option):
    global step_loaded
    ckpt, step = opts[option]
    if step != step_loaded:
        pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", prediction_type="sample" if step == 1 else "epsilon")
        pipe.unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device=device))
        step_loaded = step
    return pipe(prompt, num_inference_steps=step, guidance_scale=0).images[0]

with gr.Blocks() as demo:
    gr.HTML(
        "<h1><center>SDXL-Lightning</center></h1>" +
        "<p><center>Lightning-fast text-to-image generation</center></p>" +
        "<p><center><a href='https://huggingface.co/ByteDance/SDXL-Lightning'>https://huggingface.co/ByteDance/SDXL-Lightning</a></center></p>"
    )
    
    with gr.Row():
        prompt = gr.Textbox(
            label="Text prompt",
            scale=8
        )
        option = gr.Dropdown(
            label="Inference steps",
            choices=["1 Step", "2 Steps", "4 Steps", "8 Steps"],
            value="4 Steps",
            interactive=True
        )
        submit = gr.Button(
            scale=1,
            variant="primary"
        )
    
    img = gr.Image(label="SDXL-Lightening Generated Image")

    prompt.submit(
        fn=generate_image,
        inputs=[prompt, option],
        outputs=img,
    )
    submit.click(
        fn=generate_image,
        inputs=[prompt, option],
        outputs=img,
    )
    
demo.queue().launch()