app.py

import gradio as gr
import torch
from diffusers import DiffusionPipeline, UNet2DConditionModel, LCMScheduler
from huggingface_hub import hf_hub_download
import spaces
from PIL import Image
import requests
from translatepy import Translator

translator = Translator()

# Constants
base = "stabilityai/stable-diffusion-xl-base-1.0"
repo = "tianweiy/DMD2"
checkpoints = {
    "1-Step" : ["dmd2_sdxl_1step_unet_fp16.bin", 1],
    "4-Step" : ["dmd2_sdxl_4step_unet_fp16.bin", 4],
}
loaded = None

CSS = """
.gradio-container {
  max-width: 690px !important;
}
"""

# Ensure model and scheduler are initialized in GPU-enabled function
if torch.cuda.is_available():
    unet = UNet2DConditionModel.from_config(base, subfolder="unet").to("cuda", torch.float16)
    pipe = DiffusionPipeline.from_pretrained(base, torch_dtype=torch.float16, variant="fp16").to("cuda")


# Function 
@spaces.GPU()
def generate_image(prompt, ckpt="4-Step"):
    global loaded
    
    prompt = str(translator.translate(prompt, 'English'))

    print(prompt)
    
    checkpoint = checkpoints[ckpt][0]
    num_inference_steps = checkpoints[ckpt][1]

    if loaded != num_inference_steps:
        pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")
        pipe.unet.load_state_dict(torch.load(hf_hub_download(repo, checkpoint), map_location="cuda"))
        loaded = num_inference_steps

    if loaded == 1:
        timesteps=[399]
    else:
        timesteps=[999, 749, 499, 249]

    results = pipe(prompt, num_inference_steps=num_inference_steps, guidance_scale=0, timesteps=timesteps)
    return results.images[0]


examples = [
    "a cat eating a piece of cheese",
    "a ROBOT riding a BLUE horse on Mars, photorealistic",
    "Ironman VS Hulk, ultrarealistic",
    "a CUTE robot artist painting on an easel",
    "Astronaut in a jungle, cold color palette, oil pastel, detailed, 8k",
    "An alien holding sign board contain word 'Flash', futuristic, neonpunk",
    "Kids going to school, Anime style"
]


# Gradio Interface

with gr.Blocks(css=CSS, theme="soft") as demo:
    gr.HTML("<h1><center>Adobe DMD2🦖</center></h1>")
    gr.HTML("<p><center><a href='https://huggingface.co/tianweiy/DMD2'>DMD2</a> text-to-image generation</center><br><center>Multi-Languages, 4-step is higher quality & 2X slower</center></p>")
    with gr.Group():
        with gr.Row():
            prompt = gr.Textbox(label='Enter Your Prompt', scale=8)
            ckpt = gr.Dropdown(label='Steps',choices=['1-Step', '4-Step'], value='4-Step', interactive=True)
            submit = gr.Button(scale=1, variant='primary')
    img = gr.Image(label='DMD2 Generated Image')    
    gr.Examples(
        examples=examples,
        inputs=prompt,
        outputs=img,
        fn=generate_image,
        cache_examples="lazy",
    )

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