app.py

import random
import spaces

import gradio as gr
import numpy as np
import torch
from diffusers import LCMScheduler, PixArtAlphaPipeline, Transformer2DModel
from peft import PeftModel
import os

device = "cuda" if torch.cuda.is_available() else "cpu"
IS_SPACE = os.environ.get("SPACE_ID", None) is not None

transformer = Transformer2DModel.from_pretrained(
    "PixArt-alpha/PixArt-XL-2-1024-MS",
    subfolder="transformer",
    torch_dtype=torch.float16,
)
transformer = PeftModel.from_pretrained(transformer, "jasperai/flash-pixart")


if torch.cuda.is_available():
    torch.cuda.max_memory_allocated(device=device)
    pipe = PixArtAlphaPipeline.from_pretrained(
        "PixArt-alpha/PixArt-XL-2-1024-MS",
        transformer=transformer,
        torch_dtype=torch.float16,
    )
    if not IS_SPACE:
        pipe.enable_xformers_memory_efficient_attention()
    pipe = pipe.to(device)
else:
    pipe = PixArtAlphaPipeline.from_pretrained(
        "PixArt-alpha/PixArt-XL-2-1024-MS",
        transformer=transformer,
        torch_dtype=torch.float16,
    )
    pipe = pipe.to(device)

pipe.text_encoder.to_bettertransformer()

pipe.scheduler = LCMScheduler.from_pretrained(
    "PixArt-alpha/PixArt-XL-2-1024-MS",
    subfolder="scheduler",
    timestep_spacing="trailing",
)

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1024
NUM_INFERENCE_STEPS = 4


@spaces.GPU
def infer(prompt, seed, randomize_seed):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    generator = torch.Generator().manual_seed(seed)

    image = pipe(
        prompt=prompt,
        guidance_scale=0,
        num_inference_steps=NUM_INFERENCE_STEPS,
        generator=generator,
        width=464,
        height=1024
    ).images[0]

    return image


examples = [
    "The image showcases a freshly baked bread, possibly focaccia, with rosemary sprigs and red pepper flakes sprinkled on top. It's sliced and placed on a wire cooling rack, with a bowl of mixed peppercorns beside it.",
    "A raccoon reading a book in a lush forest.",
    "A small cactus with a happy face in the Sahara desert.",
    "A super-realistic close-up of a snake eye",
    "A cute cheetah looking amazed and surprised",
    "Pirate ship sailing on a sea with the milky way galaxy in the sky and purple glow lights",
    "a cute fluffy rabbit pilot walking on a military aircraft carrier, 8k, cinematic",
    "A close up of an old elderly man with green eyes looking straight at the camera",
    "A beautiful sunflower in rainy day",
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 1080px;
}
"""

if torch.cuda.is_available():
    power_device = "GPU"
else:
    power_device = "CPU"

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        with gr.Row():
            with gr.Column():
                gr.Markdown(
                    f"""
                # ⚡ Flash Diffusion for Phone Wallpapers ⚡
                This space version will generate 464x1024 images for your smartphones :)
                
                This is an interactive demo of [Flash Diffusion](https://gojasper.github.io/flash-diffusion-project/), a diffusion distillation method proposed in [Flash Diffusion: Accelerating Any Conditional
                Diffusion Model for Few Steps Image Generation](http://arxiv.org/abs/2406.02347) *by Clément Chadebec, Onur Tasar, Eyal Benaroche and Benjamin Aubin* from Jasper Research.
                [This model](https://huggingface.co/jasperai/flash-pixart) is a **66.5M** LoRA distilled version of [Pixart-α](https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS) model that is able to generate 1024x1024 images in **4 steps**. 
                Currently running on {power_device}.
                """
                )
                gr.Markdown(
                    "If you enjoy the space, please also promote *open-source* by giving a ⭐ to the <a href='https://github.com/gojasper/flash-diffusion' target='_blank'>Github Repo</a>."
                )
                gr.Markdown(
                    "💡 *Hint:* To better appreciate the low latency of our method, run the demo locally !"
                )

                with gr.Row():
                    prompt = gr.Text(
                        label="Prompt",
                        show_label=False,
                        max_lines=1,
                        placeholder="Enter your prompt",
                        container=False,
                    )

                    run_button = gr.Button("Run", scale=0)

                with gr.Accordion("Advanced Settings", open=False):
                    seed = gr.Slider(
                        label="Seed",
                        minimum=0,
                        maximum=MAX_SEED,
                        step=1,
                        value=0,
                    )

                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
                
                examples = gr.Examples(examples=examples, inputs=[prompt])

                gr.Markdown("**Disclaimer:**")
                gr.Markdown(
                    "This demo is only for research purpose. Jasper cannot be held responsible for the generation of NSFW (Not Safe For Work) content through the use of this demo. Users are solely responsible for any content they create, and it is their obligation to ensure that it adheres to appropriate and ethical standards. Jasper provides the tools, but the responsibility for their use lies with the individual user."
                )

            result = gr.Image(label="Result", show_label=False, format="png")

        

    gr.on(
        [run_button.click, seed.change, randomize_seed.change, prompt.submit],
        fn=infer,
        inputs=[prompt, seed, randomize_seed],
        outputs=[result],
        show_progress="minimal",
        show_api=False,
        trigger_mode="always_last",
    )

demo.queue().launch(show_api=False)