app.py

import argparse
import os
from PIL import Image
import gradio as gr
import spaces
from imagenet_en_cn import IMAGENET_1K_CLASSES
from omegaconf import OmegaConf
from huggingface_hub import snapshot_download

import torch
from transformers import T5EncoderModel, AutoTokenizer

from pixelflow.scheduling_pixelflow import PixelFlowScheduler
from pixelflow.pipeline_pixelflow import PixelFlowPipeline
from pixelflow.utils import config as config_utils
from pixelflow.utils.misc import seed_everything


parser = argparse.ArgumentParser(description='Gradio Demo', add_help=False)
parser.add_argument('--checkpoint', type=str, help='checkpoint folder path')
parser.add_argument('--class_cond', action='store_true', help='use class conditional generation')
args = parser.parse_args()

# deploy
args.checkpoint = "pixelflow_t2i"
args.class_cond = False

output_dir = args.checkpoint

if args.class_cond:
    if not os.path.exists(output_dir):
        snapshot_download(repo_id="ShoufaChen/PixelFlow-Class2Image", local_dir=output_dir)
    config = OmegaConf.load(f"{output_dir}/config.yaml")
    model = config_utils.instantiate_from_config(config.model)
    print(f"Num of parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
    ckpt = torch.load(f"{output_dir}/model.pt", map_location="cpu", weights_only=True)
    text_encoder = None
    tokenizer = None
    resolution = 256
    NUM_EXAMPLES = 4
else:
    if not os.path.exists(output_dir):
        snapshot_download(repo_id="ShoufaChen/PixelFlow-Text2Image", local_dir=output_dir)
    config = OmegaConf.load(f"{output_dir}/config.yaml")
    model = config_utils.instantiate_from_config(config.model)
    print(f"Num of parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}")
    ckpt = torch.load(f"{output_dir}/model.pt", map_location="cpu", weights_only=True)
    text_encoder = T5EncoderModel.from_pretrained("google/flan-t5-xl")
    tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-xl")
    resolution = 1024
    NUM_EXAMPLES = 1
model.load_state_dict(ckpt, strict=True)
model.eval()

print(f"outside space.GPU. {torch.cuda.is_available()=}")
if torch.cuda.is_available():
    model = model.cuda()
    text_encoder = text_encoder.cuda() if text_encoder else None
    device = torch.device("cuda")
else:
    raise ValueError("No GPU")

scheduler = PixelFlowScheduler(config.scheduler.num_train_timesteps, num_stages=config.scheduler.num_stages, gamma=-1/3)

pipeline = PixelFlowPipeline(
    scheduler,
    model,
    text_encoder=text_encoder,
    tokenizer=tokenizer,
    max_token_length=512,
)

@spaces.GPU(duration=120) 
def infer(noise_shift, cfg_scale, class_label, seed, *num_steps_per_stage):
    print(f"inside space.GPU. {torch.cuda.is_available()=}")
    seed_everything(seed)
    with torch.autocast("cuda", dtype=torch.bfloat16), torch.no_grad():
        samples = pipeline(
            prompt=[class_label] * NUM_EXAMPLES,
            height=resolution,
            width=resolution,
            num_inference_steps=list(num_steps_per_stage),
            guidance_scale=cfg_scale,         # The guidance for the first frame, set it to 7 for 384p variant
            device=device,
            shift=noise_shift,
            use_ode_dopri5=False,
        )
    samples = (samples * 255).round().astype("uint8")
    samples = [Image.fromarray(sample) for sample in samples]
    return samples


css = """
h1 {
    text-align: center;
    display: block;
}

.follow-link {
    margin-top: 0.8em;
    font-size: 1em;
    text-align: center;
}
"""


with gr.Blocks(css=css) as demo:
    gr.Markdown("# PixelFlow: Pixel-Space Generative Models with Flow")
    gr.HTML("""
        <div class="follow-link">
            For online class-to-image generation, please try
            <a href="https://huggingface.co/spaces/ShoufaChen/PixelFlow">class-to-image</a>.
            For more details, refer to our 
                <a href="https://arxiv.org/abs/2504.07963">arXiv paper</a> and <a href="https://github.com/ShoufaChen/PixelFlow">GitHub repo</a>.
        </div>
    """)

    with gr.Tabs():
        with gr.TabItem('Generate'):
            with gr.Row():
                with gr.Column():
                    with gr.Row():
                        if args.class_cond:
                            user_input = gr.Dropdown(
                                list(IMAGENET_1K_CLASSES.values()),
                                value='daisy [雏菊]',
                                type="index", label='ImageNet-1K Class'
                            )
                        else:
                            # text input
                            user_input = gr.Textbox(label='Enter your prompt', show_label=False, max_lines=1, placeholder="Enter your prompt",)
                    noise_shift = gr.Slider(minimum=1.0, maximum=100.0, step=1, value=1.0, label='Noise Shift')
                    cfg_scale = gr.Slider(minimum=1, maximum=25, step=0.1, value=4.0, label='Classifier-free Guidance Scale')
                    num_steps_per_stage = []
                    for stage_idx in range(config.scheduler.num_stages):
                        num_steps = gr.Slider(minimum=1, maximum=100, step=1, value=5, label=f'Num Inference Steps (Stage {stage_idx})')
                        num_steps_per_stage.append(num_steps)
                    seed = gr.Slider(minimum=0, maximum=1000, step=1, value=42, label='Seed')
                    button = gr.Button("Generate", variant="primary")
                with gr.Column():
                    output = gr.Gallery(label='Generated Images', height=700)
                    button.click(infer, inputs=[noise_shift, cfg_scale, user_input, seed, *num_steps_per_stage], outputs=[output])
    demo.queue()
    demo.launch(share=True, debug=True)