app.py

# Reference: https://huggingface.co./spaces/FoundationVision/LlamaGen/blob/main/app.py
from PIL import Image
import gradio as gr
from imagenet_classes import imagenet_idx2classname
import torch
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
import time
import argparse
import demo_util
import os

device = "cuda" if torch.cuda.is_available() else "cpu"


model2ckpt = {
    "TiTok-L-32": ("tokenizer_titok_l32.bin", "generator_titok_l32.bin"),
}

if not os.path.exists("tokenizer_titok_l32.bin"):
    os.system("gdown 1I_m2Vm4JgQsa7bZVORj-nVhP8fgQLngd")
if not os.path.exists("generator_titok_l32.bin"):
    os.system("gdown 1IgqZ_vwGIj2ZWOPuCzilxeQ2UrMVY93l")

parser = argparse.ArgumentParser()
parser.add_argument("--precision", type=str, default='bf16', choices=["none", "fp16", "bf16"]) 
parser.add_argument("--guidance_scale", type=float, default=3.5)
parser.add_argument("--randomize_temperature", type=float, default=1.0)
parser.add_argument("--num_sample_steps", type=int, default=8)
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--temperature", type=float, default=1.0, help="temperature value to sample with")
args = parser.parse_args()

config = demo_util.get_config("configs/titok_l32.yaml")
print(config)
titok_tokenizer = demo_util.get_titok_tokenizer(config)
print(titok_tokenizer)
titok_generator = demo_util.get_titok_generator(config)
print(titok_generator)

titok_tokenizer = titok_tokenizer.to(device)
titok_generator = titok_generator.to(device)


def demo_infer(guidance_scale, randomize_temperature, num_sample_steps,
               class_label, seed):
    n = 4
    class_labels = [class_label for _ in range(n)]
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    t1 = time.time()
    generated_image = demo_util.sample_fn(
        generator=titok_generator,
        tokenizer=titok_tokenizer,
        labels=class_labels,
        guidance_scale=guidance_scale,
        randomize_temperature=randomize_temperature,
        num_sample_steps=num_sample_steps,
        device=device
    )
    sampling_time = time.time() - t1
    print(f"generation takes about {sampling_time:.2f} seconds.")    
    samples = [Image.fromarray(sample) for sample in generated_image]
    return samples


with gr.Blocks() as demo:
    gr.Markdown("<h1 style='text-align: center'>An Image is Worth 32 Tokens for Reconstruction and Generation</h1>")

    with gr.Tabs():
        with gr.TabItem('Generate'):
            with gr.Row():
                with gr.Column():
                    with gr.Row():
                        i1k_class = gr.Dropdown(
                            list(imagenet_idx2classname.values()),
                            value='macaw',
                            type="index", label='ImageNet-1K Class'
                        )
                    guidance_scale = gr.Slider(minimum=1, maximum=25, step=0.1, value=3.5, label='Classifier-free Guidance Scale')
                    randomize_temperature = gr.Slider(minimum=0., maximum=10.0, step=0.1, value=1.0, label='randomize_temperature')
                    num_sample_steps = gr.Slider(minimum=1, maximum=32, step=1, value=8, label='num_sample_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(demo_infer, inputs=[
                        guidance_scale, randomize_temperature, num_sample_steps,
                        i1k_class, seed],
                        outputs=[output])
    demo.queue()
    demo.launch(debug=True)