app.py

import random
import os

import spaces
import numpy as np
import torch
from PIL import Image
import huggingface_hub
import gradio as gr

from src.pipeline_flux_nag import NAGFluxPipeline
from src.transformer_flux import NAGFluxTransformer2DModel


MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 2048


theme = gr.themes.Base(
    font=[gr.themes.GoogleFont('Libre Franklin'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'],
)

transformer = NAGFluxTransformer2DModel.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)
pipe = NAGFluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
)

device = "cuda"
pipe = pipe.to(device)

examples = [
    ["Portrait of AI researcher.", "Glasses.", 5],
    ["Portrait of AI researcher.", "Male.", 5],
    ["A baby phoenix made of fire and flames is born from the smoking ashes.", "Low resolution, blurry, lack of details, illustration, cartoon, painting.", 5],
    ["A tiny astronaut hatching from an egg on the moon.", "Low resolution, blurry, lack of details, illustration, cartoon, painting.", 9]
]

@spaces.GPU
def sample(
        prompt,
        negative_prompt=None, guidance_scale=3.5,
        nag_negative_prompt=None, nag_scale=5.0,
        width=1024, height=1024,
        num_inference_steps=25,
        seed=2025, randomize_seed=False,
        compare=True,
):
    prompt = prompt.strip()
    negative_prompt = negative_prompt.strip() if negative_prompt and negative_prompt.strip() else None
    guidance_scale = float(guidance_scale)
    width, height = int(width), int(height)
    num_inference_steps = int(num_inference_steps)
    
    if (randomize_seed):
        seed = random.randint(0, MAX_SEED)
    else:
        seed = int(seed)

    generator = torch.Generator(device="cuda").manual_seed(seed)
    image_nag = pipe(
        prompt,
        negative_prompt=negative_prompt,
        guidance_scale=guidance_scale,
        nag_negative_prompt=nag_negative_prompt,
        nag_scale=nag_scale,
        generator=generator,
        width=width,
        height=height,
        num_inference_steps=num_inference_steps,
    ).images[0]

    if compare:
        generator = torch.Generator(device="cuda").manual_seed(seed)
        image_normal = pipe(
            prompt,
            negative_prompt=negative_prompt,
            guidance_scale=guidance_scale,
            generator=generator,
            width=width,
            height=height,
            num_inference_steps=num_inference_steps,
        ).images[0]
    else:
        image_normal = Image.new("RGB", image_nag.size, color=(0, 0, 0))

    return (image_normal, image_nag), seed


def sample_example(
        prompt,
        nag_negative_prompt,
        nag_scale,
):
    outputs, seed = sample(
        prompt=prompt,
        nag_negative_prompt=nag_negative_prompt,
        nag_scale=nag_scale,
    )
    return outputs, 3.5, 1024, 1024, 25, seed, True


css = '''
.gradio-container{
max-width: 768px !important;
margin: 0 auto;
}
'''

with gr.Blocks(css=css, theme=theme) as demo:
    gr.Markdown('''# Normalized Attention Guidance (NAG) Flux-Dev
    Implementation of [Normalized Attention Guidance](https://chendaryen.github.io/NAG.github.io/)
    ''')
    with gr.Group():
        prompt = gr.Textbox(
            label="Prompt",
            max_lines=1,
            placeholder="Enter your prompt",
        )
        nag_negative_prompt = gr.Textbox(
            label="Negative Prompt for NAG",
            value="Low resolution, blurry, lack of details, illustration, cartoon, painting.",
            max_lines=1,
        )
        nag_scale = gr.Slider(label="NAG Scale", minimum=1., maximum=20., step=0.25, value=5.)
        compare = gr.Checkbox(label="Compare with baseline", info="If unchecked, only sample with NAG will be generated.", value=True)
        button = gr.Button("Generate", min_width=120)
        output = gr.ImageSlider(label="Left: Baseline, Right: With NAG", interactive=False)
        with gr.Accordion("Advanced Settings", open=False):
            negative_prompt = gr.Textbox(label="Negative Prompt", value=None, visible=False)
            guidance_scale = gr.Slider(label="Guidance Scale", minimum=1., maximum=15., step=0.1, value=3.5)
            with gr.Row():
                width = gr.Slider(
                    label="Width",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,
                )
                height = gr.Slider(
                    label="Height",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=32,
                    value=1024,
                )
            num_inference_steps = gr.Slider(label="Inference Steps", minimum=1, maximum=50, step=1, value=25)
            seed = gr.Slider(label="Seed", minimum=1, maximum=MAX_SEED, step=1, randomize=True)
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

        gr.Examples(
            examples=examples,
            fn=sample_example,
            inputs=[
                prompt,
                nag_negative_prompt,
                nag_scale,
            ],
            outputs=[output, guidance_scale, width, height, num_inference_steps, seed, compare],
            cache_examples="lazy",
        )

    gr.on(
        triggers=[
            button.click,
            prompt.submit
        ],
        fn=sample,
        inputs=[
            prompt,
            negative_prompt, guidance_scale,
            nag_negative_prompt, nag_scale,
            width, height,
            num_inference_steps,
            seed, randomize_seed,
            compare,
        ],
        outputs=[output, seed],
    )
if __name__ == "__main__":
    huggingface_hub.login(os.getenv('HF_TOKEN'))
    demo.launch(share=True)