app.py

import random

import gradio as gr
import numpy as np
import spaces
import torch
from diffusers import DiffusionPipeline
from PIL import Image

device = "cuda" if torch.cuda.is_available() else "cpu"
repo_id = "black-forest-labs/FLUX.1-dev"
adapter_id = "alvarobartt/ghibli-characters-flux-lora"

pipeline = DiffusionPipeline.from_pretrained(repo_id, torch_dtype=torch.bfloat16)
pipeline.load_lora_weights(adapter_id)
pipeline = pipeline.to(device)


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


@spaces.GPU(duration=120)
def inference(
    prompt: str,
    seed: int,
    randomize_seed: bool,
    width: int,
    height: int,
    guidance_scale: float,
    num_inference_steps: int,
    lora_scale: float,
    progress: gr.Progress = gr.Progress(track_tqdm=True),
):
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)
    
    image = pipeline(
        prompt=prompt,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        width=width,
        height=height,
        generator=generator,
        joint_attention_kwargs={"scale": lora_scale},
    ).images[0]

    return image, seed


examples = [
    (
        "Ghibli style futuristic stormtrooper with glossy white armor and a sleek helmet,"
        " standing heroically on a lush alien planet, vibrant flowers blooming around, soft"
        " sunlight illuminating the scene, a gentle breeze rustling the leaves"
    ),
]

css = """
#col-container {
    margin: 0 auto;
    max-width: 640px;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("# FLUX.1 Studio Ghibli LoRA")
        gr.Markdown(
            "LoRA fine-tune of [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev)"
            " with [alvarobartt/ghibli-characters](https://huggingface.co/datasets/alvarobartt/ghibli-characters)."
        )

        with gr.Accordion("How to generate nice prompts?"):
            gr.Markdown(
                "To generate high-quality prompts what worked the best for me was to prompt"
                " Anthropic Claude with the following: \"You are an expert prompt writer for"
                " diffusion models, and you've been provided the following prompt template:\n"
                "```text\nGhibli style [character description] with [distinctive features],"
                " [action or pose], [environment or background], [lighting or atmosphere],"
                " [additional details]\n```\n\nCould you generate a prompt following that template"
                " to generate [CHARACTER NAME] as a Studio Ghibli character? [MORE DETAILS IF NEEDED]\""
            )

        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)

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

        with gr.Accordion("Advanced Settings", open=False):
            seed = gr.Slider(
                label="Seed",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=42,
            )

            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)

            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=768,
                )

            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance scale",
                    minimum=0.0,
                    maximum=10.0,
                    step=0.1,
                    value=3.5,
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=30,
                )

                lora_scale = gr.Slider(
                    label="LoRA scale",
                    minimum=0.0,
                    maximum=1.0,
                    step=0.1,
                    value=1.0,
                )

        gr.Examples(
            examples=examples,
            fn=lambda x: (Image.open("./example.jpg"), 42),
            inputs=[prompt],
            outputs=[result, seed],
            run_on_click=True,
        )

        gr.Markdown(
            "### Disclaimer\n\n"
            "License is non-commercial for both FLUX.1-dev and the Studio Ghibli dataset;"
            " but free to use for personal and non-commercial purposes."
        )

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=inference,
        inputs=[
            prompt,
            seed,
            randomize_seed,
            width,
            height,
            guidance_scale,
            num_inference_steps,
            lora_scale,
        ],
        outputs=[result, seed],
    )

demo.queue()
demo.launch()