app.py

import os
import random

import gradio as gr
import numpy as np
import spaces
import torch
from diffusers import Flux2KleinPipeline
from PIL import Image

BASE_MODEL = "black-forest-labs/FLUX.2-klein-9B"
LORA_REPO = "dx8152/Flux2-Klein-9B-Migration"
LORA_FILE = "Klein-Migration.safetensors"

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

PROMPT_PRESETS = {
    "Light migration": "将图1角色的光线替换为图2的光线",
    "Style migration": "将图1变为图2的画风",
}
DEFAULT_PRESET = "Light migration"

dtype = torch.bfloat16
device = "cuda" if torch.cuda.is_available() else "cpu"

pipe = Flux2KleinPipeline.from_pretrained(
    BASE_MODEL,
    torch_dtype=dtype,
    token=os.environ.get("HF_TOKEN"),
).to(device)

pipe.load_lora_weights(LORA_REPO, weight_name=LORA_FILE, adapter_name="migration")
pipe.set_adapters(["migration"], adapter_weights=[1.0])
pipe.fuse_lora(adapter_names=["migration"], lora_scale=1.0)
pipe.unload_lora_weights()


@spaces.GPU
def infer(
    image,
    reference,
    prompt,
    seed=0,
    randomize_seed=True,
    num_inference_steps=4,
    width=0,
    height=0,
    progress=gr.Progress(track_tqdm=True),
):
    """
    Migrate lighting or art style from a reference image (Image 2)
    onto a target image (Image 1) using FLUX.2 Klein 9B with the
    dx8152/Flux2-Klein-9B-Migration LoRA.

    Args:
        image: Target image (Image 1).
        reference: Reference image (Image 2) providing the lighting
            or style to migrate.
        prompt: Instruction describing the migration (Chinese prompts work
            best, e.g. 将图1角色的光线替换为图2的光线).
        seed: Random seed. Ignored when randomize_seed is True.
        randomize_seed: Pick a random seed per call.
        num_inference_steps: Denoising steps (Klein is distilled to 4).
        width: Output width. 0 = derive from Image 1.
        height: Output height. 0 = derive from Image 1.

    Returns:
        The edited image and the seed used.
    """
    if image is None:
        raise gr.Error("Please upload a target image (Image 1).")

    if reference is None:
        raise gr.Error("Please upload a reference image (Image 2).")

    source = image.convert("RGB")
    ref = reference.convert("RGB")

    if not prompt or not prompt.strip():
        raise gr.Error("Please provide a prompt.")

    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    generator = torch.Generator(device=device).manual_seed(seed)

    result = pipe(
        image=[source, ref],
        prompt=prompt,
        width=width if width else None,
        height=height if height else None,
        num_inference_steps=num_inference_steps,
        guidance_scale=1.0,
        generator=generator,
    ).images[0]

    return result, seed


def apply_preset(preset):
    return PROMPT_PRESETS.get(preset, "")


def update_dimensions(image):
    if image is None:
        return 0, 0

    w, h = image.size
    if w * h > 1024 * 1024:
        scale = (1024 * 1024 / (w * h)) ** 0.5
        w, h = int(w * scale), int(h * scale)
    return (w // 16) * 16, (h // 16) * 16


css = """
#col-container { max-width: 1100px; margin: 0 auto; }
.dark .progress-text { color: white !important }
#examples { max-width: 1100px; margin: 0 auto; }
"""

with gr.Blocks() as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("## 🪄 FLUX.2 Klein — Anything Migration")
        gr.Markdown("""
            Migrate **lighting** or **art style** from a reference image onto your target image ⚡
            Using [dx8152's Flux2-Klein-9B-Migration LoRA](https://huggingface.co/dx8152/Flux2-Klein-9B-Migration)
            on [FLUX.2 Klein 9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B) — 4-step distilled inference 💨
        """)

        with gr.Row():
            with gr.Column():
                with gr.Row():
                    image = gr.Image(
                        label="Image 1 (Target)",
                        type="pil",
                    )
                    reference = gr.Image(
                        label="Image 2 (Reference)",
                        type="pil",
                    )

                preset = gr.Radio(
                    label="Migration type",
                    choices=list(PROMPT_PRESETS.keys()),
                    value=DEFAULT_PRESET,
                )
                prompt = gr.Textbox(
                    label="Prompt",
                    value=PROMPT_PRESETS[DEFAULT_PRESET],
                    lines=2,
                    info="Editable — Chinese prompts work best. 图1 = Image 1, 图2 = Image 2.",
                )

                run_btn = gr.Button("🪄 Migrate", variant="primary", size="lg")

                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)
                    num_inference_steps = gr.Slider(
                        label="Inference Steps", minimum=1, maximum=16, step=1, value=4
                    )
                    width = gr.Slider(
                        label="Width (0 = auto from Image 1)",
                        minimum=0, maximum=2048, step=16, value=0,
                    )
                    height = gr.Slider(
                        label="Height (0 = auto from Image 1)",
                        minimum=0, maximum=2048, step=16, value=0,
                    )

            with gr.Column():
                result = gr.Image(label="Output Image", interactive=False)
                seed_used = gr.Number(label="Seed Used", interactive=False)

        gr.Examples(
            examples=[
                ["character_1.png", "light_1.png", PROMPT_PRESETS["Light migration"]],
                ["character_2.png", "light_4.png", PROMPT_PRESETS["Light migration"]],
                ["place_1.png", "light_6.png", PROMPT_PRESETS["Light migration"]],
                ["character_1.png", "metropolis.jpg", PROMPT_PRESETS["Style migration"]],
            ],
            inputs=[image, reference, prompt],
            outputs=[result, seed_used],
            fn=infer,
            cache_examples=True,
            cache_mode="lazy",
            elem_id="examples",
        )

    preset.change(fn=apply_preset, inputs=[preset], outputs=[prompt])
    image.change(fn=update_dimensions, inputs=[image], outputs=[width, height])

    run_btn.click(
        fn=infer,
        inputs=[image, reference, prompt, seed, randomize_seed, num_inference_steps, width, height],
        outputs=[result, seed_used],
    )

demo.launch(mcp_server=True, theme=gr.themes.Citrus(), css=css, footer_links=["api", "gradio", "settings"])