app.py

import secrets
from typing import cast

import gradio as gr
import numpy as np
import spaces
import torch
from diffusers import FluxFillPipeline
from gradio.components.image_editor import EditorValue
from PIL import Image, ImageFilter, ImageOps

DEVICE = "cuda"

MAX_SEED = np.iinfo(np.int32).max
# FIXED_DIMENSION = 900
FIXED_DIMENSION = 512 + (512 // 2)
FIXED_DIMENSION = (FIXED_DIMENSION // 16) * 16

SYSTEM_PROMPT = r"""This two-panel split-frame image showcases a furniture in as a product shot versus styled in a room.
[LEFT] standalone product shot image the furniture on a white background.
[RIGHT] integrated example within a room scene."""

if not torch.cuda.is_available():

    def _dummy_pipe(image: list[Image.Image], *args, **kwargs):  # noqa: ARG001
        return {"images": image}

    pipe = _dummy_pipe
else:
    state_dict, network_alphas = FluxFillPipeline.lora_state_dict(
        pretrained_model_name_or_path_or_dict="blanchon/FluxFillFurniture",
        weight_name="pytorch_lora_weights3.safetensors",
        return_alphas=True,
    )

    if not all(("lora" in key or "dora_scale" in key) for key in state_dict):
        msg = "Invalid LoRA checkpoint."
        raise ValueError(msg)

    pipe = FluxFillPipeline.from_pretrained(
        "black-forest-labs/FLUX.1-Fill-dev", torch_dtype=torch.bfloat16
    ).to(DEVICE)
    FluxFillPipeline.load_lora_into_transformer(
        state_dict=state_dict,
        network_alphas=network_alphas,
        transformer=pipe.transformer,
    )
    pipe.to(DEVICE)


@spaces.GPU
def infer(
    furniture_image: Image.Image,
    room_image: EditorValue,
    prompt: str = "",
    seed: int = 42,
    randomize_seed: bool = False,
    guidance_scale: float = 3.5,
    num_inference_steps: int = 28,
    progress: gr.Progress = gr.Progress(track_tqdm=True),  # noqa: ARG001, B008
):
    _room_image = room_image["background"]
    if _room_image is None:
        msg = "Room image is required"
        raise ValueError(msg)
    _room_image = cast(Image.Image, _room_image)
    _room_image = ImageOps.fit(
        _room_image,
        (FIXED_DIMENSION, FIXED_DIMENSION),
        method=Image.Resampling.LANCZOS,
        centering=(0.5, 0.5),
    )

    _room_mask = room_image["layers"][0]
    if _room_mask is None:
        msg = "Room mask is required"
        raise ValueError(msg)
    _room_mask = cast(Image.Image, _room_mask)
    _room_mask = ImageOps.fit(
        _room_mask,
        (FIXED_DIMENSION, FIXED_DIMENSION),
        method=Image.Resampling.LANCZOS,
        centering=(0.5, 0.5),
    )

    furniture_image = ImageOps.fit(
        furniture_image,
        (FIXED_DIMENSION, FIXED_DIMENSION),
        method=Image.Resampling.LANCZOS,
        centering=(0.5, 0.5),
    )
    _furniture_image = Image.new(
        "RGB",
        (FIXED_DIMENSION, FIXED_DIMENSION),
        (255, 255, 255),
    )
    _furniture_image.paste(furniture_image, (0, 0))

    _furniture_mask = Image.new(
        "RGB", (FIXED_DIMENSION, FIXED_DIMENSION), (255, 255, 255)
    )

    image = Image.new(
        "RGB",
        (FIXED_DIMENSION * 2, FIXED_DIMENSION),
        (255, 255, 255),
    )
    # Paste on the center of the image
    image.paste(_furniture_image, (0, 0))
    image.paste(_room_image, (FIXED_DIMENSION, 0))

    mask = Image.new(
        "RGB",
        (FIXED_DIMENSION * 2, FIXED_DIMENSION),
        (255, 255, 255),
    )
    mask.paste(_furniture_mask, (0, 0))
    mask.paste(_room_mask, (FIXED_DIMENSION, 0), _room_mask)
    # Invert the mask
    mask = ImageOps.invert(mask)
    # Blur the mask
    mask = mask.filter(ImageFilter.GaussianBlur(radius=10))
    # Convert to 3 channel
    mask = mask.convert("L")

    if randomize_seed:
        seed = secrets.randbelow(MAX_SEED)

    prompt = prompt + ".\n" + SYSTEM_PROMPT if prompt else SYSTEM_PROMPT
    batch_size = 4
    results_images = pipe(
        prompt=[prompt] * batch_size,
        image=[image] * batch_size,
        mask_image=[mask] * batch_size,
        height=FIXED_DIMENSION,
        width=FIXED_DIMENSION * 2,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        generator=torch.Generator("cpu").manual_seed(seed),
    )["images"]

    print(len(results_images))

    cropped_images = [
        image.crop((FIXED_DIMENSION, 0, FIXED_DIMENSION * 2, FIXED_DIMENSION))
        for image in results_images
    ]

    return cropped_images, seed


intro_markdown = """
# AnyFurnish

AnyFurnish is a tool that allows you to generate furniture images using Flux.1 Fill Dev.
"""

css = """
#col-container {
    margin: 0 auto;
    max-width: 1000px;
}
"""

with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown(intro_markdown)
        with gr.Row():
            with gr.Column():
                with gr.Column():
                    furniture_image = gr.Image(
                        label="Furniture Image",
                        type="pil",
                        sources=["upload"],
                        image_mode="RGB",
                        height=300,
                    )
                    room_image = gr.ImageEditor(
                        label="Room Image - Draw mask for inpainting",
                        type="pil",
                        sources=["upload"],
                        image_mode="RGBA",
                        layers=False,
                        crop_size="1:1",
                        brush=gr.Brush(colors=["#FFFFFF"], color_mode="fixed"),
                        height=300,
                    )
                prompt = gr.Text(
                    label="Prompt",
                    show_label=False,
                    max_lines=1,
                    placeholder="Enter a custom furniture description (optional)",
                    container=False,
                )
                run_button = gr.Button("Run")

            results = gr.Gallery(
                label="Results",
                format="png",
                show_label=False,
                columns=2,
                height=600,
            )

        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)

            with gr.Row():
                guidance_scale = gr.Slider(
                    label="Guidance Scale",
                    minimum=1,
                    maximum=30,
                    step=0.5,
                    value=50,
                )

                num_inference_steps = gr.Slider(
                    label="Number of inference steps",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=28,
                )

    gr.on(
        triggers=[run_button.click, prompt.submit],
        fn=infer,
        inputs=[
            furniture_image,
            room_image,
            prompt,
            seed,
            randomize_seed,
            guidance_scale,
            num_inference_steps,
        ],
        outputs=[results, seed],
    )

demo.launch()