app.py

import os
import torch
import spaces
import gradio as gr
from diffusers import FluxFillPipeline
import random
import numpy as np
from huggingface_hub import hf_hub_download
from PIL import Image, ImageOps


CSS = """
h1 {
    margin-top: 10px
}
"""

os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
MAX_SEED = np.iinfo(np.int32).max

repo_id = "black-forest-labs/FLUX.1-Fill-dev"

if torch.cuda.is_available():
    pipe = FluxFillPipeline.from_pretrained(repo_id, torch_dtype=torch.bfloat16).to("cuda")

@spaces.GPU()
def inpaintGen(
        imgMask,
        inpaint_prompt: str,
        guidance: float,
        num_steps: int,
        seed: int,
        randomize_seed: bool,
        progress=gr.Progress(track_tqdm=True)):

    source_path = imgMask["background"]
    mask_path = imgMask["layers"][0]

    if not source_path:
        raise gr.Error("Please upload an image.")

    if not mask_path:
        raise gr.Error("Please draw a mask on the image.")

    source_img = Image.open(source_path).convert("RGB")
    mask_img = Image.open(mask_path)
    alpha_channel=mask_img.split()[3]
    binary_mask = alpha_channel.point(lambda p: p > 0 and 255)
    
    width, height = source_img.size

    new_width = (width // 16) * 16
    new_height = (height // 16) * 16
        
    # If the image size is not already divisible by 16, resize it
    if width != new_width or height != new_height:
        source_img = source_img.resize((new_width, new_height), Image.LANCZOS)

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

    result = pipe(
        prompt=inpaint_prompt,
        image=source_img,
        mask_image=binary_mask,
        width=new_width,
        height=new_height,
        num_inference_steps=num_steps,
        generator=generator,
        guidance_scale=guidance,
        max_sequence_length=512,
    ).images[0]
    
    return result, seed



@spaces.GPU()
def outpaintGen(
    img,
    outpaint_prompt: str,
    overlap_top: int,
    overlap_right: int,
    overlap_bottom: int,
    overlap_left: int,
    op_guidance: float,
    op_num_steps: int,
    op_seed: int,
    op_randomize_seed: bool
):
    image = Image.open(img)

    # Convert input to PIL Image if it's a numpy array
    if isinstance(image, np.ndarray):
        image = Image.fromarray(image)

    # Get original dimensions
    original_width, original_height = image.size
    
    # Calculate new dimensions
    new_width = original_width + overlap_left + overlap_right
    new_height = original_height + overlap_top + overlap_bottom

    # Create new blank mask image (black background)
    mask_image = Image.new('RGB', (new_width, new_height), color='black')

    # Create white rectangle for original image area
    white_area = Image.new('RGB', (original_width, original_height), color='white')

    # Paste white rectangle at the appropriate position
    mask_image.paste(white_area, (overlap_left, overlap_top))

    # Create a new image with the same size as the original image
    new_image = Image.new('RGB', (new_width, new_height), color='black')

    # Paste the original image onto the new image
    new_image.paste(image, (overlap_left, overlap_top))

    # Convert to grayscale
    mask_image = mask_image.convert('L')
    mask_image = Image.eval(mask_image, lambda x: 255 - x)
    
    fix_width = (new_width // 16) * 16
    fix_height = (new_height // 16) * 16
        
    # If the image size is not already divisible by 16, resize it
    # if new_width != fix_width or new_height != fix_height:
    #     mask_image = mask_image.resize((fix_width, fix_height), Image.LANCZOS)

    if op_randomize_seed:
        op_seed = random.randint(0, MAX_SEED)
    generator = torch.Generator("cpu").manual_seed(op_seed)

    result = pipe(
        prompt=outpaint_prompt,
        image=new_image,
        mask_image=mask_image,
        width=fix_width,
        height=fix_height,
        num_inference_steps=op_num_steps,
        generator=generator,
        guidance_scale=op_guidance,
        max_sequence_length=512,
    ).images[0]
    
    return result, seed


with gr.Blocks(theme="ocean", title="Flux.1 Fill dev", css=CSS) as demo:
    gr.HTML("<h1><center>Flux.1 Fill dev</center></h1>")
    gr.HTML("""
        <p>
            <center>
                FLUX.1 Fill [dev] is a 12 billion parameter rectified flow transformer capable of filling areas in existing images based on a text description.
            </center>
        </p>
    """)
    with gr.Tab("Inpainting"):
        with gr.Row():
            with gr.Column():
                imgMask = gr.ImageMask(type="filepath", label="Image", layers=False, height=800)
                inpaint_prompt = gr.Textbox(label='Prompts ✏️', placeholder="A hat...")
                with gr.Row():
                    Inpaint_sendBtn = gr.Button(value="Submit", variant='primary')
                    Inpaint_clearBtn = gr.ClearButton([imgMask, inpaint_prompt], value="Clear")
            image_out = gr.Image(type="pil", label="Output", height=960)
        with gr.Accordion("Advanced ⚙️", open=False):
            guidance = gr.Slider(label="Guidance scale", minimum=1, maximum=50, value=30.0, step=0.1)
            num_steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=20, step=1)
            seed = gr.Number(label="Seed", value=42, precision=0)
            randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
    
        gr.on(
            triggers = [
                inpaint_prompt.submit,
                Inpaint_sendBtn.click,
            ],
            fn = inpaintGen,
            inputs = [
                imgMask,
                inpaint_prompt,
                guidance,
                num_steps,
                seed,
                randomize_seed
            ],
            outputs = [image_out, seed]
        )

    with gr.Tab("Outpainting"):
        with gr.Row():
            with gr.Column():
                img = gr.Image(type="filepath", label="Image", height=800)
                outpaint_prompt = gr.Textbox(label='Prompts ✏️', placeholder="In city...")
                with gr.Row():
                    outpaint_sendBtn = gr.Button(value="Submit", variant='primary')
                    outpaint_clearBtn = gr.ClearButton([img, outpaint_prompt], value="Clear")
            image_exp = gr.Image(type="pil", label="Output", height=960)
        with gr.Accordion("Advanced ⚙️", open=False):
            with gr.Row():
                overlap_top = gr.Number(label="Top", value=64, precision=0)
                overlap_right = gr.Number(label="Right", value=64, precision=0)
                overlap_bottom = gr.Number(label="Bottom", value=64, precision=0)
                overlap_left = gr.Number(label="Left", value=64, precision=0)
            op_guidance = gr.Slider(label="Guidance scale", minimum=1, maximum=50, value=30.0, step=0.1)
            op_num_steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=20, step=1)
            op_seed = gr.Number(label="Seed", value=42, precision=0)
            op_randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
    
        gr.on(
            triggers = [
                outpaint_prompt.submit,
                outpaint_sendBtn.click,
            ],
            fn = outpaintGen,
            inputs = [
                img,
                outpaint_prompt,
                overlap_top,
                overlap_right,
                overlap_bottom,
                overlap_left,
                op_guidance,
                op_num_steps,
                op_seed,
                op_randomize_seed
            ],
            outputs = [image_exp, op_seed]
        )


if __name__ == "__main__":
    demo.launch(show_api=False, share=False)