image_styler_v2.py

import gradio as gr
from rembg import remove
from PIL import Image, ImageOps, ImageEnhance, ImageStat
import torch
import torchvision.transforms.functional as tf
from torchvision import transforms
import numpy as np
from src import model

# Load the harmonizer model
def load_harmonization_model(pretrained_path):
    harmonizer = model.Harmonizer()
    if torch.cuda.is_available():
        harmonizer = harmonizer.cuda()
    harmonizer.load_state_dict(torch.load(pretrained_path), strict=True)
    harmonizer.eval()
    return harmonizer

# Load the enhancer model
def load_enhancement_model(pretrained_path):
    enhancer = model.Enhancer()
    if torch.cuda.is_available():
        enhancer = enhancer.cuda()
    enhancer.load_state_dict(torch.load(pretrained_path), strict=True)
    enhancer.eval()
    return enhancer

# Function to unify the image using the custom AI harmonization model
def unify_image(combined_img, harmonizer):
    original_size = combined_img.size

    # Create a mask for the composite image
    mask = Image.new("L", original_size, 255)
    mask = mask.point(lambda p: p > 0 and 255)

    preprocess = transforms.Compose([
        transforms.ToTensor(),
    ])

    # Preprocess the images
    comp = preprocess(combined_img.convert("RGB")).unsqueeze(0)
    mask = preprocess(mask).unsqueeze(0)

    if torch.cuda.is_available():
        comp = comp.cuda()
        mask = mask.cuda()

    # Harmonization
    with torch.no_grad():
        arguments = harmonizer.predict_arguments(comp, mask)
        harmonized = harmonizer.restore_image(comp, mask, arguments)[-1]

    # Postprocess the output
    harmonized = np.transpose(harmonized[0].cpu().numpy(), (1, 2, 0)) * 255
    harmonized_img = Image.fromarray(harmonized.astype(np.uint8)).convert("RGBA")
    harmonized_img = harmonized_img.resize(original_size)
    
    return harmonized_img

# Function to enhance the image using the custom AI enhancement model
def enhance_unified_image(harmonized_img, enhancer):
    original_size = harmonized_img.size

    preprocess = transforms.Compose([
        transforms.ToTensor(),
    ])

    # Preprocess the image
    original = preprocess(harmonized_img.convert("RGB")).unsqueeze(0)

    # Create a mask (not used in enhancement, so all pixels are equal to 1)
    mask = original * 0 + 1

    if torch.cuda.is_available():
        original = original.cuda()
        mask = mask.cuda()

    # Enhancement
    with torch.no_grad():
        arguments = enhancer.predict_arguments(original, mask)
        enhanced = enhancer.restore_image(original, mask, arguments)[-1]

    # Postprocess the output
    enhanced = np.transpose(enhanced[0].cpu().numpy(), (1, 2, 0)) * 255
    enhanced_img = Image.fromarray(enhanced.astype(np.uint8)).convert("RGBA")
    enhanced_img = enhanced_img.resize(original_size)
    
    return enhanced_img

def embed_person_on_background(person_img, background_img, position_x, position_y, scale):
    # Scale the person image while keeping proportions
    person_width, person_height = person_img.size
    new_width = int(person_width * scale)
    new_height = int(person_height * scale)
    person_img = person_img.resize((new_width, new_height), Image.LANCZOS)
    
    # Calculate the position based on bottom-center transformation point
    background_width, background_height = background_img.size
    
    # Default position: bottom-center of the background
    default_x = (background_width - new_width) // 2
    default_y = background_height - new_height

    # Adjust the position based on sliders
    position_x = default_x + int(position_x)
    position_y = default_y + int(position_y)

    # Create a new image with the same size as the background and paste the person image onto it
    combined_img = Image.new("RGBA", background_img.size)
    combined_img.paste(background_img, (0, 0))
    combined_img.paste(person_img, (position_x, position_y), person_img)

    return combined_img

def auto_match_enhancers(person_img, background_img):
    # Calculate the enhancement factors based on the background image
    stat = ImageStat.Stat(background_img)
    mean = stat.mean[:3]  # Mean color of the background

    # Simple logic to calculate enhancement factors
    contrast = 1.5 if mean[0] < 128 else 1.2
    brightness = 1.2 if mean[1] < 128 else 1.1
    color = 1.3 if mean[2] < 128 else 1.0

    enhancers = [
        (ImageEnhance.Contrast(person_img), contrast),
        (ImageEnhance.Brightness(person_img), brightness),
        (ImageEnhance.Color(person_img), color),
    ]

    enhanced_image = person_img
    for enhancer, factor in enhancers:
        enhanced_image = enhancer.enhance(factor)
    return enhanced_image

def enhance_image(image, contrast, brightness, color):
    # Enhance the image based on the provided parameters
    enhancers = [
        (ImageEnhance.Contrast(image), contrast),
        (ImageEnhance.Brightness(image), brightness),
        (ImageEnhance.Color(image), color),
    ]
    enhanced_image = image
    for enhancer, factor in enhancers:
        enhanced_image = enhancer.enhance(factor)
    return enhanced_image

def process_images(person_img, background_img, enhance, auto_match, contrast, brightness, color, unify, position_x, position_y, scale):
    # Remove background from the person image
    person_no_bg = remove(person_img)
    
    if enhance and auto_match:
        print("Auto-matching enhancers based on the background color...")
        person_no_bg = auto_match_enhancers(person_no_bg, background_img)
    elif enhance:
        print(f"Applying enhancement with contrast={contrast}, brightness={brightness}, color={color}...")
        person_no_bg = enhance_image(person_no_bg, contrast, brightness, color)
    
    combined_img = embed_person_on_background(person_no_bg, background_img, position_x, position_y, scale)

    if unify:
        print("Unifying image with AI...")
        harmonizer = load_harmonization_model('pretrained/harmonizer.pth')
        combined_img = unify_image(combined_img, harmonizer)
        enhancer = load_enhancement_model('pretrained/enhancer.pth')
        combined_img = enhance_unified_image(combined_img, enhancer)
    
    return combined_img

def gradio_interface(person_img, background_img, enhance, auto_match, contrast, brightness, color, unify, position_x, position_y, scale):
    try:
        result = process_images(person_img, background_img, enhance, auto_match, contrast, brightness, color, unify, position_x, position_y, scale)
        return result
    except Exception as e:
        return str(e)

def update_enhancement_controls(auto_match):
    # Disable enhancement sliders if auto-match is checked
    return {
        contrast_slider: gr.update(interactive=not auto_match),
        brightness_slider: gr.update(interactive=not auto_match),
        color_slider: gr.update(interactive=not auto_match),
    }

# Create Gradio interface
with gr.Blocks(css='#output_image {max-width: 800px !important; width: auto !important; height: auto !important;}') as interface:
    with gr.Row():
        person_img = gr.Image(type="pil", label="Upload Person Image")
        background_img = gr.Image(type="pil", label="Upload Background Image")
    enhance = gr.Checkbox(label="Enhance Image", value=False)
    auto_match = gr.Checkbox(label="Auto-Match Enhancers", value=False)
    contrast_slider = gr.Slider(minimum=0.5, maximum=3.0, step=0.1, value=1.0, label="Contrast")
    brightness_slider = gr.Slider(minimum=0.5, maximum=3.0, step=0.1, value=1.0, label="Brightness")
    color_slider = gr.Slider(minimum=0.5, maximum=3.0, step=0.1, value=1.0, label="Color")
    
    auto_match.change(fn=update_enhancement_controls, inputs=auto_match, outputs=[contrast_slider, brightness_slider, color_slider])
    
    unify = gr.Checkbox(label="Unify Image with AI", value=True)
    position_x = gr.Slider(minimum=-500, maximum=500, step=1, value=0, label="Horizontal Position (pixels)")
    position_y = gr.Slider(minimum=-500, maximum=500, step=1, value=0, label="Vertical Position (pixels)")
    scale = gr.Slider(minimum=0.1, maximum=3.0, step=0.1, value=1.0, label="Scale")

    output = gr.Image(type="pil", label="Generated Image", elem_id="output_image")
    run_button = gr.Button("Run")
    
    run_button.click(
        fn=gradio_interface,
        inputs=[person_img, background_img, enhance, auto_match, contrast_slider, brightness_slider, color_slider, unify, position_x, position_y, scale],
        outputs=output
    )

if __name__ == "__main__":
    interface.launch()