app.py

import os
import random
import tensorflow as tf
from keras import models
import numpy as np
import gradio as gr
import cv2

# Load the model
try:
    generator = models.load_model("generator.keras")
    print("Model loaded successfully!")
except Exception as e:
    print("Error loading model:", e)


# Function to preprocess the image (resize, normalize)
def preprocess_image(img):
    img = cv2.resize(img, (256, 256))

    # Convert L to range [-1, 1]
    img = img.astype("float32")
    img = (img / 127.5) - 1

    # Convert to tensor
    img = tf.convert_to_tensor(img, dtype=tf.float32)

    img = tf.expand_dims(img, axis=-1)  # Add image dimension
    img = tf.expand_dims(img, axis=0)  # Add batch dimension

    return img


# Function to postprocess the image (denormalize)
def postprocess_image(img):
    return cv2.cvtColor(((img + 1) * 127.5).numpy().astype(np.uint8), cv2.COLOR_LAB2RGB)


# Function to adjust brightness
def adjust_brightness(img, brightness=0.0):
    # Apply brightness adjustment
    img = cv2.convertScaleAbs(img, beta=int(brightness * 127.0 / 4.0))
    return np.uint8(np.clip(img, 0, 255))


# Function to adjust contrast
def adjust_contrast(img, contrast=0.0):
    # Apply contrast adjustment
    img = cv2.convertScaleAbs(img, alpha=(contrast * 0.75 + 1.0))
    return np.uint8(np.clip(img, 0, 255))


# Function to adjust hue
def adjust_hue(img, hue_shift=0.0):
    # Convert the image to HSV
    hsv_img = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)

    # Adjust the hue channel (value is between 0 and 179 in OpenCV's HSV)
    hsv_img[:, :, 0] = (
        hsv_img[:, :, 0] + hue_shift * 90
    ) % 180  # Hue is wrapped in OpenCV HSV format

    # Convert back to BGR
    img = cv2.cvtColor(hsv_img, cv2.COLOR_HSV2BGR)

    return np.uint8(np.clip(img, 0, 255))


def adjust_saturation(img, saturation_factor=0.0):
    # Convert the image to HSV
    hsv_img = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)

    # Adjust the saturation channel (index 1 in HSV)
    hsv_img[:, :, 1] = np.clip(hsv_img[:, :, 1] * (saturation_factor + 1.0), 0, 255)

    # Convert back to BGR
    img = cv2.cvtColor(hsv_img, cv2.COLOR_HSV2BGR)

    return np.uint8(np.clip(img, 0, 255))


# Define the inference function
def colorize_image(input_image):
    # Preprocess the image for the model
    preprocessed_image = preprocess_image(input_image)

    # Predict using the model
    output_ab = generator.predict(preprocessed_image)
    output = tf.concat([preprocessed_image[0], output_ab[0]], axis=-1)

    # Postprocess the output
    output_image = postprocess_image(output)

    return output_image


# Function to colorize and store the result for further manipulation
def colorize_and_store(img, bright_slider, cont_slider, sat_slider, hue_slider):
    # Colorize the image
    colorized_image = colorize_image(img)

    output_image = adjust_brightness(colorized_image, bright_slider)
    output_image = adjust_contrast(output_image, cont_slider)
    output_image = adjust_saturation(output_image, sat_slider)
    output_image = adjust_hue(output_image, hue_slider)

    # Return the colorized image for further manipulation (no model call)
    return colorized_image, output_image


def make_grayscale_256(img):
    img = cv2.resize(img, (256, 256))
    # img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

    return img


css = """
h1 {
    text-align: center;
    display:block;
    font-size: 3rem;
    margin: 0;
    padding: 0.5rem;
    line-height: 1;
    overflow: hidden;
}
p {
    text-align: center;
    display:block;
    font-size:1.5rem;
    margin: 0;
    padding: 0.5rem;
    line-height: 1;
    overflow: hidden;
}
#input-image img {
    filter: grayscale(1);
}
"""

# Get all image file paths in the folder
image_files = [
    os.path.join("examples", file)
    for file in os.listdir("examples")
    if file.lower().endswith((".png", ".jpg", ".jpeg", ".webp"))
]

# Gradio Interface
with gr.Blocks(css=css) as demo:
    demo.title = "Portrait Colorizer"

    # title
    gr.HTML("<h1>Portrait Colorizer</h1>")
    # description
    gr.HTML("<p>Upload a grayscale image to colorize it and fine-tune the output using the sliders below.</p>")


    with gr.Row():
        input_image = gr.Image(
            type="numpy",
            label="Grayscale Image",
            image_mode="L",
            height=256,
            width=256,
            elem_id="input-image",
        )
        examples_gallery = gr.Examples(
            examples=image_files, inputs=[input_image], label="Example Images"
        )
        output_image = gr.Image(
            type="numpy",
            label="Colorized Image",
            image_mode="RGB",
            height=256,
            width=256,
        )
    process_button = gr.Button("Colorize")
    bright_slider = gr.Slider(-1.0, 1.0, value=0.0, label="Brightness")
    cont_slider = gr.Slider(-1.0, 1.0, value=0.0, label="Contrast")
    sat_slider = gr.Slider(-1.0, 1.0, value=0.0, label="Saturation")
    hue_slider = gr.Slider(-1.0, 1.0, value=0.0, label="Hue")

    # Initially colorize and display the image when it is uploaded
    colorized_image = gr.State()

    # Button click triggers processing
    process_button.click(
        fn=colorize_and_store,
        inputs=[input_image, bright_slider, cont_slider, sat_slider, hue_slider],
        outputs=[colorized_image, output_image],
    )

    # Apply hue adjustment to the stored colorized image (no re-generation)
    bright_slider.change(
        fn=adjust_brightness,
        inputs=[colorized_image, bright_slider],
        outputs=output_image,  # Update output image
    )

    # Apply hue adjustment to the stored colorized image (no re-generation)
    cont_slider.change(
        fn=adjust_contrast,
        inputs=[colorized_image, cont_slider],
        outputs=output_image,  # Update output image
    )

    # Apply hue adjustment to the stored colorized image (no re-generation)
    hue_slider.change(
        fn=adjust_hue,
        inputs=[colorized_image, hue_slider],
        outputs=output_image,  # Update output image
    )

    # Apply saturation adjustment to the stored colorized image (no re-generation)
    sat_slider.change(
        fn=adjust_saturation,
        inputs=[colorized_image, sat_slider],
        outputs=output_image,  # Update output image
    )

# Launch the app
demo.launch(share=True, ssr_mode=False)