app.py

import os
import cv2
import numpy as np
import onnxruntime as ort
import base64
from io import BytesIO
import gradio as gr
import spaces
import re

# Setup
API_KEY = os.getenv("API_KEY", "demo")
INPUT_SIZE = (512, 512)
MODEL_PATH = "BiRefNet-general-resolution_512x512-fp16-epoch_216.onnx"

# Load ONNX model
assert os.path.exists(MODEL_PATH), f"Model not found: {MODEL_PATH}"
session = ort.InferenceSession(MODEL_PATH, providers=["CPUExecutionProvider"])
input_name = session.get_inputs()[0].name

# Preprocess
def preprocess_image(image: np.ndarray):
    original_shape = image.shape[:2]
    resized = cv2.resize(image, INPUT_SIZE)
    normalized = (resized.astype(np.float32) / 255.0 - 0.5) / 0.5
    transposed = np.transpose(normalized, (2, 0, 1))
    input_tensor = np.expand_dims(transposed, axis=0).astype(np.float32)
    return input_tensor, original_shape, image

# Mask logic
def apply_mask(original_img, mask_array, original_shape):
    mask = np.squeeze(mask_array)

    # Resize the mask back to the original image size
    resized_mask = cv2.resize(mask, (original_shape[1], original_shape[0]))

    # Create binary mask
    binary_mask = (resized_mask > 0.5).astype(np.uint8)

    # Create alpha channel
    alpha = (binary_mask * 255).astype(np.uint8)

    # Convert to BGRA (preserve original image without resizing)
    if original_img.shape[2] == 3:
        bgra = cv2.cvtColor(original_img, cv2.COLOR_BGR2BGRA)
    else:
        bgra = original_img.copy()

    bgra[:, :, 3] = alpha
    return bgra

# ============ UI ============ #
@spaces.GPU
def remove_background_ui(image, bg=None):
    input_tensor, original_shape, original_img = preprocess_image(image)
    mask = session.run(None, {input_name: input_tensor})[0]
    result = apply_mask(original_img, mask, original_shape)

    if bg is not None:
        bg_resized = cv2.resize(bg, (original_shape[1], original_shape[0]))
        alpha = result[:, :, 3] / 255.0
        fg = result[:, :, :3]
        blended = (fg * alpha[..., None] + bg_resized * (1 - alpha[..., None])).astype(np.uint8)
        return blended

    return result

# ============ File API ============ #
@spaces.GPU
def remove_background_api(image_file, api_key=""):
    if api_key != API_KEY:
        raise gr.Error("❌ Invalid API Key")

    np_arr = np.fromfile(image_file, np.uint8)
    original_img = cv2.imdecode(np_arr, cv2.IMREAD_UNCHANGED)

    if original_img is None:
        raise gr.Error("❌ Unable to decode image.")

    if len(original_img.shape) == 2:
        original_img = cv2.cvtColor(original_img, cv2.COLOR_GRAY2BGR)

    input_tensor, original_shape, _ = preprocess_image(original_img)
    mask = session.run(None, {input_name: input_tensor})[0]
    result = apply_mask(original_img, mask, original_shape)

    success, buffer = cv2.imencode(".png", result)
    return f"data:image/png;base64,{base64.b64encode(buffer).decode('utf-8')}"

# ============ Base64 API ============ #
def clean_base64_string(b64_string):
    if b64_string.startswith('data:image'):
        b64_string = b64_string.split(',', 1)[1]

    b64_string = re.sub(r'\s+', '', b64_string)

    missing_padding = len(b64_string) % 4
    if missing_padding:
        b64_string += '=' * (4 - missing_padding)

    return b64_string

@spaces.GPU
def remove_background_base64_api(base64_image, api_key=""):
    try:
        if api_key != API_KEY:
            raise gr.Error("❌ Invalid API Key")

        if not re.match(r'^(data:image\/[a-zA-Z]+;base64,)?[A-Za-z0-9+/=\s]+$', base64_image):
            raise gr.Error("❌ Invalid base64 image format")

        cleaned_b64 = clean_base64_string(base64_image)

        image_data = base64.b64decode(cleaned_b64)
        np_arr = np.frombuffer(image_data, np.uint8)
        original_img = cv2.imdecode(np_arr, cv2.IMREAD_UNCHANGED)

        if original_img is None:
            raise gr.Error("❌ Unable to decode image.")

        if len(original_img.shape) == 2:
            original_img = cv2.cvtColor(original_img, cv2.COLOR_GRAY2BGR)

        input_tensor, original_shape, _ = preprocess_image(original_img)
        mask = session.run(None, {input_name: input_tensor})[0]
        result = apply_mask(original_img, mask, original_shape)

        success, buffer = cv2.imencode('.png', result)
        result_base64 = base64.b64encode(buffer).decode('utf-8')

        return f"data:image/png;base64,{result_base64}"

    except Exception as e:
        raise gr.Error(f"❌ Error processing image: {str(e)}")

# ============ Gradio Interfaces ============ #
ui = gr.Interface(
    fn=remove_background_ui,
    inputs=[
        gr.Image(type="numpy", label="Main Image"),
        gr.Image(type="numpy", label="Optional Background")
    ],
    outputs=gr.Image(type="numpy", label="Result"),
    title="🖼️ Background Remover",
    description="Upload a photo (and optionally a background)."
)

api = gr.Interface(
    fn=remove_background_api,
    inputs=[
        gr.Image(type="filepath", label="Upload Image"),
        gr.Text(label="API Key", type="password")
    ],
    outputs=gr.Text(label="Base64 PNG"),
    title="🔐 File API Access",
    description="POST to `/run/predict` with file + API key."
)

api_base64 = gr.Interface(
    fn=remove_background_base64_api,
    inputs=[
        gr.Text(label="Base64 Image String"),
        gr.Text(label="API Key", type="password")
    ],
    outputs=gr.Text(label="Base64 PNG"),
    title="🔐 Base64 API Access",
    description="POST to `/run/predict` with base64 image + API key."
)

# Final Gradio app
demo = gr.TabbedInterface([ui, api, api_base64], ["Web UI", "File API", "Base64 API"])

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