app.py

import os
import math
import torch
import spaces
import gradio as gr
import numpy as np
from PIL import Image, ImageOps, ImageFilter, ImageEnhance
from functools import lru_cache

# ============================================================
# 🎨 AI Image Editor - Powered by HuggingFace Models
# Features:
#   1. ✏️ Instruction-based Editing (InstructPix2Pix)
#   2. 🖌️ Inpainting (SDXL Inpainting)
#   3. ✂️ Background Removal (RMBG-2.0 / BiRefNet)
#   4. 🔍 Image Upscaling (Swin2SR)
# ============================================================

# --- Global model holders (lazy loaded) ---
_edit_pipe = None
_inpaint_pipe = None
_rmbg_model = None
_rmbg_transform = None
_upscale_processor = None
_upscale_model = None


def get_edit_pipe():
    """Lazy load InstructPix2Pix pipeline"""
    global _edit_pipe
    if _edit_pipe is None:
        from diffusers import StableDiffusionInstructPix2PixPipeline, EulerAncestralDiscreteScheduler
        _edit_pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
            "timbrooks/instruct-pix2pix",
            torch_dtype=torch.float16,
            safety_checker=None,
        ).to("cuda")
        _edit_pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(
            _edit_pipe.scheduler.config
        )
    return _edit_pipe


def get_inpaint_pipe():
    """Lazy load SDXL Inpainting pipeline"""
    global _inpaint_pipe
    if _inpaint_pipe is None:
        from diffusers import AutoPipelineForInpainting
        _inpaint_pipe = AutoPipelineForInpainting.from_pretrained(
            "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
            torch_dtype=torch.float16,
            variant="fp16",
        ).to("cuda")
    return _inpaint_pipe


def get_rmbg_model():
    """Lazy load background removal model"""
    global _rmbg_model, _rmbg_transform
    if _rmbg_model is None:
        from torchvision import transforms
        from transformers import AutoModelForImageSegmentation
        _rmbg_model = AutoModelForImageSegmentation.from_pretrained(
            "ZhengPeng7/BiRefNet", trust_remote_code=True
        ).to("cuda").eval()
        _rmbg_transform = transforms.Compose([
            transforms.Resize((1024, 1024)),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ])
    return _rmbg_model, _rmbg_transform


def get_upscale_model():
    """Lazy load Swin2SR upscaling model"""
    global _upscale_processor, _upscale_model
    if _upscale_model is None:
        from transformers import AutoImageProcessor, Swin2SRForImageSuperResolution
        _upscale_processor = AutoImageProcessor.from_pretrained(
            "caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr"
        )
        _upscale_model = Swin2SRForImageSuperResolution.from_pretrained(
            "caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr"
        ).to("cuda").eval()
    return _upscale_processor, _upscale_model


# ============================================================
# Feature 1: Instruction-based Editing
# ============================================================
@spaces.GPU(duration=120)
def instruct_edit(input_image, instruction, text_cfg, image_cfg, steps, seed):
    if input_image is None:
        raise gr.Error("⚠️ Vui lòng upload ảnh trước!")
    if not instruction or instruction.strip() == "":
        raise gr.Error("⚠️ Vui lòng nhập lệnh chỉnh sửa!")

    pipe = get_edit_pipe()

    # Resize to be compatible with the model (multiples of 64)
    width, height = input_image.size
    factor = 512 / max(width, height)
    factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
    w = int((width * factor) // 64) * 64
    h = int((height * factor) // 64) * 64
    input_resized = ImageOps.fit(input_image, (w, h), method=Image.Resampling.LANCZOS)

    generator = torch.Generator("cuda").manual_seed(int(seed))

    result = pipe(
        instruction,
        image=input_resized,
        guidance_scale=text_cfg,
        image_guidance_scale=image_cfg,
        num_inference_steps=int(steps),
        generator=generator,
    ).images[0]

    # Resize back to original
    result = result.resize(input_image.size, Image.Resampling.LANCZOS)
    return result


# ============================================================
# Feature 2: Inpainting
# ============================================================
@spaces.GPU(duration=120)
def inpaint(input_dict, prompt, negative_prompt, guidance_scale, steps, strength, seed):
    if input_dict is None:
        raise gr.Error("⚠️ Vui lòng upload ảnh và vẽ mask!")

    # Extract image and mask from ImageEditor
    init_image = input_dict["background"].convert("RGB")
    # Get mask from the drawn layer
    if len(input_dict["layers"]) > 0:
        mask_layer = input_dict["layers"][0]
        # The alpha channel of the layer IS the mask
        if mask_layer.mode == "RGBA":
            mask = mask_layer.getchannel("A")
        else:
            mask = mask_layer.convert("L")
    else:
        raise gr.Error("⚠️ Vui lòng vẽ vùng cần chỉnh sửa trên ảnh!")

    # Check if mask has any painted area
    mask_array = np.array(mask)
    if mask_array.max() == 0:
        raise gr.Error("⚠️ Vui lòng vẽ vùng cần chỉnh sửa (brush) trên ảnh!")

    if not prompt or prompt.strip() == "":
        raise gr.Error("⚠️ Vui lòng nhập mô tả nội dung muốn tạo!")

    pipe = get_inpaint_pipe()

    # Resize to 1024x1024 for SDXL
    init_resized = init_image.resize((1024, 1024), Image.Resampling.LANCZOS)
    mask_resized = mask.resize((1024, 1024), Image.Resampling.LANCZOS)

    generator = torch.Generator("cuda").manual_seed(int(seed))

    output = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt if negative_prompt else None,
        image=init_resized,
        mask_image=mask_resized,
        guidance_scale=guidance_scale,
        num_inference_steps=int(steps),
        strength=strength,
        generator=generator,
    ).images[0]

    # Resize back
    output = output.resize(init_image.size, Image.Resampling.LANCZOS)
    return (init_image, output)


# ============================================================
# Feature 3: Background Removal
# ============================================================
@spaces.GPU(duration=60)
def remove_background(input_image):
    if input_image is None:
        raise gr.Error("⚠️ Vui lòng upload ảnh!")

    model, transform = get_rmbg_model()

    # Prepare input
    input_tensor = transform(input_image).unsqueeze(0).to("cuda")

    # Inference
    with torch.no_grad():
        preds = model(input_tensor)[-1].sigmoid().cpu()

    # Create mask
    pred_pil = Image.fromarray((preds[0].squeeze().numpy() * 255).astype(np.uint8))
    mask = pred_pil.resize(input_image.size, Image.Resampling.LANCZOS)

    # Apply alpha channel
    result = input_image.copy().convert("RGBA")
    result.putalpha(mask)

    return result


# ============================================================
# Feature 4: Image Upscaling
# ============================================================
@spaces.GPU(duration=120)
def upscale_image(input_image, scale_factor):
    if input_image is None:
        raise gr.Error("⚠️ Vui lòng upload ảnh!")

    processor, model = get_upscale_model()

    # Limit input size to prevent OOM
    max_dim = 512
    w, h = input_image.size
    if max(w, h) > max_dim:
        ratio = max_dim / max(w, h)
        new_w = int(w * ratio)
        new_h = int(h * ratio)
        input_image_resized = input_image.resize((new_w, new_h), Image.Resampling.LANCZOS)
    else:
        input_image_resized = input_image

    inputs = processor(input_image_resized, return_tensors="pt").to("cuda")

    with torch.no_grad():
        outputs = model(**inputs)

    output = outputs.reconstruction.data.squeeze().float().cpu().clamp(0, 1).numpy()
    output = np.moveaxis(output, 0, -1)  # CHW -> HWC
    output = (output * 255.0).round().astype(np.uint8)
    result = Image.fromarray(output)

    # If user wants 2x instead of 4x, resize down
    if scale_factor == "2x":
        target_w = input_image.size[0] * 2
        target_h = input_image.size[1] * 2
        result = result.resize((target_w, target_h), Image.Resampling.LANCZOS)

    orig_size = f"{input_image.size[0]}×{input_image.size[1]}"
    new_size = f"{result.size[0]}×{result.size[1]}"
    info = f"📐 Gốc: {orig_size} → Kết quả: {new_size}"

    return result, info


# ============================================================
# Feature 5: Basic adjustments (no GPU needed)
# ============================================================
def basic_adjust(input_image, brightness, contrast, saturation, sharpness, blur_radius):
    if input_image is None:
        raise gr.Error("⚠️ Vui lòng upload ảnh!")

    img = input_image.copy()

    if brightness != 1.0:
        img = ImageEnhance.Brightness(img).enhance(brightness)
    if contrast != 1.0:
        img = ImageEnhance.Contrast(img).enhance(contrast)
    if saturation != 1.0:
        img = ImageEnhance.Color(img).enhance(saturation)
    if sharpness != 1.0:
        img = ImageEnhance.Sharpness(img).enhance(sharpness)
    if blur_radius > 0:
        img = img.filter(ImageFilter.GaussianBlur(radius=blur_radius))

    return img


# ============================================================
# Custom CSS
# ============================================================
css = """
.gradio-container {
    max-width: 1200px !important;
    margin: auto !important;
}
.main-header {
    text-align: center;
    padding: 20px 0;
    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
    -webkit-background-clip: text;
    -webkit-text-fill-color: transparent;
    font-size: 2.5em;
    font-weight: bold;
}
.sub-header {
    text-align: center;
    color: #666;
    margin-bottom: 20px;
}
.feature-icon {
    font-size: 1.3em;
}
footer {
    text-align: center;
    padding: 20px;
    color: #999;
}
"""


# ============================================================
# Gradio UI
# ============================================================
with gr.Blocks(css=css, theme=gr.themes.Soft(), title="🎨 AI Image Editor") as demo:

    gr.HTML("""
    <div class="main-header">🎨 AI Image Editor</div>
    <div class="sub-header">
        Chỉnh sửa ảnh thông minh với AI | Powered by HuggingFace 🤗
    </div>
    """)

    # === Tab 1: Instruction Edit ===
    with gr.Tab("✏️ Chỉnh sửa bằng lệnh"):
        gr.Markdown("""
        ### Mô tả thay đổi bạn muốn bằng tiếng Anh
        Ví dụ: *"make it snowy"*, *"turn the sky to sunset"*, *"add sunglasses"*, *"make it a watercolor painting"*
        """)
        with gr.Row():
            with gr.Column(scale=1):
                edit_input = gr.Image(type="pil", label="📷 Ảnh gốc", height=400)
                edit_instruction = gr.Textbox(
                    label="✏️ Lệnh chỉnh sửa (tiếng Anh)",
                    placeholder="e.g. make it look like a painting...",
                    lines=2,
                )
                with gr.Accordion("⚙️ Cài đặt nâng cao", open=False):
                    edit_text_cfg = gr.Slider(
                        label="Text Guidance Scale",
                        minimum=1.0, maximum=15.0, value=7.5, step=0.5,
                        info="Mức độ tuân theo lệnh (cao = thay đổi nhiều hơn)"
                    )
                    edit_image_cfg = gr.Slider(
                        label="Image Guidance Scale",
                        minimum=0.5, maximum=3.0, value=1.5, step=0.1,
                        info="Mức độ giữ lại ảnh gốc (cao = giữ nhiều hơn)"
                    )
                    edit_steps = gr.Slider(
                        label="Số bước", minimum=10, maximum=100, value=30, step=5,
                    )
                    edit_seed = gr.Number(label="Seed", value=42, precision=0)
                edit_btn = gr.Button("🚀 Chỉnh sửa", variant="primary", size="lg")

            with gr.Column(scale=1):
                edit_output = gr.Image(label="🖼️ Kết quả", height=400)

        edit_btn.click(
            fn=instruct_edit,
            inputs=[edit_input, edit_instruction, edit_text_cfg, edit_image_cfg, edit_steps, edit_seed],
            outputs=edit_output,
        )

        gr.Examples(
            examples=[
                ["https://raw.githubusercontent.com/timbrooks/instruct-pix2pix/main/imgs/example.jpg", "turn him into a cyborg"],
                ["https://raw.githubusercontent.com/timbrooks/instruct-pix2pix/main/imgs/example.jpg", "make it a watercolor painting"],
            ],
            inputs=[edit_input, edit_instruction],
            label="💡 Ví dụ",
        )

    # === Tab 2: Inpainting ===
    with gr.Tab("🖌️ Inpainting (Tô vẽ lại)"):
        gr.Markdown("""
        ### Vẽ lên vùng cần thay đổi, sau đó mô tả nội dung mới
        Dùng **brush** để tô lên vùng muốn chỉnh sửa → nhập mô tả → nhấn Inpaint
        """)
        with gr.Row():
            with gr.Column(scale=1):
                inpaint_editor = gr.ImageEditor(
                    type="pil",
                    label="🖌️ Vẽ mask lên ảnh",
                    height=450,
                    brush=gr.Brush(
                        colors=["#FFFFFF"],
                        default_color="#FFFFFF",
                        color_mode="fixed",
                        default_size=30,
                    ),
                    eraser=gr.Eraser(default_size=30),
                    layers=True,
                )
                inpaint_prompt = gr.Textbox(
                    label="📝 Mô tả nội dung mới (tiếng Anh)",
                    placeholder="e.g. a cute cat sitting here...",
                    lines=2,
                )
                inpaint_neg = gr.Textbox(
                    label="🚫 Negative prompt (tùy chọn)",
                    placeholder="e.g. blurry, low quality, distorted...",
                    lines=1,
                )
                with gr.Accordion("⚙️ Cài đặt nâng cao", open=False):
                    inpaint_cfg = gr.Slider(
                        label="Guidance Scale", minimum=1.0, maximum=20.0, value=8.0, step=0.5,
                    )
                    inpaint_steps = gr.Slider(
                        label="Số bước", minimum=10, maximum=50, value=25, step=5,
                    )
                    inpaint_strength = gr.Slider(
                        label="Strength", minimum=0.5, maximum=1.0, value=0.99, step=0.01,
                        info="Mức độ thay đổi (1.0 = thay đổi hoàn toàn)",
                    )
                    inpaint_seed = gr.Number(label="Seed", value=42, precision=0)
                inpaint_btn = gr.Button("🎨 Inpaint", variant="primary", size="lg")

            with gr.Column(scale=1):
                inpaint_output = gr.ImageSlider(
                    label="📊 So sánh Trước / Sau",
                    height=450,
                )

        inpaint_btn.click(
            fn=inpaint,
            inputs=[inpaint_editor, inpaint_prompt, inpaint_neg, inpaint_cfg, inpaint_steps, inpaint_strength, inpaint_seed],
            outputs=inpaint_output,
        )

    # === Tab 3: Background Removal ===
    with gr.Tab("✂️ Xóa nền"):
        gr.Markdown("""
        ### Xóa nền ảnh tự động bằng AI
        Upload ảnh → nhận ảnh nền trong suốt (PNG)
        """)
        with gr.Row():
            with gr.Column(scale=1):
                bg_input = gr.Image(type="pil", label="📷 Ảnh gốc", height=400)
                bg_btn = gr.Button("✂️ Xóa nền", variant="primary", size="lg")
            with gr.Column(scale=1):
                bg_output = gr.Image(label="🖼️ Kết quả (nền trong suốt)", height=400)

        bg_btn.click(fn=remove_background, inputs=bg_input, outputs=bg_output)

    # === Tab 4: Image Upscaling ===
    with gr.Tab("🔍 Phóng to ảnh"):
        gr.Markdown("""
        ### Phóng to ảnh chất lượng cao với AI
        Tăng độ phân giải ảnh lên 2x hoặc 4x mà không bị mờ
        """)
        with gr.Row():
            with gr.Column(scale=1):
                upscale_input = gr.Image(type="pil", label="📷 Ảnh gốc", height=400)
                upscale_factor = gr.Radio(
                    choices=["2x", "4x"], value="4x", label="📏 Mức phóng to",
                )
                upscale_btn = gr.Button("🔍 Phóng to", variant="primary", size="lg")
            with gr.Column(scale=1):
                upscale_output = gr.Image(label="🖼️ Kết quả", height=400)
                upscale_info = gr.Textbox(label="📐 Thông tin", interactive=False)

        upscale_btn.click(
            fn=upscale_image,
            inputs=[upscale_input, upscale_factor],
            outputs=[upscale_output, upscale_info],
        )

    # === Tab 5: Basic Adjustments ===
    with gr.Tab("🎚️ Chỉnh sửa cơ bản"):
        gr.Markdown("""
        ### Điều chỉnh các thông số cơ bản
        Độ sáng, tương phản, bão hòa, sắc nét, làm mờ
        """)
        with gr.Row():
            with gr.Column(scale=1):
                adj_input = gr.Image(type="pil", label="📷 Ảnh gốc", height=400)
                adj_brightness = gr.Slider(label="☀️ Độ sáng", minimum=0.1, maximum=3.0, value=1.0, step=0.05)
                adj_contrast = gr.Slider(label="🔲 Tương phản", minimum=0.1, maximum=3.0, value=1.0, step=0.05)
                adj_saturation = gr.Slider(label="🎨 Bão hòa", minimum=0.0, maximum=3.0, value=1.0, step=0.05)
                adj_sharpness = gr.Slider(label="🔪 Sắc nét", minimum=0.0, maximum=3.0, value=1.0, step=0.05)
                adj_blur = gr.Slider(label="💨 Làm mờ", minimum=0, maximum=10, value=0, step=0.5)
                adj_btn = gr.Button("✨ Áp dụng", variant="primary", size="lg")
            with gr.Column(scale=1):
                adj_output = gr.Image(label="🖼️ Kết quả", height=400)

        adj_btn.click(
            fn=basic_adjust,
            inputs=[adj_input, adj_brightness, adj_contrast, adj_saturation, adj_sharpness, adj_blur],
            outputs=adj_output,
        )

    # === Footer ===
    gr.HTML("""
    <footer>
        <hr>
        <p>
            🤗 <strong>AI Image Editor</strong> | 
            Models: <a href="https://huggingface.co/timbrooks/instruct-pix2pix" target="_blank">InstructPix2Pix</a> · 
            <a href="https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1" target="_blank">SDXL Inpainting</a> · 
            <a href="https://huggingface.co/ZhengPeng7/BiRefNet" target="_blank">BiRefNet</a> · 
            <a href="https://huggingface.co/caidas/swin2SR-realworld-sr-x4-64-bsrgan-psnr" target="_blank">Swin2SR</a>
        </p>
        <p>Powered by 🤗 HuggingFace Diffusers, Transformers & Gradio</p>
    </footer>
    """)


if __name__ == "__main__":
    demo.queue(max_size=20, api_open=False).launch()