app.py

import os
import gc
import random
from typing import Iterable, List, Tuple

from huggingface_hub import login as hf_login
_hf_token = os.environ.get("HF_TOKEN")
if _hf_token:
    hf_login(token=_hf_token)

import gradio as gr
import numpy as np
import spaces
import torch
from PIL import Image
from gradio.themes import Soft
from gradio.themes.utils import colors, fonts, sizes

# =========================================================
# THEME
# =========================================================
colors.fire_red = colors.Color(
    name="fire_red",
    c50="#FFF5F0",
    c100="#FFE8DB",
    c200="#FFD0B5",
    c300="#FFB088",
    c400="#FF8C5A",
    c500="#FF6B35",
    c600="#E8531F",
    c700="#CC4317",
    c800="#A63812",
    c900="#80300F",
    c950="#5C220A",
)


class FireRedTheme(Soft):
    def __init__(
        self,
        *,
        primary_hue: colors.Color | str = colors.gray,
        secondary_hue: colors.Color | str = colors.fire_red,
        neutral_hue: colors.Color | str = colors.slate,
        text_size: sizes.Size | str = sizes.text_md,
        font: fonts.Font | str | Iterable[fonts.Font | str] = (
            fonts.GoogleFont("Inter"),
            "system-ui",
            "sans-serif",
        ),
        font_mono: fonts.Font | str | Iterable[fonts.Font | str] = (
            fonts.GoogleFont("JetBrains Mono"),
            "ui-monospace",
            "monospace",
        ),
    ):
        super().__init__(
            primary_hue=primary_hue,
            secondary_hue=secondary_hue,
            neutral_hue=neutral_hue,
            text_size=text_size,
            font=font,
            font_mono=font_mono,
        )
        super().set(
            body_background_fill="#f0f2f6",
            body_background_fill_dark="*neutral_950",
            background_fill_primary="white",
            background_fill_primary_dark="*neutral_900",
            block_background_fill="white",
            block_background_fill_dark="*neutral_800",
            block_border_width="1px",
            block_border_color="*neutral_200",
            block_border_color_dark="*neutral_700",
            block_shadow="0 1px 4px rgba(0,0,0,0.05)",
            block_shadow_dark="0 1px 4px rgba(0,0,0,0.25)",
            block_title_text_weight="600",
            block_label_background_fill="*neutral_50",
            block_label_background_fill_dark="*neutral_800",
            button_primary_text_color="white",
            button_primary_text_color_hover="white",
            button_primary_background_fill="linear-gradient(135deg, *secondary_500, *secondary_600)",
            button_primary_background_fill_hover="linear-gradient(135deg, *secondary_600, *secondary_700)",
            button_primary_background_fill_dark="linear-gradient(135deg, *secondary_500, *secondary_600)",
            button_primary_background_fill_hover_dark="linear-gradient(135deg, *secondary_600, *secondary_700)",
            button_primary_shadow="0 4px 14px rgba(232, 83, 31, 0.25)",
            button_secondary_text_color="*secondary_700",
            button_secondary_text_color_dark="*secondary_300",
            button_secondary_background_fill="*secondary_50",
            button_secondary_background_fill_hover="*secondary_100",
            button_secondary_background_fill_dark="rgba(255, 107, 53, 0.1)",
            button_secondary_background_fill_hover_dark="rgba(255, 107, 53, 0.2)",
            button_large_padding="12px 24px",
            slider_color="*secondary_500",
            slider_color_dark="*secondary_500",
            input_border_color_focus="*secondary_400",
            input_border_color_focus_dark="*secondary_500",
            color_accent_soft="*secondary_50",
            color_accent_soft_dark="rgba(255, 107, 53, 0.15)",
        )


theme = FireRedTheme()

# =========================================================
# MODEL
# =========================================================
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("CUDA_VISIBLE_DEVICES =", os.environ.get("CUDA_VISIBLE_DEVICES"))
print("torch.__version__ =", torch.__version__)
print("device =", device)

from diffusers import FlowMatchEulerDiscreteScheduler, QwenImageEditPlusPipeline  # noqa: E402,F401
from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3  # noqa: E402
from transformers import AutoModelForImageSegmentation  # noqa: E402
from torchvision import transforms  # noqa: E402
import torch.nn.functional as F  # noqa: E402

dtype = torch.bfloat16

# ── FireRed 编辑模型（官方原生加载）──
pipe = QwenImageEditPlusPipeline.from_pretrained(
    "FireRedTeam/FireRed-Image-Edit-1.1",
    torch_dtype=dtype,
).to(device)

pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

try:
    pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
    print("Flash Attention 3 Processor set successfully.")
except Exception as e:
    print(f"Warning: Could not set FA3 processor: {e}")

# ── Lightning LoRA（4步加速，与 ComfyUI Rebels.json 完全一致）──
try:
    pipe.load_lora_weights(
        "Osrivers/Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors",
        weight_name="Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors",
        adapter_name="lightning",
    )
    pipe.set_adapters(["lightning"], adapter_weights=[1.0])
    print("Lightning LoRA (4steps V2.0) loaded successfully.")
except Exception as e:
    print(f"Warning: Could not load Lightning LoRA: {e}")

# ── RMBG 2.0 抠图模型 ──
rmbg = AutoModelForImageSegmentation.from_pretrained(
    "briaai/RMBG-2.0",
    trust_remote_code=True,
)
rmbg.to(device)
rmbg.eval()

MAX_SEED = np.iinfo(np.int32).max
DEFAULT_NEGATIVE_PROMPT = (
    "worst quality, low quality, bad anatomy, bad hands, text, error, "
    "missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, "
    "signature, watermark, username, blurry"
)

# =========================================================
# SAFE BUCKETS (~1MP each)
# =========================================================
SAFE_BUCKETS: List[Tuple[int, int]] = [
    # 标准桶 (~1MP)
    (1024, 1024),
    (1184, 880),
    (880, 1184),
    (1392, 752),
    (752, 1392),
    (1568, 672),
    (672, 1568),
    # 宽图桶（综艺花字等长条形图）
    (1920, 640),   # 3:1
    (1600, 400),   # 4:1  ← Rebels.json 同款
    (2048, 512),   # 4:1
    (1920, 384),   # 5:1
    (2560, 512),   # 5:1
    (2048, 336),   # ~6:1
]

UPSCALE_SMALL_IMAGES = True

_rmbg_normalize = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

RMBG_SIZE = 1024


@spaces.GPU
def run_rmbg(pil_image: Image.Image) -> Image.Image:
    """用 RMBG-2.0 去除背景，与 ComfyUI comfyui-rmbg 完全一致：
    squish 到 1024×1024，sigmoid 激活，bilinear resize 回原尺寸。
    """
    orig_w, orig_h = pil_image.size
    inp = _rmbg_normalize(pil_image.convert("RGB").resize((RMBG_SIZE, RMBG_SIZE), Image.LANCZOS))
    inp = inp.unsqueeze(0).to(device)

    with torch.no_grad():
        outputs = rmbg(inp)

    # 与 ComfyUI 完全一致：取最后输出层，sigmoid 激活
    if isinstance(outputs, list):
        result = outputs[-1].sigmoid().cpu()
    elif isinstance(outputs, dict) and 'logits' in outputs:
        result = outputs['logits'].sigmoid().cpu()
    else:
        result = outputs.sigmoid().cpu()

    result = torch.clamp(result.squeeze(), 0, 1)
    result = F.interpolate(result.unsqueeze(0).unsqueeze(0), size=(orig_h, orig_w), mode='bilinear').squeeze()

    mask_pil = Image.fromarray((result.numpy() * 255).astype(np.uint8))
    out = pil_image.convert("RGBA")
    out.putalpha(mask_pil)
    return out


def color_match_reinhard(source: Image.Image, result: Image.Image) -> Image.Image:
    """Reinhard RGB 均值/标准差色彩迁移：将 result 的色调对齐 source。"""
    src = np.array(source.convert("RGB")).astype(np.float32)
    res = np.array(result.convert("RGB")).astype(np.float32)
    out = np.zeros_like(res)
    for c in range(3):
        s_mean, s_std = src[:, :, c].mean(), src[:, :, c].std()
        r_mean, r_std = res[:, :, c].mean(), res[:, :, c].std()
        ratio = s_std / (r_std + 1e-6) if r_std > 0.5 else 1.0
        out[:, :, c] = (res[:, :, c] - r_mean) * ratio + s_mean
    return Image.fromarray(np.clip(out, 0, 255).astype(np.uint8))


def remove_black_bg(pil_image: Image.Image, dark_thresh: int = 40) -> Image.Image:
    """仅去除与四边连通的黑色背景，保留文字内部的黑色元素。
    用连通区域标记（flood fill）实现，不依赖 AI 模型。
    """
    from scipy import ndimage as ndi
    arr = np.array(pil_image.convert("RGB"))
    dark_mask = np.all(arr <= dark_thresh, axis=2)

    labeled, _ = ndi.label(dark_mask)

    # 找所有与图片边缘相连的连通区域
    border_labels = set()
    border_labels.update(labeled[0, :].tolist())
    border_labels.update(labeled[-1, :].tolist())
    border_labels.update(labeled[:, 0].tolist())
    border_labels.update(labeled[:, -1].tolist())
    border_labels.discard(0)  # 0 = 非黑色区域

    bg_mask = np.zeros(arr.shape[:2], dtype=bool)
    for lbl in border_labels:
        bg_mask |= (labeled == lbl)

    alpha = np.where(bg_mask, 0, 255).astype(np.uint8)
    out = pil_image.convert("RGBA")
    out.putalpha(Image.fromarray(alpha))
    return out


def add_image_watermark(result: Image.Image, ref: Image.Image, size: int = 200, padding: int = 16) -> Image.Image:
    result = result.copy().convert("RGBA")
    thumb = ref.convert("RGBA")
    thumb.thumbnail((size, size), Image.LANCZOS)
    result.paste(thumb, (padding, padding), thumb)
    return result.convert("RGB")


def paste_png_into_mask(editor_value: dict, png_image) -> Image.Image:
    """
    从 ImageEditor 的 mask 层提取 bounding box，
    把 PNG 等比缩放（最长边 = mask 最长边）后居中贴入。
    """
    if editor_value is None:
        raise gr.Error("⚠️ Please upload and draw a mask on the source image.")
    if png_image is None:
        raise gr.Error("⚠️ Please upload a PNG to place.")

    # 取底图和 mask 层
    background: Image.Image = editor_value.get("background")
    layers: list = editor_value.get("layers", [])

    if background is None:
        raise gr.Error("⚠️ No source image found.")
    if not layers:
        raise gr.Error("⚠️ Please draw a mask area on the image first.")

    if isinstance(background, np.ndarray):
        background = Image.fromarray(background)
    background = background.convert("RGBA")

    mask_layer = layers[0]
    if isinstance(mask_layer, np.ndarray):
        mask_layer = Image.fromarray(mask_layer)
    mask_layer = mask_layer.convert("RGBA")

    # 从 mask 层的 alpha 通道找 bounding box
    alpha = mask_layer.split()[3]
    bbox = alpha.getbbox()
    if bbox is None:
        raise gr.Error("⚠️ Mask area is empty. Please draw on the image.")

    x1, y1, x2, y2 = bbox
    mask_w = x2 - x1
    mask_h = y2 - y1
    mask_longest = max(mask_w, mask_h)

    # 加载 PNG
    if isinstance(png_image, str):
        png = Image.open(png_image).convert("RGBA")
    else:
        png = Image.fromarray(png_image).convert("RGBA")

    png_w, png_h = png.size
    png_longest = max(png_w, png_h)

    # 等比缩放：最长边对齐 mask 最长边
    scale = mask_longest / png_longest
    new_w = max(1, int(png_w * scale))
    new_h = max(1, int(png_h * scale))
    png_resized = png.resize((new_w, new_h), Image.LANCZOS)

    # 居中贴入 mask 区域
    paste_x = x1 + (mask_w - new_w) // 2
    paste_y = y1 + (mask_h - new_h) // 2

    result = background.copy()
    result.paste(png_resized, (paste_x, paste_y), png_resized)
    return result.convert("RGB")


# =========================================================
# HELPERS
# =========================================================
def load_pil_image(item) -> Image.Image:
    if item is None:
        return None
    if isinstance(item, Image.Image):
        return item.convert("RGB")
    if isinstance(item, str):
        return Image.open(item).convert("RGB")
    if isinstance(item, (tuple, list)):
        path = item[0]
        if isinstance(path, Image.Image):
            return path.convert("RGB")
        return Image.open(path).convert("RGB")
    return Image.open(item.name).convert("RGB")


def pick_best_bucket(
    orig_w: int,
    orig_h: int,
    buckets: List[Tuple[int, int]] = SAFE_BUCKETS,
    allow_upscale: bool = UPSCALE_SMALL_IMAGES,
) -> Tuple[int, int]:
    if orig_w <= 0 or orig_h <= 0:
        return 1024, 1024

    orig_ratio = orig_w / orig_h

    def score(bucket):
        bw, bh = bucket
        ratio_diff = abs((bw / bh) - orig_ratio)
        area_diff = abs((bw * bh) - (orig_w * orig_h))
        return (ratio_diff, area_diff)

    sorted_buckets = sorted(buckets, key=score)

    if allow_upscale:
        return sorted_buckets[0]

    not_larger = [b for b in sorted_buckets if b[0] <= orig_w and b[1] <= orig_h]
    return not_larger[0] if not_larger else sorted_buckets[0]


def prepare_images_before_pipe(
    pil_images: List[Image.Image],
    allow_upscale: bool = UPSCALE_SMALL_IMAGES,
    divisible_by: int = 16,
) -> Tuple[List[Image.Image], int, int, tuple]:
    """准备图片：等比缩放 + 补边到最佳 bucket，保留原始比例。
    返回 (processed_images, width, height, pad_info)
    pad_info = (pad_left, pad_top, content_w, content_h) 用于推理后裁剪补边。
    """
    if not pil_images:
        raise ValueError("No input images.")

    base_w, base_h = pil_images[0].size

    # 选最佳 bucket（~1MP，比例最接近）
    bucket_w, bucket_h = pick_best_bucket(base_w, base_h, SAFE_BUCKETS, allow_upscale)

    # 等比缩放 fit 到 bucket 内（不拉伸）
    scale = min(bucket_w / base_w, bucket_h / base_h)
    content_w = max(divisible_by, round(base_w * scale))
    content_h = max(divisible_by, round(base_h * scale))

    # 居中补边到 bucket 尺寸
    pad_left = (bucket_w - content_w) // 2
    pad_top = (bucket_h - content_h) // 2
    pad_info = (pad_left, pad_top, content_w, content_h)

    processed = []
    for img in pil_images:
        # 等比缩放
        resized = img.resize((content_w, content_h), Image.LANCZOS)
        # 创建 bucket 大小的画布，边缘用镜像填充减少接缝
        canvas = Image.new("RGB", (bucket_w, bucket_h), (0, 0, 0))
        canvas.paste(resized, (pad_left, pad_top))

        # 用边缘像素填充补边区域（比纯黑效果好）
        import numpy as _np
        arr = np.array(canvas)
        res_arr = np.array(resized)
        # 填充左右
        if pad_left > 0:
            left_col = res_arr[:, 0:1, :]
            arr[pad_top:pad_top+content_h, :pad_left, :] = np.broadcast_to(left_col, (content_h, pad_left, 3))
        right_start = pad_left + content_w
        if right_start < bucket_w:
            right_col = res_arr[:, -1:, :]
            arr[pad_top:pad_top+content_h, right_start:, :] = np.broadcast_to(right_col, (content_h, bucket_w - right_start, 3))
        # 填充上下
        if pad_top > 0:
            top_row = arr[pad_top:pad_top+1, :, :]
            arr[:pad_top, :, :] = np.broadcast_to(top_row, (pad_top, bucket_w, 3))
        bottom_start = pad_top + content_h
        if bottom_start < bucket_h:
            bottom_row = arr[bottom_start-1:bottom_start, :, :]
            arr[bottom_start:, :, :] = np.broadcast_to(bottom_row, (bucket_h - bottom_start, bucket_w, 3))

        processed.append(Image.fromarray(arr))

    return processed, bucket_w, bucket_h, pad_info


def extract_pil_from_source(source) -> Image.Image:
    """从 gr.ImageEditor dict 或普通路径/PIL 中提取图片（使用 composite 保留涂色标注）。"""
    if source is None:
        return None
    if isinstance(source, dict):
        img = source.get("composite")
        if img is None:
            img = source.get("background")
        if img is None:
            return None
        if isinstance(img, np.ndarray):
            return Image.fromarray(img).convert("RGB")
        return img.convert("RGB")
    return load_pil_image(source)


def format_info(seed_val, source_img, ref_img):
    lines = [f"**Seed:** `{int(seed_val)}`"]

    for label, img in [("Source", source_img), ("Reference", ref_img)]:
        if img is None:
            continue
        try:
            pil = extract_pil_from_source(img) if label == "Source" else load_pil_image(img)
            ow, oh = pil.size
            nw, nh = pick_best_bucket(ow, oh, SAFE_BUCKETS, UPSCALE_SMALL_IMAGES)
            lines.append(
                f"\n**{label}:** {ow}×{oh} → **{nw}×{nh}** "
                f"(ratio {ow/oh:.3f} → {nw/nh:.3f})"
            )
        except Exception:
            pass

    return "\n\n".join(lines)


# =========================================================
# INFERENCE
# =========================================================
@spaces.GPU
def infer(
    source_image,
    ref_image,
    prompt,
    negative_prompt,
    seed,
    randomize_seed,
    guidance_scale,
    steps,
    color_match,
    out_width=0,
    out_height=0,
    progress=gr.Progress(track_tqdm=True),
):
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    if source_image is None:
        raise gr.Error("⚠️ Please upload a source image.")
    if not prompt or not prompt.strip():
        raise gr.Error("⚠️ Please enter an edit prompt.")

    # 提取原图（兼容 ImageEditor dict 和普通路径）
    try:
        src_pil = extract_pil_from_source(source_image)
    except Exception as e:
        raise gr.Error(f"⚠️ Could not load source image: {e}")
    if src_pil is None:
        raise gr.Error("⚠️ Please upload a source image.")

    # 记录原始尺寸，推理后 resize 回来，避免 16 对齐导致裁剪
    orig_size = src_pil.size  # (w, h)

    # ── 路由：抠图 ──
    if "抠" in prompt:
        if "黑底" in prompt:
            # 黑底花字：连通区域去除外围黑色，保留文字内部黑色
            result = remove_black_bg(src_pil)
        else:
            # 普通抠图：RMBG 2.0 语义分割
            result = run_rmbg(src_pil)
        return result, seed

    # 收集图片：原图必须，参考图可选
    pil_images = [src_pil]

    if ref_image is not None:
        try:
            pil_images.append(load_pil_image(ref_image))
        except Exception as e:
            print(f"Warning: could not load reference image: {e}")

    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    generator = torch.Generator(device=device).manual_seed(int(seed))

    processed_images, width, height, pad_info = prepare_images_before_pipe(
        pil_images, allow_upscale=UPSCALE_SMALL_IMAGES
    )

    # 显式指定输出尺寸（对齐 ComfyUI EmptyLatentImage 行为）
    if out_width > 0:
        width = (out_width // 16) * 16
    if out_height > 0:
        height = (out_height // 16) * 16

    try:
        result = pipe(
            image=processed_images,
            prompt=prompt,
            negative_prompt=negative_prompt,
            height=height,
            width=width,
            num_inference_steps=steps,
            generator=generator,
            true_cfg_scale=guidance_scale,
        ).images[0]

        # ── 裁掉补边，还原到原始比例内容区域 ──
        pad_left, pad_top, content_w, content_h = pad_info
        if pad_left > 0 or pad_top > 0 or content_w < width or content_h < height:
            result = result.crop((pad_left, pad_top, pad_left + content_w, pad_top + content_h))

        # ── 还原到原始尺寸 ──
        if result.size != orig_size:
            result = result.resize(orig_size, Image.LANCZOS)

        if ref_image is not None and len(pil_images) > 1:
            result = add_image_watermark(result, pil_images[1])

        if color_match:
            # 用原图背景（无笔迹）作为色彩参考
            if isinstance(source_image, dict):
                bg = source_image.get("background")
                if bg is not None:
                    ref_pil = Image.fromarray(bg).convert("RGB") if isinstance(bg, np.ndarray) else bg.convert("RGB")
                else:
                    ref_pil = src_pil
            else:
                ref_pil = src_pil
            ref_pil_resized = ref_pil.resize(result.size, Image.LANCZOS)
            result = color_match_reinhard(ref_pil_resized, result)

        return result, seed

    finally:
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()


# =========================================================
# UI
# =========================================================
# JS：等 ImageEditor 渲染完后，把绝对定位的工具栏改为相对定位，
# 使其不再悬浮覆盖画布（CSS 选择器会被 Svelte 作用域哈希阻挡，
# 所以用 JS 通过 getComputedStyle 精确检测并强制修改）
_FIX_TOOLBAR_JS = """
() => {
    const setup = (ed) => {
        if (ed.dataset.toggleReady) return;
        // 找工具栏元素（Gradio/Svelte 会给 class 加哈希，用 includes 匹配）
        const toolbar = Array.from(ed.querySelectorAll('*')).find(el => {
            const cls = el.getAttribute('class') || '';
            return cls.includes('toolbar') || cls.includes('tool-bar');
        });
        if (!toolbar) return;
        ed.dataset.toggleReady = '1';

        // 插入切换按钮，放在 toolbar 的父容器第一位
        const btn = document.createElement('button');
        btn.className = 'toolbar-toggle-btn';
        btn.textContent = '🎨 隐藏画笔工具栏';
        let hidden = false;
        btn.onclick = () => {
            hidden = !hidden;
            // 用 visibility 而非 display，避免画布区域跳动
            toolbar.style.visibility = hidden ? 'hidden' : '';
            toolbar.style.pointerEvents = hidden ? 'none' : '';
            btn.textContent = hidden ? '🎨 显示画笔工具栏' : '🎨 隐藏画笔工具栏';
        };
        toolbar.parentNode.insertBefore(btn, toolbar);
    };

    const mo = new MutationObserver(() => {
        document.querySelectorAll('.src-editor').forEach(setup);
    });
    mo.observe(document.body, { childList: true, subtree: true });
    setTimeout(() => document.querySelectorAll('.src-editor').forEach(setup), 1000);
}
"""

with gr.Blocks(
    theme=theme,
    js=_FIX_TOOLBAR_JS,
    css="""
.gradio-container {
    max-width: 1400px !important;
    margin: 0 auto;
    padding-top: 20px;
}
.hero {
    text-align: center;
    padding: 24px 0 12px 0;
}
.hero h1 {
    font-size: 2.2rem;
    font-weight: 800;
    margin-bottom: 8px;
}
.hero p {
    font-size: 1rem;
    color: #666;
    margin-bottom: 0;
}
/* 工具栏隐藏时，隐藏按钮仍可点击 */
.toolbar-toggle-btn {
    display: block;
    width: 100%;
    padding: 4px 10px;
    margin-bottom: 2px;
    background: #f0f0f0;
    border: 1px solid #ddd;
    border-radius: 4px;
    font-size: 12px;
    cursor: pointer;
    text-align: left;
    color: #555;
}
""",
) as demo:
    gr.HTML("""
    <div class="hero">
        <h1>🔥 FireRed Image Edit 1.1 Fast</h1>
    </div>
    """)

    with gr.Tabs():

        # ══════════════════════════════════════════════════════
        # Tab 1: AI 编辑
        # ══════════════════════════════════════════════════════
        with gr.Tab("AI Edit"):
            with gr.Row():
                with gr.Column(scale=1):
                    source_input = gr.ImageEditor(
                        label="Source Image — 可用画笔标注区域（红/绿/蓝等），提示词中引用颜色",
                        elem_classes=["src-editor"],
                        brush=gr.Brush(
                            colors=["#FF0000", "#00CC00", "#0066FF", "#FFFF00", "#FF00FF", "#FFFFFF"],
                            color_mode="defaults",
                        ),
                    )
                    gr.Markdown(
                        "<small>🔴红 🟢绿 🔵蓝 🟡黄 🟣紫 ⬜白 — 画好后提示词写：*去掉红色标注的区域* 等</small>"
                    )
                    with gr.Row():
                        ref_input = gr.Image(
                            label="Reference Image（参考图，可选）",
                            type="filepath",
                            sources=["upload", "clipboard"],
                        )

                    prompt_input = gr.Textbox(
                        label="Prompt",
                        placeholder="Describe how you want to edit the image...",
                        lines=4,
                    )
                    negative_prompt_input = gr.Textbox(
                        label="Negative Prompt",
                        value=DEFAULT_NEGATIVE_PROMPT,
                        lines=3,
                    )

                    color_match_input = gr.Checkbox(label="Color Match — 色彩对齐原图", value=True)

                    with gr.Accordion("Advanced Settings", open=False):
                        seed_input = gr.Slider(
                            label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0,
                        )
                        randomize_seed_input = gr.Checkbox(label="Randomize Seed", value=True)
                        guidance_scale_input = gr.Slider(
                            label="Guidance Scale", minimum=1.0, maximum=10.0, step=0.1, value=1.0,
                        )
                        steps_input = gr.Slider(
                            label="Inference Steps", minimum=1, maximum=50, step=1, value=4,
                        )

                    run_button = gr.Button("Generate", variant="primary")
                    info_markdown = gr.Markdown()

                with gr.Column(scale=1):
                    output_image = gr.Image(label="Result", type="pil")

            for trigger in [source_input, ref_input, seed_input]:
                trigger.change(
                    fn=format_info,
                    inputs=[seed_input, source_input, ref_input],
                    outputs=[info_markdown],
                )

            run_button.click(
                fn=infer,
                inputs=[
                    source_input, ref_input, prompt_input, negative_prompt_input,
                    seed_input, randomize_seed_input, guidance_scale_input, steps_input,
                    color_match_input,
                ],
                outputs=[output_image, seed_input],
            ).then(
                fn=format_info,
                inputs=[seed_input, source_input, ref_input],
                outputs=[info_markdown],
            )

        # ══════════════════════════════════════════════════════
        # Tab 2: PNG 贴图（画 mask → 等比贴入）
        # ══════════════════════════════════════════════════════
        with gr.Tab("PNG Placement"):
            gr.Markdown("**用法：** 上传底图后在图上涂抹出放置区域，再上传 PNG，点击 Apply。PNG 会等比缩放，最长边对齐 mask 最长边，居中贴入。")
            with gr.Row():
                with gr.Column(scale=1):
                    mask_editor = gr.ImageEditor(
                        label="Source Image — 在图上涂抹出放置区域",
                        brush=gr.Brush(colors=["#FF6B35"], color_mode="fixed"),
                    )
                    png_input = gr.Image(
                        label="PNG to place（支持透明背景）",
                        type="numpy",
                        sources=["upload", "clipboard"],
                        image_mode="RGBA",
                    )
                    apply_button = gr.Button("Apply", variant="primary")

                with gr.Column(scale=1):
                    placement_output = gr.Image(label="Result", type="pil")

            apply_button.click(
                fn=paste_png_into_mask,
                inputs=[mask_editor, png_input],
                outputs=[placement_output],
            )

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