AgentSkillsCN

remove-background

利用 AI 分割技术去除图像背景。当用户希望从照片或图像中移除、删除背景,或将其设为透明时,可选用此方法。

SKILL.md
--- frontmatter
name: remove-background
description: Remove backgrounds from images using AI segmentation. Use when user asks to remove, delete, or make transparent backgrounds from photos or images.
allowed-tools: Bash, Read

Remove Background

Remove image backgrounds using BiRefNet_lite model. Runs locally on CPU, MPS (Apple Silicon), or CUDA.

When to Use

  • User wants to remove background from an image
  • User wants a transparent PNG version of an image
  • User wants to isolate a subject from its background

Requirements

No API keys required. Model downloads automatically on first run (~100MB, cached).

Command

bash
./scripts/remove_background.py <input_image> [options]

Options

OptionDescription
inputInput image path (required)
-o, --outputOutput path (default: {name}_nobg.png)
-c, --cropSmart crop to foreground bounding box
-p, --paddingPadding around crop in pixels (default: 0)
--deviceForce device: cuda/mps/cpu (default: auto-detect)
-f, --formatOutput: json (default) or table

Examples

bash
# Basic usage - outputs photo_nobg.png
./scripts/remove_background.py photo.jpg

# Smart crop to subject
./scripts/remove_background.py photo.jpg --crop

# Smart crop with 20px padding
./scripts/remove_background.py photo.jpg --crop --padding 20

# Custom output path
./scripts/remove_background.py photo.jpg -o transparent.png

# Force CPU (if MPS has issues)
./scripts/remove_background.py photo.jpg --device cpu

# Human-readable output
./scripts/remove_background.py photo.jpg --format table

Output Format

JSON (default):

json
{
  "input": "photo.jpg",
  "output": "photo_nobg.png",
  "device": "mps",
  "original_size": [1920, 1080],
  "output_size": [800, 600],
  "cropped": true,
  "crop_box": [120, 80, 920, 680],
  "model": "ZhengPeng7/BiRefNet_lite"
}

Notes

  • First run downloads model (~100MB), subsequent runs use cache
  • Output is always PNG with alpha channel (transparency)
  • Device auto-detection: CUDA > MPS > CPU
  • MPS (Apple Silicon) may have some op fallbacks to CPU