dsh-imagedit: DeepSeek Harness Local Image Processing and Smooth Image Generation Plugin

Introduction¶

The plugin system for DeepSeek Harness (DSH) is designed to solve automation problems in specific scenarios. In image processing, model-generated images often have issues such as unclean backgrounds, jagged edges, or fragmented shapes. The dsh-imagedit plugin provides a deterministic local image-editing pipeline that can perform precise cutout, cropping, and color-adjustment operations, and it can use the generate_image tool to optimize image-generation prompts, reducing broken-shape generation artifacts.

Plugin Overview¶

Name: dsh-imagedit
Maintainer: bbbz123
Category: Workflow
License: MIT

This plugin integrates two modes: flood-fill-based quick cutout and rembg-based AI cutout, and supports Sprite Sheet generation with multi-format export.

Core Features¶

The plugin mainly includes the following capabilities:

Cutout: Supports remove_bg: quick (flood-fill based, no model required, fast) or remove_bg: rembg (AI mode).
Basic operations: trim (trimming), flip h/v (flip), rotate (rotation), padding (padding), canvas WxH (canvas rescaling and centering).
Color adjustment: Adjusts brightness, contrast, and saturation.
Filters: Blur and sharpen.
Finishing: Rounded corners, border (width and color), automatic orientation correction (EXIF).
Batch processing: Supports JSON manifest configuration, recursive directory processing, sprite sheet generation, and sprites.json output.
Export: PNG (automatic compression using oxipng/pngquant), JPEG, WebP (configurable quality).

Installation and Enabling¶

Install the plugin:

    dsh plugin --profile web add github:bbbz123/dsh-imagedit

After restarting `dsh web`, the `image_edit` tool appears in the toolbar.

Adjust the Python interpreter (if needed):
Set the environment variable DSH_IMAGEDIT_PYTHON to point to the target interpreter.

Typical Usage¶

1. Agent Tool Invocation¶

The Agent can invoke image_edit with structured parameters. For example, generating a background-ready icon:

{
  "input": "gen/sword.png",
  "remove_bg": "quick",
  "bg_color": "auto",
  "canvas": "256x256",
  "padding": 16,
  "formats": ["png", "webp"]
}

2. CLI Command Execution¶

The plugin also provides a CLI skill, suitable for local batch processing:

# 单图处理：快速抠图 + 画布调整
python scripts/asset_pipeline.py run --input item.png --remove-bg-quick auto --canvas 256x256 --padding 16 --out-dir output/images/edited

# 组合编辑：旋转、翻转、调色、圆角、边框
python scripts/asset_pipeline.py run --input item.png --rotate 90 --flip h --brightness 1.1 --saturation 1.2 --rounded 30 --border "4,#FF0000" --exports png jpg --out-dir output/images/edited

# 批量处理目录
python scripts/asset_pipeline.py batch --dir ./photos --rotate 90 --exports jpg

# 生成精灵图
python scripts/asset_pipeline.py batch --manifest manifest.json

Recommended Workflow¶

Combining generation with dsh-draw or image_generate, followed by post-processing with image_edit, is the standard workflow for obtaining high-quality asset images:

dsh-draw / image_generate (prompt: "flat solid white background, no shadows")
  → image_edit (remove_bg: quick, canvas, padding)
  → batch manifest with "sheet" for sprite atlas

Environment Dependencies¶

Python 3.10+
Pillow (required)
rembg (optional, only for AI cutout)
oxipng / pngquant (optional, for PNG compression)

Notes¶

The plugin runs with the permissions of the current DSH process.
It is recommended to review the source code and license before installing.
The generate_image tool automatically appends a smoothing suffix to the prompt to correct generation artifacts of GPT-image-2.

GitHub repository: https://github.com/bbbz123/dsh-imagedit
Plugin directory: https://www.skillhub.cn/plugins/bbbz123/dsh-imagedit