dsh-image-zoom

English | 中文

Smart image splitting and zooming for Vision-Language Models (VLMs) inside DeepSeek Harness (dsh).

Most VLMs downsample large images to ~640k pixels (≈800x800) before inference, making fine details in schematics, PCBs, documents, or dense UI screens illegible. This plugin gives your model two surgical tools (split_image / zoom_detail) to see the big picture, then zoom into any numbered region at full resolution — all without leaving the chat, and with persistent attachments.

Why

• Vision models downscale high-res images. Text on a 4K screen or tiny components on a PCB become a blurry mess.
• Asking the model to guess x,y,w,h coordinates on a blurry thumbnail is unreliable.
• This plugin provides a number-labeled grid (drawn directly on the thumbnail). The model just says "Zoom into tile 9", and the engine crops the original full-resolution image exactly there.
• Results are returned as inline images (durable attachments), not temporary file paths.

How it works

You attach a large image (or pass image_path)
        │
        ▼
[split_image] ──► downscaled context + numbered grid overlay
        │              (solid red = cut lines, dashed blue = overlap)
        │
        ▼
Model inspects the labeled overview ──► picks a number (e.g., tile 5)
        │
        ▼
[zoom_detail] ──► crops the ORIGINAL full-res image at that region
        │
        ▼
Inline cropped image returned (can be saved and chained further)

Complete call paths

• Session-gated: remembers per-session enable/disable state via the composer toggle.
• Vision-gated: automatically detects if the current main model accepts image input; refuses to run on pure-text models.
• Coordinate abstraction: you provide pixel coordinates in what you see (the downscaled view); the engine auto-scales them to the original image resolution.
• File chaining: save a zoom result with save_to, then pass that file as image_path to zoom even deeper.

Install

From npm

dsh plugin --profile web add dsh-image-zoom

dsh plugin forwards to pnpm inside the profile, so the package lands in the profile's node_modules. Then append the composition row to ~/.dsh/profiles/web/cordis.patch.yml:

- insert:
    - id: image-zoom
      name: dsh-image-zoom

From source (local checkout)

Clone the repository, then run the cross-platform installer (plain Node — works on Windows, macOS and Linux):

git clone https://github.com/your-username/dsh-image-zoom.git
cd dsh-image-zoom
node install.mjs

The installer copies the package to ~/.dsh/profiles/web/dsh-image-zoom (honors DSH_HOME), links it into the profile's node_modules (junction on Windows, symlink elsewhere), registers the link: dependency in package.json, and appends the required insert row to cordis.patch.yml. It also migrates away the pre-release dsh-photo-zoom name if present.

To uninstall:

node install.mjs --remove

After installation, restart your DSH web server:

dsh --profile web

Tools Overview

1. split_image

Generates a labeled overview grid (attached inline). The grid has:

• Solid red lines: Exact tile cut boundaries.
• Dashed blue rectangles: Actual capture extent including overlap margin (content between the blue edge and red line also appears in the neighboring tile).

Parameters:

• image_index (int, default 0): Most recent attachment.
• image_path (string): Alternative local file (PNG/JPEG/WebP/GIF).
• save_to (string): Optional directory or file path to save the overview.
• type (enum): document, photo, webpage, object, misc — hints for grid strategy.
• rotate (0/90/180/270): Force rotation (default respects EXIF).

2. zoom_detail

Crops a specific region at full original resolution and returns it inline.

Parameters (either view-space coords or tile):

• x, y, w, h (int): Pixels in the downscaled view (auto-scaled to full-res).
• tile (int): Numbered tile from a prior split_image call (easiest).
• label (string): Optional short name for the output image.
• image_index / image_path: Same as above.
• save_to (string): Save crop to disk for chaining.

Configuration

A small composer toggle (sliding switch) appears in the chat input bar when a vision model is active. It controls whether the tools are available for the current session.

• Enabled (default): Tools execute normally.
• Disabled: Tools refuse with a clear error ("smart split is switched off").

The toggle communicates with the host via two simple HTTP routes:

• GET /image-zoom/state – returns current vision status and enabled state.
• POST /image-zoom/enabled – updates the per-session toggle.

No complex settings panel is required; it works out-of-the-box.

Model Experience

What the model sees

• split_image result: A downscaled JPEG with the grid overlay. The model reads the numbers and visual layout.
• zoom_detail result: A PNG of the exact cropped region at native resolution.

Token / KV Cache effect

The tools themselves do not modify the LLM prompt or KV cache. They produce image attachments that are rendered as standard image content blocks in the conversation. The token cost is purely the cost of the rendered image attachments (managed by DSH's native attachment system).

Project Structure

dsh-image-zoom/
├── lib/
│   ├── index.js             # Host plugin (tools, HTTP toggle, subprocess orchestration)
│   └── client.js            # Client half (composer toggle UI)
├── helper/
│   └── split_helper.cjs     # Sharp-based image processing engine (standalone)
├── LICENSE
└── package.json             # DSH bundle manifest (dsh.client declaration)
install.mjs                  # Cross-platform installer (Windows / macOS / Linux)

Development

The helper uses sharp for image manipulation. It is spawned as a subprocess to handle heavy lifting without blocking the main event loop. sharp itself is NOT bundled: the engine discovers the copy that ships with the DSH installation (available on every platform DSH supports), so this package stays pure JavaScript.

To test changes, re-run the installer and restart dsh web.

Publish (maintainer)

cd dsh-image-zoom
npm publish

License

MIT