zyh20041227/improved_vision_for_deepseek
Turns one high-resolution image into an overview plus overlapping tiles and a dense-region crop before the model reads it, for receipts and screenshots.
The premise is the documented 384-vision-token cap per image: large images are scaled before inference, which is fine for ordinary photos but erases small characters in receipts, tables, diagrams and long screenshots. The plugin builds three views instead — a global overview, overlapping coverage tiles, and an optional 512x512 detail crop for dense regions — and returns them as bounded batches the model can take in one pass. Coverage tiles are audited against every source pixel for 100% geometric coverage, seams overlap so text crossing a tile edge survives in a neighbour, and document mode moves horizontal seams toward low-ink areas. Each tile carries its source coordinates, role, batch state, coverage and a conservative token cap so the reading can be traced back to a region of the original. The README includes a dense-text benchmark and a technical report on whether the 384-token cap actually degrades reading.
Install
dsh plugin --profile web add github:zyh20041227/improved_vision_for_deepseek#v0.2.3The pinned GitHub release from the README; the same block runs dsh --profile web --dump-config to confirm the row. No allow-build entry is needed. An npm release is planned but not published, so the README's github: form is the only install route — do not substitute a same-named npm package.
Compatibility
DeepSeek Harness plus a vision-capable DSH model profile and Node.js 22 or later. v0.2.3 is pure JavaScript with bundled WebAssembly, so it avoids sharp/libvips conflicts inside DSH Web and needs no native image build; runtime packages are declared in package.json and locked in package-lock.json.
Details
- Repo: zyh20041227/improved_vision_for_deepseek
- Category: Vision & Multimodal
- Stars: 4
- Version: 0.2.3 (GitHub tag); MIT
Alternatives
ethanweave/dsh-vision-better · TZHR-invest/dsh-plugins#dsh-vision-tool · Harvey-Will/dsh-vision-analysis