READMESource: main@b9c25a64
dsh-cutout-workbench
A persistent DeepSeek Harness Web plugin that opens a dark, interactive cutout editor from the left sidebar.
Repository: https://github.com/GYL9999/dsh-cutout-workbench
Install
From npm after publication:
dsh plugin --profile <name> add dsh-cutout-workbench
From a tarball:
dsh plugin --profile <name> add ./dsh-cutout-workbench-0.1.0.tgz
Requires a DSH release in >=0.2.1-rc.4 <0.3.0.
What it does
- Upload JPG / PNG / WebP. The image is decoded locally and the working copy is capped at 1800 px on the long edge.
- Run the vendored ISNet quantized ONNX model through
onnxruntime-webin the browser. No image is sent to a server. - If the model cannot run, fall back to a local colour/edge separation matte and say so explicitly. Low-confidence edge cells are reported by location; they are never silently skipped.
- Preview on a dark checkerboard, or switch to black / white / light gray / custom colour for edge checks. The preview colour is never baked into export.
- Refine with erase and restore brushes, adjustable size and feather. Wheel zoom, space/middle-drag or hand-tool pan, undo/redo and reset to the automatic matte.
- Export a true RGBA PNG: transparent pixels keep alpha 0.
- A natural-language command bar executes edits on the current canvas. The same
operations are available to the agent through the
cutout_workbenchtool; the tool relays an instruction to the open panel over SSE and waits for the panel's result. It does not generate a replacement image.
Local layout
assets/containsresources.json, the 44 MBisnet_quint8model chunks and the ONNX Runtime Web CPU WASM chunks. They are served from/api/cutout-workbench/assets/.client.jsis an esbuild bundle ofsrc/entry.js. Install the build dependencies once withcd _build && pnpm install --registry=https://registry.npmmirror.com, then rebuild withnode build.mjs.index.jsserves the assets, exposes the tool and relays commands.- The package is AGPL-3.0-only because of the bundled IMG.LY code and model
data. See
LICENSE,SOURCE-OFFER.mdandTHIRD_PARTY_NOTICES.md.
Notes
- The model is quantized; hair, fur and translucent edges can still be imperfect. Use the restore brush for the subject edge and erase for residue.
- The browser caches the model chunks after the first run.
- The cutout state lives in the browser tab, not in the session log. Reloading the page clears the current image.
- The interface is intentionally dark and does not follow a light host theme.
No comments yet. Be the first to write one.