Deskhand models

The screen-reading models behind Deskhand, a computer-use agent for Windows. They come from StudioApprentice (Kunal Gupta, UC San Diego), and are published here with permission so that Deskhand can be set up on any Windows PC with one command (setup.ps1, which runs scripts/fetch_models.py).

file	what it is
`weights/l0/`	L0 detector `round-l0-20260906-190904`: classes ['blocking', 'passive'], mAP50 0.972, trained on 11440 screenshots
`weights/l1/`	L1 detector `round-l1-20260907-010157`: classes ['box'], mAP50 0.987, trained on 6650 screenshots
`icon-library/20261002-153904/`	hand-labelled toolbar icons (601 icons in 163 panels) for Blender, Unreal Editor, DaVinci Resolve and others
`SHA256SUMS`	checksums; `fetch_models.py` checks every file against them

The detectors

Both are RF-DETR small (31.8M parameters), input 512 px, fine-tuned on screenshots of creative desktop applications (Unreal Editor, Blender, DaVinci Resolve, MetaHuman, web tools) and Windows itself.

L0 finds what changes how the screen can be used: blocking (a modal, menu or progress dialog the user must deal with first) and passive (tooltips, toasts, status text that cannot be clicked).
L1 finds the application's regions (box): every panel, toolbar, editor and window part a person can act on.

registry_row.json holds each model's training record (classes in slot_names order, mAP, data size); ddp_spec.json its training settings. Load either with the rfdetr package: RFDETRSmall(pretrain_weights=..., resolution=512).

Downloads last month: -; Downloads are not tracked for this model. How to track