README / README.md
sjeozq's picture
Links: GitHub repository Glance-X-Lite is now Glance-X-MLX
987f5c2 verified
|
Raw History Blame Contribute Delete
3.45 kB
metadata
title: README
emoji: 👁️
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false

Demo: 256 images, one request, a scan picks the right image, the view zooms in, and new requests find each target in it

Vision-Decision builds Glance, which answers many questions about one image in one forward pass, returning typed decisions: a box, yes or no, a choice, or not there.

视觉决策模型:一次前向,回答一张图上的许多问题。

Why. Agents mostly need decisions, not prose: Where is the button? Is it there at all? Which clip shows it? Generative VLMs write every answer token by token, which is slow and turns "not there" into a guess. Glance reads them all from one pass.

Models

model what it is
Glance-4B-preview a 4B decision VLM on Qwen3.5-4B: boxes, yes/no, choices and scores, with calibrated "not there", from one prefill
Glance-X-4B bidirectional grounding: reuse the image across requests or the request across images, with multi-box output
Glance-X-Lite a 0.8B Glance-X for Apple silicon and other small devices

Data

dataset what it is
Glance-COCO grounding on COCO images with exact boxes, many phrasings, Chinese requests, hard negatives and multi-target queries; plus training supplements, trainer-format records and the robust evaluation set (formerly Glance-Data)
Glance-GUI synthetic GUI grounding: 18.5k screenshots, 1.08M requests with verified hard negatives
Glance-Video typed-question records for short video clips (annotations only)
Glance-Experiments results and model outputs: the Decision Index run, the referring-negative audit, scan-engine measurements

Browse by collection: Glance models · Glance training data · Glance evaluation · Collaborator projects

Code

Glance (inference package) · Glance-Core (training, data builders, evaluation) · Glance-Engine (serving engine) · Glance-Video (video and clip scanning) · Glance-X-MLX · Glance-COCO (data pipeline)