# Glance

> Ask an open vision-language model typed questions about an image (yes/no, pick-one, rating) and get probabilities back,
> on your own machine. The answer is read from the logits of one forward pass of a frozen open model (Qwen3-VL-4B by
> default); nothing is generated or trained. Glance is a calibration and measurement harness, not a model.

Use it for private or offline image decisions, many questions per image, probabilities to threshold on, and image-quality
rubrics you can fit from a few dozen images. Zero-shot, pooled over three sets of fresh photographs, the open 4B model is level with the
best hosted models on coarse pick-one questions and about two points behind the best (both Gemini models) on yes/no; on exact rating levels a cheap hosted model (Gemini 3.1
Flash-Lite, 0.763) is ahead zero-shot (0.669), about level once Glance has seen 16 unlabeled images of the rubric (0.758), and
behind after 32 labels (0.857). Know its limits: it confuses mirror-image diagonal directions, misjudges relative size (the best hosted models are
perfect on both), slips when counting eight objects, and on rubrics that are not image quality (tilt, how much of a subject is cut off) even a 300-label
fit reaches only 0.55. For image-quality scoring alone, a small trained quality model or hand-built features do as well or better.

- README (install, usage, scorecard): README.md in the repository
- Operating guide for coding agents: AGENTS.md in the repository
- Every claim with evidence and caveats: docs/CLAIMS.md
- Results, generated from result files: docs/paper/RESULTS_ZEROSHOT.md, results/lab/matrix.md
- Pre-registration notebook: lab/NOTES.md