# Taste

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Santosh Shakya](https://cerebralvalley.ai/u/asantosh)
- **GitHub:** https://github.com/asantosh666/taste
- **Demo video:** https://youtu.be/zlapPZEAKgQ?si=HEpSqi-1q-IN-S16
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/991

Every shoot ends the same way: hundreds of frames, and only a handful of keepers. Finding them is the job, and it is slow, repetitive, and easy to get wrong when near-identical bursts blur together. Taste is a culling assistant built for exactly that pass: rank the shoot, pick one winner per burst, and export the keep-set.

What it does:

- Ranks every frame best-first from sharpness, exposure, and live vision scores.
- Finds near-duplicate groups (35 groups across the 152-frame Nepal shoot in the demo) so you compare within a burst instead of across the whole shoot.
- Reviews keyboard-first: P keep, X reject, U clear, M maybe, arrows to move, Z for 100% zoom. Flagging auto-advances, so a pass stays in flow.
- Compares group members side by side with a synced view, a per-factor delta table, deterministic reason cards, and a contested flag when the scores disagree.
- Builds a coverage-aware keep-set: ask for 20 keeps and it covers 20 groups and 20 scenes, one winner per group first, then the best of the rest.
- Exports the keep-set plus a contact sheet.

How it uses Meta's AI tools, explicitly:

- Muse Spark (via the Meta Model API) scores composition and moment for every frame — the qualities heuristics cannot read. Those scores appear beside each frame in the UI.
- SAM 3.1 (via the Meta Model API) segments the subject so Taste measures sharpness on the subject, not just somewhere in the frame, and powers the subject strip in compare view.

The honest part — and the reason the demo shows its evaluation panel instead of hiding it:

- In-sample separation is +0.072 (keepers 0.536 vs rejects 0.464), but held-out accuracy on unseen frames is about one half. The ranking fits the labels it trained on and does not generalize.
- SAM covers only 58 of 152 frames here and gave no ranking lift.
- A rule discovered from the data (lowest contrast wins inside bursts) then failed a pre-registered fresh-photo test: 3 of 11 against ~3.45 by chance.

So Taste shows its evaluation rather than claiming it learned your eye. The workflow — ranking, grouping, keyboard review, keep-set, export — is real and working; the learning claim is one we tested and did not make.

## More from Meta Global AI Developer Hackathon

- [Sahaj](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/988)
- [Saliga](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/989)
- [SiteSidekick](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/990)
- [IncidentLab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/992)
- [Relay Systems](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/993)
- [Vary_artifacts](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/994)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/991. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt