Eye: On-device Image Understanding
On-device image understanding that scores any image in one call: best-frame pick, per-axis quality, scene, tags, and a perceptual embedding for burst dedupe.
Eye scores any image in a single call, returning a best-frame pick score, per-axis quality (interestingness, composition, thumbnail quality, subject clarity, aesthetic), a scene class, tags and concepts, an optional identity-blind face analysis, and a perceptual embedding used for near-duplicate burst dedupe. Use it to pick the best photo from a burst, the best frame from a video, or the best thumbnail from a clip.
Use cases
Best-of selection
Pick the best photo from a burst, the best frame from a video, or the best thumbnail from a clip.
Identity-blind face scoring
Score faces as visual content (geometry, expression, attention) for a best headshot, no blinks, eyes on camera. This scoring does not perform face recognition.
What it does
- One call: pick score, quality axes, scene class, tags and concepts.
- Perceptual embedding for near-duplicate burst dedupe.
- Optional
EyeFacesmodule (+22 MB), identity-blind by design.
Specs
- Size
- ~27 MB core, +22 MB optional face module
- Backbone
- TinyCLIP-40M (distilled from a LAION-400M teacher)
- Platform
- Apple (Swift); Android and web planned
FAQ
What is Eye?
On-device image understanding that scores any image in one call: best-frame pick, per-axis quality, scene, tags, and a perceptual embedding for burst dedupe.
Does Eye run on device?
Yes. Eye runs entirely on device: inference happens locally with no server call, so data never leaves the device.
Is Eye available yet?
Eye is in closed beta. You can request early access from its page.
How much does Eye cost?
Every model is free up to 100k monthly active devices per SDK. Unlimited inference per user. Contact us for custom licenses.
Early access
Tell us what you are building and we will get you set up.