vision

Safe Rust bindings for Apple's Vision framework — on-device OCR, object detection, face landmarks, and other computer vision tasks on macOS.

Status: v0.14 ships the full Apple Vision request surface, including all five stateful tracking requests.

Quick start — OCR

use apple_vision::prelude::*;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let recognizer = TextRecognizer::new()
        .with_recognition_level(RecognitionLevel::Accurate)
        .with_language_correction(true);

    let observations = recognizer.recognize_in_path("/tmp/screenshot.png")?;
    for obs in &observations {
        println!("[{:.2}] '{}'", obs.confidence, obs.text);
    }
    Ok(())
}

Composes with the rest of the doom-fish stack

screencapturekit-rs / capture ──► IOSurface / PNG ──► vision ──► text
                                                          │
                                                          ▼
                                                  foundation-models
                                                  ("summarise this")

Feature flags

All request-type modules can be enabled independently, but the default feature set now enables the full Vision surface, including the new tracking module in v0.14.

Roadmap

Single-image Vision requests (OCR, faces, landmarks, pose, contours, saliency, segmentation, Core ML, and the rest of the stateless request surface)
Pairwise image-registration requests (VNTranslationalImageRegistrationRequest, VNHomographicImageRegistrationRequest)
Stateful tracking requests (VNTrackObjectRequest, VNTrackRectangleRequest, VNTrackOpticalFlowRequest, VNTrackTranslationalImageRegistrationRequest, VNTrackHomographicImageRegistrationRequest)
Async API (VNRequest completion handlers exposed via async fn)

License