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

Status: v0.16.0 keeps the full Vision request surface, adds a Tier-1 async_api module for one-shot OCR / face / barcode / segmentation workflows, and ships a fully-implemented COVERAGE.md + COVERAGE_AUDIT.md matrix plus a gold-standard multi-file Swift bridge.

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

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

All request-type modules can be enabled independently, and the default feature set still enables the full Vision surface. v0.16.0 also adds an optional async feature for executor-agnostic Future wrappers around the Tier-1 one-shot request surface.

Enable async plus the request features you need:

apple-vision = { version = "0.16.0", features = ["async", "recognize_text"] }

use apple_vision::async_api::AsyncRecognizeText;
use apple_vision::RecognitionLevel;

# async fn example() -> Result<(), Box<dyn std::error::Error>> {
let texts = AsyncRecognizeText::new(RecognitionLevel::Accurate, true)
    .recognize_in_path("screenshot.png")
    .await?;
println!("found {} text observations", texts.len());
# Ok(())
# }

Tier-1 currently covers background-queue wrappers for OCR, face detection, barcode detection, and person segmentation. Multi-fire delegate / stream-style Vision APIs remain future Tier-2 work.

• 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)
• Header-audited request + observation coverage matrix (COVERAGE.md) with dedicated wrappers for every current request/observation type and a split Swift bridge (all bridge files stay under 500 lines)
• Explicit VNRequest / VNObservation / request-handler / VNVideoProcessor wrappers for OCR pipelines, plus base request/observation helpers reused across the rest of the crate
• Async API (Tier-1 Future wrappers for OCR, face detection, barcode detection, and person segmentation; Tier-2 stream/delegate surfaces still TBD)

Licensed under either of Apache-2.0 or MIT at your option.