rlx-ocr

OCR engine for RLX — native compiled text detection + recognition graphs. Loads HuggingFace robertknight/ocrs weights as .safetensors and runs detection + recognition on every standard RLX backend. No Python and (on the default rlx feature) no RTen engine — host-side pad/resize plus RLX graphs.

Quick start

# Detection + recognition on a single image (models auto-resolved from --model-dir).
just ocr --model-dir .cache/ocrs --image page.png --device cpu

# or directly:
cargo run -p rlx-ocr --bin rlx-ocr --release -- \
  --model-dir .cache/ocrs --image page.png --device metal

CLI flags: --model-dir DIR (expects ocr-*-full.safetensors) or both --detection-model PATH and --recognition-model PATH; --image PATH; --device cpu|metal|mlx|cuda|rocm|gpu|vulkan; --dry (skip inference). The rlx-ocr-convert bin (feature convert-rten) exports legacy .rten checkpoints to .safetensors.

Public API

use rlx_ocr::{OcrRunner, OcrOutput};
use rlx_runtime::Device;

let runner = OcrRunner::builder()
    .model_dir(".cache/ocrs")
    .device(Device::Cpu)
    .build()?;

let out: OcrOutput = runner.predict_path("page.png".as_ref())?;
println!("{}", out.text);

// Or straight from decoded RGB bytes:
let text: String = runner.predict_text("page.png".as_ref())?;
# anyhow::Ok(())

Lower-level entry points: OcrEngine / OcrEngineParams (detection + recognition pipeline), ocr_rgb_bytes, and prepare_image. Text geometry is returned as TextLine / TextWord / TextChar with RotatedRect boxes; CTC decoding lives in ctc.

Backends & features

Tests & benches

just test-ocr-backends       # detection + recognition + pipeline on each backend (needs OCR_MODEL_DIR)
just bench-ocr-rlx           # native rlx-ocr latency vs upstream ocrs
just test-ocr-parity         # rlx pipeline vs ocrs reference (parity-ocrs feature)