Carmenta

Optical character recognition in pure Rust — documents, screens, and live frames.

Carmenta reads text out of images: detection, mixed-case recognition, and reading order computed from geometry rather than a layout model. It runs on candle, downloads its weights from hash-verified manifests, and has no Python runtime, no C/C++ by default, no gated models and nothing GPL.

It is the OCR component of FFai, named for the Roman goddess who adapted the Greek alphabet into Latin letters.

[dependencies]
ffai-carmenta = "0.6"
ffai-core = "0.6"
ffai-media = "0.6"

---

What it does well, measured

Documents

On the OmniDocBench English holdout — 236 real pages, Apache-2.0, the benchmark Baidu's Unlimited-OCR states its record on:

	CER	correctness
mobiledet-crnn	18.88 % micro · 30.25 % macro	236/236

Reading order is computed by a recursive XY-cut over detector boxes, routed per-node between a column grid and a projection cut. No layout model, no extra weights. It is worth 5.9 points of CER, and 3.4x against reading top-to-bottom. The page selects among three orderings at run time on a column-coherence score.

Against the state of the art

43 pages, all three engines through the same harness and the same metric:

	CER	pages/s	memory
Unlimited-OCR (Baidu, 3B MoE, GPU)	15.51 %	0.01	8745 MiB peak
PP-StructureV3	19.14 %	0.02	1481 MiB steady
Carmenta (CPU)	23.76 %	0.15	425 MiB steady

Carmenta's row was 25.91 % before per-node reading-order routing landed; the same 43 pages, same harness, same metric now read 23.76 %. The deficit against the 3B GPU reference fell from 10.40 pp to 8.25 pp — a 21 % reduction, from ordering alone, with no new weights.

Behind on quality. 17x the throughput with no GPU, on 4.7 MB of detector weights against 6.4 GB — a deployment class neither reference can enter.

LIVE — a streaming mode nothing mainstream ships

Point it at a sequence of frames and it behaves like a stream instead of a loop over images. A frame whose pixels have not moved reuses the previous result, so text cannot flicker on an unchanged frame — the failure that makes naive per-frame OCR unusable for screen reading.

On a pinned 180-frame screencast:

	Carmenta LIVE	per-frame Tesseract
OCR calls for 24 text changes	24	180
churn on 156 unchanged frames	0	24
CER on change frames	1.21 %	—

LIVE also beats the same engine in batch mode on the same frames (1.21 % vs 1.74 %).

Photographs — the open front

On photographed receipts the full pipeline trails PaddleOCR (20.9 % vs 15.6 %) even though the recognition stage beats theirs on identical crops (1.5 % vs 3.0 %). The cause is diagnosed — tilt-sensitive line grouping — and deskew is the named fix. It is stated here because a page that only lists wins is not a measurement.

---

Use it

use ffai_core::engine::{OcrEngine, OcrOptions};
use ffai_carmenta::engine::{CraftCrnn, RecStage};

let engine = CraftCrnn::new_mobiledet(RecStage::Crnn);
let out = engine.recognize(&ffai_media::load_image("page.png")?,
                           &OcrOptions::default())?;
println!("{}", out.text());

Engines are chosen by lineage name, like codecs in ffmpeg — mobiledet-crnn for documents, craft-crnn for screens, craft-parseq for photographs. The choice stays yours; only the per-image segmentation strategy auto-dispatches, on a measured signal with an empty band between the content classes.

ffai ocr -i page.png --engine mobiledet-crnn
ffai ocr --live --watch 5 -i captures/ -o screen.srt

---

Honest scope

• Frame sources. LiveSession takes pixels and is decoder-agnostic. A directory of stills decodes through rusty_png/rusty_jpeg; video ingest is not built; camera/MJPEG input is untested. "Point it at a screen" today means frames on disk.
• Determinism. Run-to-run CER varies ~0.5 pp on identical inputs (non-deterministic parallel reduction). Differences below that are not results.
• Batch parity in LIVE. The harness reports parity breaks against batch mode; they trace to the change gate's tolerance rather than to lost accuracy — the quality gate has LIVE ahead of batch. The gate definition is being fixed rather than explained away.
• Status is experimental. Documents are the strongest surface; photographs trail PaddleOCR.

Every figure above is a line in bench/ledger.jsonl. The mission plan records the losses beside the wins — five refuted ordering variants, two retracted findings, and the instrument errors that produced them.

License

MIT OR Apache-2.0. Model weights carry their own licenses, surfaced at selection time.