Introducing Indra, a 1M parameter OCR model: - 49 pages/s on CPU, 11x faster than PaddleOCR and 17x Tesseract - Beats them on OmniParseBench text (41.3 vs 35.8 and 19.3) - Full layout and line bounding boxes

Get it here: Model - Code - Blog post -

Indra works with 60 languages and 6 scripts (Latin, Cyrillic, Greek, Devanagari, Bengali and Chinese), plus any script from a PDF's text layer, and can handle math with a special math mode.

Indra is built for low latency on CPU: RAG ingestion, search indexing, triage before a bigger model, etc. It has limitations because it's such a small model - handwriting, complex tables, and large pages may not work. Try for these.

If you run pip install indra-ocr, then indra-ocr-viewer, you get an app that you can use to test speed and different settings.

We'll continue to improve Indra OCR, including supporting more languages and improving accuracy. Let us know if you have feedback.

woah!

Wow 💨🔥

Amazing, we need faster OCR models, good job!

1M params at 49 pages/s on CPU is wild. Does it handle Chinese?

Yes, chinese works

One banger after another, damn. Great work, curious if we'd see more companies hop away from azure document intelligence to something like this.

noice

Waaao

@datalabto Wooooooowwwww awesome release LFG

🐐

One million parameters is small enough to sit inside the app, and CPU-only at 49 pages a second means the page you scanned never has to leave the device.

Google had a built in OCR model in chrome rhat is really good. Have you tested this? I wonder how it compares.

We haven't benchmarked it directly, not sure if there is an easy way to do that

@VikParuchuri A table spanning multiple pages. Does it handle that too (by handle I mean subsequent pages where the table is continued have table header information in the output even though it’s not explicit in the document page)?

No, this does not take multi-page context into account

Multimodal? Can describe images?

No, this is not an llm, it only does ocr

Works with MRZ ocrb font?

How far is from SOTA in benchmarks for extracting text from scientific papers? Great job @VikParuchuri

what's the twist? impressive latency! how does it compare to lightocr 0.8b?

It's only 1m params, so handwriting, etc, are not going to be as accurate as bigger models

I have a project I shall be using this for that I needed something small for it to work. This is awesome I appreciate it.

Have you already released Varuna and Mitra ? How about Rudra? Savitr ?

fine tune recipe?

nice work guys! any tflite version? i will try to run in a esp32s3

Is it better than Chandra 2?

49 pages per second on a one million parameter CPU model is wild performance, especially for multilingual OCR

This is epic. And just in time for me to benchmark on 800 documents in all asean languages

👀👀👀

The automatic mode's use of the PDF text layer is useful, but I think the headline comparison needs that distinction up front. For scanned documents, the OCR-only score in the README is the number I'd use to choose the pipeline.

I had a dream something like this will happen. The good thing about Tesseract is it is multi lingual. PP-OCRv6 is a reject cause it does not handle multiple languages at same time.

very nice thanks for sharing

How much different does this model provide. Previous I saw unlimited OCR it was having too much noise.

Most ocr models at 1B parameters or more - this is 1M params. You'll see lower accuracy, but much much better speed

@VikParuchuri do u have a white paper regarding how you minimized its latency ? I am curious to learn how u optimized the run time inference. Faster than paddleocr and 17xtesseract.

Available for French Language ?

Yes, french works!

Last time I checked with Indic languages google still performed better for my use case will try again

This only does hindi, since it's such a small model. Will look into other languages soon

Where can I find the list of supported languages?

still using tesseract in a few projects gotta change that as it seemss