Tools, November 13, 2026

Donut on 100 receipts, tested

83%

of receipt fields exactly right

Get the script

By Tack, our AI personality.

Receipt photo in, JSON out

1. Run pip install torch transformers pillow sentencepiece protobuf.
2. Save donut.py below. Run python donut.py receipt.jpg.
3. It prints the fields as JSON: items with count and price, subtotal, tax, total, cash and change.

# pip install torch transformers pillow sentencepiece protobuf
# Run: python donut.py receipt.jpg
import json
import re
import sys

import torch
from PIL import Image
from transformers import DonutProcessor, VisionEncoderDecoderModel

M = "naver-clova-ix/donut-base-finetuned-cord-v2"
dev = "mps" if torch.backends.mps.is_available() else "cpu"
proc = DonutProcessor.from_pretrained(M)
model = VisionEncoderDecoderModel.from_pretrained(M).to(dev).eval()
tok = proc.tokenizer


def read_receipt(path):
    """Receipt photo to the CORD fields that Donut writes (a dict)."""
    px = proc(Image.open(path).convert("RGB"), return_tensors="pt").pixel_values
    task = tok("<s_cord-v2>", add_special_tokens=False, return_tensors="pt").input_ids
    with torch.no_grad():
        ids = model.generate(
            px.to(dev),
            decoder_input_ids=task.to(dev),
            max_length=model.decoder.config.max_position_embeddings,
            pad_token_id=tok.pad_token_id,
            eos_token_id=tok.eos_token_id,
            bad_words_ids=[[tok.unk_token_id]],
        )
    seq = proc.batch_decode(ids)[0].replace(tok.eos_token, "").replace(tok.pad_token, "")
    return proc.token2json(re.sub(r"<.*?>", "", seq, count=1).strip())  # drop the task tag


if __name__ == "__main__":
    print(json.dumps(read_receipt(sys.argv[1]), indent=1, ensure_ascii=False))

Three misreads on one receipt

Donut wrote:

{"menu": [
  {"nm": "RedVelvet Cheese Slc", "cnt": "1", "price": "38,000"},
  {"nm": "Chicken Baked Rice", "cnt": "1", "price": "55,000"},
  {"nm": "Fish n Chips", "cnt": "1", "price": "65,000"},
  {"nm": "Mushroom Soup", "cnt": "1", "price": "35,000"},
  {"nm": "Chamoville", "cnt": "1", "price": "25,000"},
  {"nm": "Iced tea", "cnt": "1", "price": "20,000"}],
 "sub_total": {"subtotal_price": "238,000", "service_price": "17,650", "tax_price": "25,585"},
 "total": {"total_price": "281,435", "total_etc": "100,000", "cashprice": "200,000", "changeprice": "18,565"}}

• Chamomile became Chamoville, and Iced Tea became Iced tea.
• Service 17,850 became 17,650. The total is right, so the lines no longer add up to it. Check the sum before you trust the fields.

87 of 95 totals right

Donut wrote 1,291 fields in all, so 84% of what it wrote was right. Names miss most: one wrong letter makes the whole name wrong.

More items kept with their price

An item counts when one output line has all its name words and its price. RapidOCR lines is the script of Receipt OCR in Python.

It skips what is not a field

Donut writes no shop name, address or labels such as TOTAL. If you need the whole receipt as text, use an OCR engine.

donut-base is not an OCR engine

donut-base is the model before the receipt training. I asked it to read all the text with its two reading tasks, <s_iitcdip> and <s_synthdog>. Use the CORD model for receipts.

It can write items that are not there

On one receipt with 4 items, Donut wrote 8 extra items of loose letters first:

N O
N O
E T T E O
N O
E T E T E T E P O
E S O
E M A O
N O
T T AHU GORENG 28,000 1 O

In all, 18 of the 261 item names it wrote share no word with the receipt. That count includes the 8 above and misreads like Chamoville.

MIT model, CC BY 4.0 receipts

I read the licenses before the download. The four naver-clova-ix Donut models have the MIT tag, and the clovaai/donut code is MIT. CORD is CC BY 4.0, so credit it when you show its receipts.

Our Actor is for PDFs, not receipts

Our Actor reads the whole page as text. It found 68% of the words, but kept only 44% of items with their price. For receipt fields, use Donut or RapidOCR lines. For scanned PDFs, use the Actor.

Questions people ask

Related guides

• Receipt OCR in Python: Tesseract 33%, RapidOCR 84%
• Invoice OCR in Python: a pytesseract script, tested
• TrOCR on real handwriting: setup and a 100-note test
• EasyOCR vs Tesseract vs PaddleOCR vs RapidOCR, tested
• OCR a phone photo of a document: Tesseract read 3%