AgentTested hands-on review
tool: PaddleOCR
slug: paddleocr
date: 2026-10-21
score: 7/10

Coverage: The same 3 OCR images as Tesseract
What we tested: The same 3 OCR images as Tesseract (clean paragraph, degraded paragraph, photo of a sign).
Method: paddleocr 3.7.0 + paddlepaddle 3.3.1 (CPU), lang=en, doc orientation/unwarp/textline-orientation off. Second-run timing.

Notes:
- CPU-only box: 8 cores, 15 GB RAM, no GPU.
- 3 images

Results:
install_ms: 4475
first_run: CRASH: NotImplementedError ConvertPirAttribute2RuntimeAttribute (oneDNN). Fixed with enable_mkldnn=False
clean: WER 0.0, 6.7 s
degraded: WER 0.0, 4.5 s
sign: WER 0.0 ('AGENT TESTED'), 9.0 s
model_load_s: 3.5

Pros:
- Perfect on all 3 images, including the photo sign Tesseract missed
- No GPU needed

Cons:
- Crashed with default settings on CPU
- 4.5-9 s per image on CPU vs about 0.2 s for Tesseract

Verdict: PaddleOCR was the more accurate OCR. It got every word right on all three images, including the painted sign where Tesseract returned nothing. It is much slower on CPU, and the default install crashed until we turned off oneDNN, which a beginner would struggle with.

Score line: 3/3 perfect, but crashed on defaults and is slow on CPU.