Image-Based Handwriting Recognition: Comparing AI Model Accuracy Across Languages

In this follow-up experiment, we tested how generative AI models handle handwritten text when provided as image input rather than PDF. Unlike in the earlier test, all models were able to process the images and return readable outputs.

Performance Comparison

🏆 Champion as of October 2: Copilot, Gemini 2.5 Pro, GPT-5

Three models—GPT-5, Copilot, and Gemini 2.5 Pro—produced fully accurate results across English, French, and Hungarian, with Gemini showing a marked improvement compared to the PDF case. Claude Sonnet 4.5, Mistral, and DeepSeek-V3.2 performed consistently with their earlier results, handling English and French reliably but introducing errors in the Hungarian version. Qwen3-Max also advanced compared to the PDF test, correctly reproducing the English and French texts from the images, though still struggling with Hungarian. Grok-4, by contrast, showed weaker performance on image input than on PDFs, with more substantial rewriting of the Hungarian text.

Input

For this experiment, we used the same three handwritten samples as in the previous test—one each in English, French, and Hungarian—but provided them as image files rather than PDFs. The instruction given to each model was deliberately minimal: “Return the text from the attached image.”

Output

GPT-5

Gemini 2.5 Pro

Claude Sonnet 4.5, Mistral, DeepSeek-V3.2, Qwen3-Max

Grok-4

Recommendations

This comparison underlines that the format of the input file can shape OCR performance. While image-based input proved more universally accessible—enabling all models to return readable text—it also raises practical considerations. For short handwritten notes, supplying images is straightforward, but for longer or multi-page documents the need to scan or photograph each page makes this approach less efficient than extracting text directly from PDFs.