Optical Character Recognition

Evidence (10)

• 7Cloudflare Blog·10/1/2026product_launch
  Cloudflare AI Search reaches general availability

  Cloudflare has made AI Search generally available, offering a managed index and retrieval pipeline built on Workers AI, Vectorize, R2, and Browser Run. The update adds native image embeddings, OCR for PDFs, and support for larger files, with billing to begin on November 1, 2026 while a free tier remains available.

• 6Hacker News·8/21/2026research
  Call to digitize rare books before physical loss

  The article argues that AI companies and other large-scale users of books for training data can contribute to the physical destruction of rare volumes, making preservation more urgent. It urges faster scanning and digitization of rare books before more items are lost or damaged.

• 6Hacker News·8/13/2026product_launch
  Mistral releases OCR 4.1 public preview

  Mistral has launched OCR 4.1, a public preview for its Document AI stack. The service adds native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores, with pricing listed per 1,000 pages and annotated pages.

• 4Hacker News·7/16/2026product_launch
  Decoy Font hides text from AI readers

  Decoy Font is a TTF typeface that uses hybrid-image style spatial frequencies to show different text to humans and AI/OCR systems depending on viewing distance. The project is positioned as an anti-scraping and anti-OCR tool, but it is not a guaranteed defense against more capable agents.

• 4Hacker News·7/11/2026research
  Ghost Font hides messages from AI models

  Ghost Font is an experimental message format that uses motion, video, noise, and decoy elements so humans can read text while AI models struggle to decode it. The project compares this approach with older anti-OCR fonts like ZXX and claims that static screenshots do not reveal the embedded message.

• 2Hacker News·7/8/2026product_launch
  Decoding the obfuscated Bash script on a Uniqlo T-shirt

  A Hacker News post describes finding an obfuscated, self-evaluating Bash script printed on the back of an Akamai-designed Uniqlo T-shirt. The script is an Easter egg that decodes a Base64-encoded payload and displays a celebratory message, discovered after OCR and manual cleanup of the printed text.

• 6Hacker News·7/3/2026open_source
  Pxpipe cuts Fable costs by converting code to images for OCR

  The pxpipe project reports a 60% reduction in Fable's costs by converting source code into images and then having a model read them via OCR. The approach is presented as a practical optimization for lowering model processing costs while preserving the underlying coding task.

• 6Hugging Face Blog·6/22/2026model_release
  PP-OCRv6 Brings 50-Language OCR to Hugging Face

  Hugging Face blog introduced PP-OCRv6, the latest PaddleOCR model family for detection and recognition across documents, screenshots, multilingual images, and scene text. The lineup ranges from 1.5M to 34.5M parameters, with the small and medium models supporting 50 languages and the medium model improving benchmark scores over PP-OCRv5_server.

• 7Hugging Face Blog·6/3/2026research
  DPO Cuts Text Degeneration in OCR Models Beyond Chatbots

  Hugging Face Blog describes how Dharma-AI used Direct Preference Optimization (DPO) after supervised fine-tuning to reduce text degeneration in structured OCR models. Across tested model families, degeneration fell in every case, with an average reduction of 59.4% and a best-case drop of 87.6%. The article argues that DPO can use a model's own failure outputs as rejection pairs, extending preference optimization beyond chatbot alignment.

• 5Hugging Face Blog·5/18/2026framework_update
  PaddleOCR 3.5 adds a Transformers inference backend

  PaddleOCR 3.5 lets supported OCR and document parsing models run with Hugging Face Transformers as an inference backend by setting engine="transformers". The update keeps PaddleOCR managing the OCR/document parsing pipeline while giving developers a more flexible way to integrate models into Transformers-centered workflows.