百度已开源Unlimited-OCR，这是一款专为长文档高效准确的光学字符识别（OCR）设计的新模型。该系统采用“一次性长视野解析”方法，能够通过恒定的KV缓存处理大量文本，从而在OmniDocBench等基准测试中取得最先进的性能。该模型支持与Hugging Face Transformers等流行库以及vLLM和SGLang等推理引擎集成，使其成为文档分析和知识提取的通用工具。 AI
影响 该模型可以显著提高长而复杂文档的文档分析和数据提取效率。
排序理由 主要科技公司（百度）发布新的OCR模型，并声称达到最先进的性能。
AI 生成摘要 · Google Gemini · 来自 8 个来源。 我们如何撰写摘要 →
image-text-to-text · 47 downloads · 55 likes
Baidu Unveils Unlimited-OCR: Constant KV Cache Delivers SOTA Performance on Long Documents
<p><a href="https://lobste.rs/s/5ej4m6/unlimited_ocr_one_shot_long_horizon_ocr">Comments</a></p>
🌘 GitHub - baidu/Unlimited-OCR：無限 OCR 時代：迎接單次長視野解析的革命 ➤ 打造高效能、長文本的工業級 OCR 解析方案 ✤ https:// github.com/baidu/Unlimited-OCR 百度開源了「Unlimited-OCR」專案，旨在進一步推進文檔解析技術的邊界。該工具專注於「單次長視野解析」（One-shot Long-horizon Parsing），能夠高效處理單頁與多頁文件的 OCR 需求。該模型不僅支援 Huggingface Transformers 的標準推理，還針對高效能需求提供了…
🚀 # GitHub and # Baidu introduce "Unlimited OCR: One-Shot Long-Horizon Parsing," proving that even # AI can get lost in its own overcomplicated jargon maze. 🙄 With promises of "direct agents" and "automate any workflow," it's like they've discovered the fax machine of the digital…
Baidu has unveiled Unlimited-OCR, a new model that solves a fundamental bottleneck in long-document transcription. By introducing Reference Sliding Window Attention, it compresses memory from linear to constant growth, achieving 93.92 percent on the OmniDocBench benchmark. The 3B…
Unlimited-OCR: One-shot Long-horizon OCR https://github.com/baidu/Unlimited-OCR # AI # OCR # ComputerVision
🔥 Unlimited OCR: One-Shot Long-Horizon Parsing Researchers have developed a new OCR (Optical Character Recognition) system that can parse long-horizon text with unprecedented accuracy. This technology has significant implications for document scanning and data extraction, and cou…