DeepSeek-OCR 2: Visual Causal Flow for VLMs скачать в хорошем качестве

DeepSeek-OCR 2: Visual Causal Flow for VLMs

In this AI Research Roundup episode, Alex discusses the paper: 'DeepSeek-OCR 2: Visual Causal Flow' DeepSeek-OCR 2 introduces DeepEncoder V2 to move beyond the rigid 2D raster-scan order used by traditional vision-language models. The architecture utilizes a Qwen2-0.5B component and a dual-stream attention mechanism to achieve what the researchers call visual causal flow. It employs a specialized vision tokenizer that provides 16x token compression to maintain high resolution while lowering computational overhead. By using learnable causal flow queries, the model reorders and distills 2D visual information into a 1D causal sequence based on image semantics. This approach allows for more semantic flexibility and improved document understanding compared to conventional methods. Paper URL: https://arxiv.org/abs/2601.20552 #AI #MachineLearning #DeepLearning #OCR #DeepSeek #ComputerVision #VLM #DocumentAI Resources: GitHub: https://github.com/deepseek-ai/DeepSe...