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

DeepSeek-OCR 2: Visual Causal Flow

DeepSeek-OCR 2 introduces a novel vision-language architecture designed to overcome the limitations of traditional raster-scan image processing by implementing a more human-like, causally driven visual flow. At the core of this system is DeepEncoder V2, which replaces standard CLIP components with a compact language model architecture that utilizes learnable queries to dynamically reorder visual tokens based on semantic content rather than rigid spatial coordinates. By employing a specialized attention mask that combines bidirectional visibility for visual tokens with causal attention for query tokens, the model achieves a two-stage cascade of causal reasoning that significantly enhances document understanding. This approach not only maintains efficient token compression rates between 256 and 1120 tokens per image but also delivers a 3.73% performance improvement on the OmniDocBench v1.5 benchmark compared to its predecessor, particularly excelling in accurately reconstructing the logical reading order of complex documents. https://github.com/deepseek-ai/DeepSe... https://huggingface.co/deepseek-ai/De... https://x.com/danielhanchen/status/20...