BingL-Li/vision-translation

README excerpt

Translation with Visual Primitives One stable visual-translation Core; any Agent can plug in through a Bridge. 一个稳定的视觉翻译 Core，任意 Agent 都可以通过 Bridge 即插即用。 This project turns an image into structured text. An auxiliary VLM looks at the image, while the Core produces norm-1000 xyxy visual primitives, computes spatial relations from geometry, and extracts OCR. The main Agent consumes text only: it does not handle pixels and does not need a model change. 本项目把图片转换成结构化的 文本：辅助 VLM 负责看图，Core 统一生成 norm-1000 xyxy 视觉基元、由几何计算空间关系并提取 OCR。主 Agent 只消费文本，不接触像素，也不需要修改模型。 The name intentionally echoes DeepSeek's paper Thinking with Visual Primitives . The paper puts visual primitives inside a model; this project reverses the direction by translating vision into primitives outside the model and handing them to an ordinary text model. 项目名有意呼应 DeepSeek 的论文 Thinking with Visual Primitives 。论文让模型在内部使用视觉基元，本项目则把方向反过来：在模型外部把视觉翻译成基元，再交给原本的文本模型。 Architecture: one Core, any Agent, one Bridge each 架构：一个 Core，任意 Agent，各自的 Bridge The call direction is a round trip: the Core calls the auxiliary VLM, the Bridge calls the Core, and the final returns through the Bridge to the Agent. A Python Bridge may import the Cor…

Category