提示缓存 on Text Matrixhttps://txtmix.com/tags/%E6%8F%90%E7%A4%BA%E7%BC%93%E5%AD%98/Recent content in 提示缓存 on Text MatrixHugozh-cnTue, 21 Jul 2026 20:06:14 +0800teamchong/pxpipe：把 Claude Code 上下文塞进 PNG，省下 70% tokenhttps://txtmix.com/posts/tech/teamchong-pxpipe-png-image-context-compression/Wed, 08 Jul 2026 19:25:00 +0800https://txtmix.com/posts/tech/teamchong-pxpipe-png-image-context-compression/<h1 id="teamchongpxpipe把-claude-code-上下文塞进-png省下-70-token">teamchong/pxpipe：把 Claude Code 上下文塞进 PNG，省下 70% token</h1> <blockquote> <p><strong>GitHub</strong>: <a href="https://github.com/teamchong/pxpipe" target="_blank" rel="noopener noreffer ">teamchong/pxpipe</a> <strong>Stars</strong>: 4,931 ⭐ <strong>Forks</strong>: 391 <strong>License</strong>: MIT <strong>核心语言</strong>: TypeScript <strong>首次提交</strong>: 2026-05-20</p></blockquote> <p>第一眼看上去反直觉：把文字渲染成图片，再让视觉通道读回去。但 vision token 按像素面积计费，不按字符数——一张 1928×1928 的 PNG 装下 9.2 万字符、只收 4761 个 token；同样的字符按 text 走要 2.5 万 token。压缩比约 5×。</p>