将上下文转成图像让肥波token减少60%

萝卜干蛤蜊 2026-07-05 17:04 1

网上看到个消息说是通过讲上下文转成图片，然后再让大多模态的模型去识别，能够减少60%的token消耗。

我看介绍用的是Fable 5，也不知道国产多模态的Kimi之类的会不会有同样的效果，工具链接是这个，有佬试过吗，是否靠谱？

https://github.com/teamchong/pxpipe

最新回复 (3)

24岁是白给 07-05 17:13

1楼

^-^幽默完了我只能说 fable自己知道自己ocr有这么厉害吗

这种东西你确定你要给肥波?

橘子 07-05 17:15

2楼

ohmypi的snapcompact上下文压缩好像也是这样子的，我用gpt用了几天，感觉压缩的不比普通的上下文压缩差，但是说是token用量更少。

github.com/can1357/oh-my-pi

crates/pi-natives/src/snapcompact.rs

38565c58c

//! Snapcompact frame rendering.

//!

//! Rasterizes pre-normalized conversation text onto a `size`-wide bitmap

//! (height hugs the rows the text actually needs) using one of the bundled

//! public-domain pixel fonts, then encodes it as PNG:

//!

//! - `5x8`  — X.org BDF font (legacy shape).

//! - `8x8`  — unscii-8 hex font (Latin-1 subset), the square cell that won the

//!   snapcompact `SQuAD` evals.

//! - `6x12` / `8x13` — X.org misc BDF fonts (higher-density eval winners).

//! - `silver` — bundled TrueType font for CJK and other non-Latin text.

//!

//! Shape controls, all eval-validated in `packages/snapcompact`:

//!

//! - **variant** — `sent` cycles glyph ink through six hues at sentence

//!   boundaries; `bw` prints plain black ink (best for Anthropic readers).

//! - **lineRepeat** — prints every text line N times; copies after the first

//!   sit on a pale highlight band. Redundancy coding: two looks per glyph at

//!   half the density ("8x8r" shapes).

//! - **cellWidth/cellHeight** — target cell size. When it differs from the

此文件已被截断。显示原始文件

这是minimax的解释：

YaLei Li 07-05 17:35

3楼

感觉都是奇巧淫技，增加响应时间，并没有多大用处。

* 帖子来源Linux.do

附近帖子