dsh-text2img-compress

Feeding long text to a vision model often means tens of thousands of input tokens and a context window that fills up before you are done. dsh-text2img-compress takes a straightforward approach: before a message is sent, any text block exceeding a configurable threshold is automatically rendered into a series of 800x800 images, each consuming a fixed 384 tokens, so the model reads the full content from pictures.

For Chinese text, rendering at 18px typically yields a 2-2.5x token reduction, and smaller sizes such as 13px can push that to 4-6x. English text sees roughly 1-2x depending on font size. Critically, every word is preserved in the image, the conversation prefix remains byte-for-byte unchanged, and cache hits are not affected. If the page limit is exceeded or rendering fails, the plugin falls back to plain text-never losing content.

It is built for developers who need a vision model to digest long documents, papers, or changelogs, or who want to reclaim context-window budget for reasoning steps and tool results. If your workflow demands character-exact fidelity-code, SQL, JSON-keep those messages as plain text.