pxpipe — Reduce LLM Token Usage by Rendering Text Context as Images

Introduction

pxpipe is an open-source tool that reduces LLM token consumption by rendering text-based context (source code, logs, documentation) as images. Vision-capable models can read these images at a fraction of the token cost of the equivalent raw text, making large-context operations significantly cheaper without sacrificing comprehension.

What pxpipe Does

• Renders source code, logs, and text files into optimized PNG images
• Compresses thousands of lines of text context into a single image
• Produces images formatted for optimal readability by vision-capable LLMs
• Integrates as a preprocessing step in agent pipelines and CLI workflows
• Supports configurable font size, syntax highlighting, and layout options

Architecture Overview

pxpipe uses a headless rendering pipeline that converts text input into styled HTML, then captures it as a PNG image using a lightweight browser engine. The rendering optimizes for LLM vision comprehension: monospace fonts, high contrast, line numbers, and syntax highlighting. The output image dimensions and quality are tuned to balance file size against readability, targeting the sweet spot where vision models extract information accurately while token costs stay low.

Self-Hosting & Configuration

• Run directly with npx pxpipe or install globally via npm
• Configure rendering options (font size, theme, columns) via CLI flags or config file
• Supports batch rendering of entire directories or file globs
• Output images can be piped directly into model API calls
• Works with any vision-capable model (Claude, GPT-4o, Gemini)

Key Features

• Dramatic token reduction — thousands of text tokens compressed into one image token block
• Syntax-highlighted code rendering optimized for LLM vision comprehension
• Batch mode for rendering entire project directories
• Configurable themes, font sizes, and layout options
• CLI-first design that integrates into existing agent workflows

Comparison with Similar Tools

• RTK — proxy-based token reducer for CLI output; pxpipe takes a visual approach by rendering to images
• Headroom — compresses tool output via text summarization; pxpipe preserves full content via image rendering
• LLMLingua — prompt compression via token pruning; pxpipe avoids information loss by rendering everything visually
• Screenshots — manual screenshots of code; pxpipe automates and optimizes the rendering for LLM consumption
• Repomix — packs repos into text files; pxpipe converts that text into token-efficient images

FAQ

Q: Which models support image-based context? A: Any vision-capable model including Claude (Sonnet, Opus), GPT-4o, and Gemini Pro Vision.

Q: How much token savings can I expect? A: Typical savings range from 60-90% depending on content type and model pricing for vision versus text tokens.

Q: Does the model lose comprehension from image-based context? A: For well-formatted code and text, vision models achieve comparable comprehension. Highly structured data like tables may benefit from staying as text.

Q: Can I customize the rendering theme? A: Yes, pxpipe supports light and dark themes, configurable font sizes, and multiple layout options for different content types.

Sources

• https://github.com/teamchong/pxpipe