HOME SERVICES ARTICLES DUMPS APPS ABOUT
take 2026-07-26 by Stephan Schielke / f59bf179 / llm.md

Fable 5 with omniglyph context compression

#ai #llm #context #compression

A picture is worth a thousand words

That old proverb was a lossy compression claim all along! A thousand words of prose collapse into one image your brain decodes in a glance. Large Language Model (LLM) can do the same trick, and the savings are absurd.

I have been squeezing agent context for a while and in my Compress Yourself Token Rich article I hinted at the possibilities of finding a better standard for lossless text compression. Although not lossless, omniglyph takes the next step and compresses whole context blocks into an image and feeds that to a vision-capable model:

omniglyph example render
@omniglyph example rendering of context as an image.

Why the image vector space packs more in

A block of API docs, config, and log records is expensive as text: every word, brace, and UUID becomes tokens. Render that same block as a compact image and a vision model reads it from a much denser latent space, because the image is projected into a handful of vision embeddings instead of hundreds of text tokens. Same information, a fraction of the token budget.

The screenshot below shows my compression pipeline and token savings using @omniglyph:

OmniRoute omniglyph dashboard showing a full tool-spec, config, and ledger-record page compressed to roughly 84 tokens
Omniglyph context compression logs from my OmniRoute gateway. A whole context page rendered as one image, ≈ 84 tokens.

Going from 150k tokens per request to just 50k really is not bad at all.

The library doing the heavy lifting

The text-to-image encoding leans on pxpipe, a tidy library for packing text into images (and back) so the pixels carry the payload. Pair it with a vision model on the read side and you have a working context-compression pipeline: text in, image on the wire, meaning out.

Same idea as caveman, one layer deeper: caveman shrinks the words, omniglyph shrinks the representation. Stack both and the machine-to-machine layers stop paying human-language prices.

Want these savings on your stack?

Omniglyph is not a toy. It is real token budget back in your pocket, faster inference, and more agents per subscription. If you want image-vector context compression wired into your own agents and gateways, that is exactly the kind of work I do.

Book my services and let us cut your context bill the same way. A picture really is worth a thousand tokens.

Glossary (7)
Tool
OmniRoute Multi-provider LLM gateway that routes, compresses, and meters AI requests across different providers.
pxpipe Library for packing text into images (and back) so the pixels carry the payload. It is the encoding layer for image-as-context compression.
Standard
context compression Reducing the token footprint of an LLM context window while preserving its semantic meaning.
embedding A numeric vector that represents text or image content in a model latent space.
Large Language Model (LLM) AI model trained on large text datasets that generates human-like text responses. Examples: GPT, Claude, Gemma, Llama. Used for coding assistance, text analysis, and increasingly for autonomous system administration.
latent space The internal vector representation a model projects inputs into; denser encodings pack more information per token.
Omniglyph Context-compression technique that encodes an LLM context block as an image so a vision-capable model reads it from a denser latent space, cutting token cost.
keyboard_command_key
/search for stuff Send a prompt to my LLMs /summarize a page Open a chat Use a /command Send a /dm to my phone Ask the /chat assistant Smash your head against the keyboard Hire me /search for stuff
...
CTRL+K