squish?!

experimental repository for testing text -> image long-context compression

inspired by

• DeepSeek-OCR: Contexts Optical Compression https://github.com/deepseek-ai/DeepSeek-OCR/blob/main/DeepSeek_OCR_paper.pdf
• Glyph: Scaling Context Windows via Visual-Text Compression https://arxiv.org/pdf/2510.17800

Successfully does -50% compression of medium-length haystack tasks (ie. 16k text tokens = 8k image tokens)

• further optimizations get me a 70k/10k ratio (-86% ish?) not sure tho

todos:

• why is image slower?
• can you further compress
• proper textwrap framing
• run on OSS model not openai

page: