Snapcompact: SoTA Compaction – Instant, Local, Free. Pick 3 (blog.can.ac)

Snapcompact introduces a novel approach for enhancing long-context interactions in AI models by converting text into dense pixel-font bitmaps. This method allows a 1568×1568 PNG image to store approximately 40,000 characters, essentially offering a cost-effective way to manage context—the model retrieves the text almost verbatim for one-third of the traditional input token cost. What began as a joke led to significant performance benchmarks, showing that with the right optimizations, particularly in font choice, Snapcompact can preserve crucial context information without degrading model performance. This technique is particularly significant for the AI/ML community as it addresses the challenge of maintaining coherent narrative continuity during extended interactions. Traditional compaction methods often compromise model output quality, whereas Snapcompact maintains high recall rates for identifiers and context. Early experiments indicate that while token savings can come with increased decoding costs, the overall efficiency—achieving an F1 score of 0.86–0.96 across different models—suggests strong applicability for both current and future AI systems. The findings not only showcase a practical advancement in content handling but also open the door for further exploration of dense visual media as effective context carriers in AI.