A new research concept is challenging one of the most fundamental assumptions in generative AI design: that text should always be processed as text. A group of researchers proposes a counterintuitive approach—converting text into images before feeding it into large language models (LLMs).
At first glance, the idea seems irrational. LLMs are built to understand and generate natural language, not visual imagery of words. Yet this unconventional perspective may hold technical merit, especially when viewed through the lens of efficiency and compression.
Rethinking Tokenization
Modern LLMs rely on a process known as tokenization, which converts text into numerical units called tokens. These tokens are the atomic elements that AI models use to interpret and generate language. However, the number of tokens a model can process at once—the context window—remains limited by hardware constraints. Even today’s advanced systems like GPT-5 and Claude manage hundreds of thousands of tokens, but the ceiling still exists.
Researchers have long pursued ways to compress tokens while retaining precision. Smaller tokens mean larger context windows, enabling models to process longer conversations or more extensive datasets. But compression often comes with a trade-off in accuracy.
Turning Text Into Images
A recent study titled “DeepSeek-OCR: Contexts Optical Compression” (Wei et al., 2025, arXiv) proposes a bold alternative. Instead of encoding text directly, the researchers fed LLMs images of text, effectively treating written language as visual data.
The results were surprising. A single image containing text could represent significantly more information using far fewer tokens. The team achieved up to 10× compression with over 96% decoding precision, suggesting that vision-based encoding could allow models to handle millions of tokens without overwhelming computational resources.
The paper argues that optical representations of language could act as a compression bridge between vision and text models, leveraging techniques from visual language models (VLMs). Essentially, instead of forcing models to interpret raw text line by line, they would process the image of the text, compressing it visually while preserving structure and semantics.
Expert Reactions and Implications
AI pioneer Andrej Karpathy commented on the approach, noting that it raises fundamental questions about whether LLMs should continue relying on text tokens at all. “Maybe it makes more sense that all inputs to LLMs should only ever be images,” he suggested, hinting at a potential shift toward pixel-based natural language processing.
If proven scalable, this technique could drastically expand AI’s capacity for context, reduce token costs, and streamline multimodal training across languages that use pictographic or non-Latin scripts. It may also blur the boundaries between language and vision models, moving toward unified architectures capable of processing all modalities seamlessly.
A Step Toward Unified AI Models
The concept remains experimental and will require further validation. Challenges such as maintaining precision, ensuring speed, and addressing new computational demands remain unresolved. Still, the idea marks a conceptual turning point.
If “optical compression” proves viable, future LLMs might process text more like the human eye—visually, holistically, and contextually. It reinforces a timeless truth that even in machine learning, a picture might truly be worth a thousand words.
Adapted by the IIGAI Editorial Team
Based on original reporting by Dr. Lance B. Eliot, Forbes Contributor