OCR Is Multimodal Coding: A Compressed Sensing View of the DeepSeek-OCR Paradigm Shift

Forum topic · ✨步子哥 · 2026-07-26

Summary

This analysis reframes OCR from a text-recognition tool into a cross-modal compression basis for large language models, drawing on compressed sensing theory (Candes, Tao, Donoho, 2006). It highlights DeepSeek-OCR (2025), which renders document pages as images and encodes content requiring roughly 7,000 visual tokens into just 100 visual tokens while preserving 97% decoding accuracy — a 10:1 compression ratio that sidesteps Transformer attention's O(n²) bottleneck. The post argues that Chinese AI companies (DeepSeek, Baidu, Tencent, Alibaba, Zhipu, iFlytek, Intsig) are not competing on recognition accuracy but racing to own the next compression basis for long-context processing. It also covers DeepSeek's proposal to simulate human forgetting via resolution-decayed conversation history images, the potential of optical computing for encoder workloads, and the extension of 'optical context compression' to code, dialogue history, and knowledge graphs. Key takeaway: finding the right sparse basis yields order-of-magnitude efficiency gains — 'compression as intelligence' is becoming engineering practice.

Overview

This forum post argues that OCR is no longer a recognition tool but a form of multimodal coding — a new compression basis for text — analyzing DeepSeek-OCR through the lens of compressed sensing theory.

Key points

The problem: *War and Peace* is ~800K tokens; Transformer attention is O(n²), so 640 billion attention operations exceed H100 memory. Rendering pages as images changes the math entirely.
DeepSeek-OCR (2025): renders documents as images and represents content needing ~7,000 visual tokens with only 100 visual tokens, keeping 97% decoding accuracy (~10:1 compression). It outperforms GOT-OCR2.0 (256 tokens) and uses less than 1/70 the tokens of MinerU2.0.

The compressed sensing analogy

Compressed sensing (Candès, Tao, Donoho, 2006) showed that if a signal is sparse in some basis, it can be reconstructed from far fewer samples than Nyquist requires — sampling is compression.
Token sequences are a redundant basis for natural language (grammar, repetition, predictable collocations). DeepSeek's insight: text rendered as an image is sparse in the visual modality, because visual encoders capture layout, glyph structure, and formatting in a highly compressible form.
Pipeline: text → render as image → DeepEncoder → 100 visual tokens → MoE decoder → text.

Why LLMs need a new basis

Attention's O(n²) cost makes long context the core bottleneck (8K → 100K → 1M contexts required heavy engineering).
OCR-as-compression effectively multiplies context capacity ~10x, shifting the problem from engineering optimization to paradigm transition. "Coding" here means mapping information from one representation to a more compact one — here, text-to-vision cross-modal encoding.

Chinese vendors' positioning

| Vendor | Approach | Representative models | |---|---|---| | DeepSeek | Optical context compression (DeepEncoder + MoE decoder) | DeepSeek-OCR / OCR-2 | | Baidu | Irregular-box localization, ERNIE-powered | PaddleOCR, PaddleOCR-VL-1.5 | | Tencent | Mobile-optimized, screenshot recognition | Tencent Youtu OCR | | Alibaba | Cloud-scale batch processing | DAMO Academy OCR | | Zhipu AI | GLM multimodal base, CogVIT encoder | GLM-OCR | | Intsig | Vertical industry customization | Qixinbao, CamCard | | iFlytek | LLM semantic understanding fusion | iFlytek LLM OCR |

The strategic divide: most vendors optimize recognition accuracy; DeepSeek redefines OCR as compression — a step-change (70x fewer tokens) vs. incremental accuracy gains. Baidu's PaddleOCR has surpassed Google Tesseract on GitHub, but the real contest is whose visual encoder becomes the standard compression basis.

Bolder ideas from the paper

Simulating human forgetting: conversation history rendered as image pages with resolution decaying exponentially over time — memory and compute shrink with a forgetting coefficient rather than sequence length, enabling effectively unbounded context.
Optical computing: ViT/CNN matrix operations are naturally suited to photonic architectures (lenses perform Fourier transforms, diffractive elements perform convolutions), suggesting encoders could run at light speed with near-zero power.
Beyond documents: code (syntax-highlighted images), dialogue history (chat-bubble images), and knowledge graphs (node-link diagrams) may all be sparse in visual space — each new cross-modal basis implies another order-of-magnitude leap.

Author's conclusion

The paper's significance is underrated: text's visual representation is a more compact basis than token sequences. "Compression is intelligence" is moving from slogan to engineering practice, and China's OCR build-out is a race to own the paradigm shift.

---

References: DeepSeek-OCR: Contexts Optical Compression; Compressed Sensing (Candès, Tao, Donoho, 2006).