DeepSeek-OCR: Contexts Optical Compression [pdf] (github.com)

DeepSeek-OCR introduces "Contexts Optical Compression," a task-aware image compression approach designed specifically for document and scene-text images so that compressed artifacts preserve the features critical for Optical Character Recognition (OCR). Instead of optimizing generic visual fidelity, the method jointly trains a learned codec with a differentiable OCR/text-recognition loss and context modules that prioritize text regions and surrounding cues. The result is a compact representation that reduces file size substantially while maintaining or improving OCR accuracy compared with traditional codecs (e.g., JPEG/PNG) and naïve neural compression when evaluated on standard document and scene-text benchmarks. Technically, DeepSeek-OCR combines a neural encoder–decoder with multi-scale context aggregation, entropy modeling for efficient bitrate control, and a task-driven loss that blends rate-distortion objectives with explicit OCR/perceptual penalties. The pipeline uses learned quantization/entropy priors and attention-like context masks to preserve glyph shapes and layout cues most relevant to recognizers. Implications for the AI/ML community include cheaper storage and transmission for large OCR corpora, more reliable on-device OCR under bandwidth constraints, and a pattern for integrating downstream-task objectives directly into compression — a strategy that can be extended to other vision-and-language tasks where semantic preservation outranks pixel-perfect reconstruction.