Unlimited OCR Works
Unlimited OCR Works is an advanced optical character recognition (OCR) model developed by Baidu Inc., designed to efficiently process extensive documents in a single pass. By introducing the Reference Sliding Window Attention (R-SWA) mechanism, it maintains a constant key-value (KV) cache throughout the decoding process, significantly reducing memory consumption and enhancing processing speed. This innovation enables the model to transcribe documents spanning dozens of pages without degradation in performance. The model's architecture is built upon DeepSeek OCR, incorporating the high compression rate of its encoder and the constant KV cache design of R-SWA. Beyond OCR applications, R-SWA serves as a versatile parsing attention mechanism applicable to tasks such as automatic speech recognition (ASR) and translation. The model's code and weights are publicly accessible on GitHub.
Unlimited OCR Works is an advanced optical character recognition (OCR) model developed by Baidu Inc., designed to efficiently process extensive documents in a single pass. By introducing the Reference Sliding Window Attention (R-SWA) mechanism, it maintains a constant key-value (KV) cache throughout the decoding process, significantly reducing memory consumption and enhancing processing speed. This innovation enables the model to transcribe documents spanning dozens of pages without degradation in performance. The model's architecture is built upon DeepSeek OCR, incorporating the high compression rate of its encoder and the constant KV cache design of R-SWA. Beyond OCR applications, R-SWA serves as a versatile parsing attention mechanism applicable to tasks such as automatic speech recognition (ASR) and translation. The model's code and weights are publicly accessible on GitHub.
- 01Efficient processing of extensive documents in a single pass
- 02Reference Sliding Window Attention (R-SWA) mechanism
- 03Constant key-value cache throughout decoding
- 04High compression rate encoder based on DeepSeek OCR
- 05Versatile parsing attention mechanism applicable to ASR and translation tasks
- 06Publicly accessible code and model weights on GitHub