# DODO: Discrete OCR Diffusion Models Source: https://www.youtube.com/watch?v=Ds8xtE_xBvY Recap page: https://rapidrecap.app/video/Ds8xtE_xBvY Generated: 2026-02-27T19:03:24.539+00:00 --- ## Quick Overview The DODO (Discrete OCR Diffusion) model successfully addresses the speed-accuracy trade-off in large-scale document processing by introducing a novel block-discrete approach that allows for parallel processing, achieving a 3x speed advantage over standard autoregressive models while maintaining high accuracy, especially when compared to models like DALL-E or Stable Diffusion which struggle with rigid, sequential tasks like OCR. **Key Points:** - DODO (Discrete OCR Diffusion) research from Technion and Amazon Web Services proposes a block-discrete method to solve the speed/accuracy trade-off in large-scale document processing. - The DODO architecture successfully avoids the primary bottleneck of sequential autoregressive models by processing document blocks in parallel, resulting in a 3x speed multiplier over the baseline. - The model achieves a Normalized Edit Distance (NED) score of 0.066 on OCR tasks, placing it in the high-precision realm, compared to the baseline autoregressive model's score of 0.093. - DODO uses a discrete diffusion approach, which is typically associated with image generation, but applies it to OCR by breaking the document into blocks and processing them simultaneously. - The inherent rigidity of OCR tasks (requiring strict sequential processing for elements like invoice numbers) makes them historically poor fits for general diffusion models, which DODO overcomes. - The technique allows for caching of processed blocks, preventing recomputation of the history and enabling faster throughput, which is crucial for enterprise-scale operations. - The paper frames the DODO approach as moving AI processing from an era of 'Typing AI' to 'Stamping AI' for document understanding. ![Screenshot at 00:04: The video introduces the core concept of the paper, which is analyzing the trade-off between speed and accuracy in large-scale document processing, highlighting the challenge of sequential processing versus parallel computation.](https://ss.rapidrecap.app/screens/Ds8xtE_xBvY/00-00-04.jpg) **Context:** This video discusses a research paper titled 'DODO: Discrete OCR Diffusion Models,' a joint effort between the Technion Israel Institute of Technology and Amazon Web Services. The paper introduces a new architecture designed to improve the speed and accuracy of Optical Character Recognition (OCR) on large documents, specifically addressing the inherent sequential processing limitations found in traditional autoregressive models. ## Detailed Analysis The paper analyzed, titled 'DODO: Discrete OCR Diffusion Models,' presents a joint research effort from Technion and AWS that tackles the fundamental trade-off between speed and accuracy in handling large documents like medical journals or financial reports. The core problem addressed is that traditional OCR models, often autoregressive, process text sequentially (token by token or word by word), leading to high latency. The DODO architecture proposes a solution by adopting a block-discrete diffusion approach, similar to image generation models but adapted for OCR. This allows the model to process large chunks of the document in parallel, leading to a significant speed increase. The researchers found that DODO achieves a 3x speed advantage over the baseline autoregressive model while maintaining high accuracy. Specifically, DODO achieved an NED score of 0.066, while the baseline autoregressive model scored 0.093. The key innovation is that DODO processes blocks in parallel, verifying the output of previous blocks to maintain context, which is crucial for structural integrity (like maintaining the correct order of invoice numbers). In contrast, standard diffusion models often fail rigid tasks because they process the entire sequence simultaneously, leading to errors or 'hallucinated padding.' DODO mitigates this by using a cache to store the results of already processed blocks, preventing redundant recomputation and maintaining context across the entire document, thus solving the problem of fractured output and ensuring overall coherence. ### Paper Identification - DODO: Discrete OCR Diffusion Models - Joint research from Technion and AWS - Addresses speed vs. accuracy in large-scale document processing ### Core Problem & Solution - Sequential bottleneck in autoregressive models - DODO uses block-discrete diffusion for parallel processing - Achieves 3x speedup over baseline autoregressive models ### Performance Metrics - DODO NED score of 0.066 - Baseline autoregressive NED score of 0.093 - DODO achieves 66 tokens/second cache rate vs. 21 tokens/second for baseline ### Architectural Advantage - DODO allows parallel processing within blocks and sequential processing between blocks - Maintains high accuracy by caching previous block representations - Avoids issues like hallucinated padding or incorrect ordering common in standard diffusion models ### Impact and Future - Paper frames this as a shift from 'Typing AI' to 'Stamping AI' - Shows that diffusion models can handle rigid tasks like OCR efficiently - The approach suggests high speed and accuracy are achievable for enterprise document operations ![Screenshot at 00:00: The initial screen displays the podcast branding and a call to action: "Become A Member Today!"](https://ss.rapidrecap.app/screens/Ds8xtE_xBvY/00-00-00.jpg) ![Screenshot at 00:17: The speakers introduce the paper being analyzed: 'DODO: Discrete OCR Diffusion Models,' noting it addresses speed vs. accuracy in document processing.](https://ss.rapidrecap.app/screens/Ds8xtE_xBvY/00-00-17.jpg) ![Screenshot at 01:05: The speaker names the core technique being examined: 'Autoregressive Decoding' and contrasts it with the new approach.](https://ss.rapidrecap.app/screens/Ds8xtE_xBvY/00-01-05.jpg) ![Screenshot at 02:26: A comparison is made between DODO and Masked Diffusion Models, suggesting DODO is the alternative approach.](https://ss.rapidrecap.app/screens/Ds8xtE_xBvY/00-02-26.jpg) ![Screenshot at 09:57: A slide or graphic illustrating the two main failure modes identified in the paper: Length Mismatch and Positional Anchoring.](https://ss.rapidrecap.app/screens/Ds8xtE_xBvY/00-09-57.jpg)