SADE: a Scene-text Autoregressive Diffusion Engine for Character Sequence Recognition

Cameron Painting, Willie Brink

Keywords: optical character recognition, diffusion models, autoregressive image generation

TL;DR: We propose SADE: a scene-text autoregressive diffusion engine for the generation of images containing short alphanumeric text sequences that may assist OCR model training in data-scare scenarios.

Abstract: We consider the problem of training an optical character recognition (OCR) model to read short alphanumeric scene-text sequences, such as number plates or vehicle type labels, in scenarios where labelled training images are limited in quantity and sequence variety. OCR models may under-perform in these scenarios, so we explore whether a diffusion model can be trained on the small set of labelled images, to generate synthetic images with similar background statistics but new character sequences. We find that a diffusion model struggles to generate characters in positions of the sequence where they did not appear during training. We address this problem by introducing SADE: a scene-text autoregressive diffusion engine that generates multiple image parts one by one, conditioned on previously generated image parts for visual coherency. This approach reduces the effective number of possible positions for a character, and increases the diffusion model's ability to generate characters in novel positions of the full sequence. Our results indicate that SADE can indeed lead to substantial improvements in OCR accuracy in data-scare scenarios, particularly on sequences with characters at positions not encountered in the original training data.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10182