Abstract
Many photo archives, e.g., personal photo libraries, suffer from poor quality due to factors such as capture with older, low-resolution cameras, motion blur, etc. Such images, while possibly holding significant sentimental value, are visually unsatisfying and unsuitable for printing or for high-resolution display. This disclosure describes techniques that generatively fill in contextually accurate detail in a low-quality image. Multimodal analysis of the image is performed to detect key regions of interest (faces, objects, etc.) that are low quality (blurred, pixelated, etc.). These regions are used to retrieve relevant, high-fidelity reference images from personal and global data corpora. The resulting multimodal data, including the original image, textual descriptions, and visual references, etc., are used to condition a generative model. The generative model synthesizes an updated image with new, contextually accurate detail.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Labzovsky, Ilia and Karmon, Danny, "Reference-guided Generative Augmentation of Low-Quality Images", Technical Disclosure Commons, (January 11, 2026)
https://www.tdcommons.org/dpubs_series/9171