Multi-Image Zero-Shot Subject Generation for Visual Storytelling
Keywords: Text-to-Image Generation, Personalization
Abstract: In text-to-image generation, the subfield of personalization deals with creating images of specific subjects/concepts in novel compositions. This is done by leveraging a set of images with a reference concept to either learn or guide a personalized generation process.
Such personalization, however, is only applicable when reference data exists - i.e. when the desired concept is present in (preferably numerous) images. This does not address the setting where consistency across generations is desired without constraints pertaining to existing data. In this work, we explore how to achieve consistency without reference images. Conditioning on sets of text prompts, we generate corresponding image sets where the shared concept has persistent appearance across generations. Our contributions are three-fold. First, we identify and define the novel task of Multi-image Zero-shot Subject Generation. Second, we create a benchmark of over 100 caption sets (corresponding to short stories) with repeated subject concepts to use as a testbed. Third, we demonstrate that existing methods cannot perform this task and propose an initial method which achieves uniform subject appearance by iteratively optimizing image sets to have similar visual content while still satisfying the text conditioning.
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/2024/AuthorGuide.
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: 6279
Loading