Amazing Combinatorial Creation: Acceptable Swap-Sampling for Combinatorial Text-to-Image Generation

Jun Li, Ze-dong Zhang, Jian Yang

Keywords: Creative Learning; Diffusion Model; Combinatorial Text-to-Image Synthesis

TL;DR: We develop an acceptable swap-sampling strategy to generate a meaningful object image with novelty and surprise from text concepts of given different objects in combinatorial text-to-image synthesis.

Abstract: Exploring a machine learning system to generate meaningful combinatorial object images from multiple textual descriptions, emulating human creativity, is a significant challenge as humans are able to construct amazing combinatorial objects, but machines strive to emulate data distribution. In this paper, we develop a straightforward yet highly effective technique, called acceptable swap-sampling, to generate a combinatorial object image that exhibits novelty and surprise, utilizing text concepts of different objects. Initially, we propose a swapping mechanism that constructs a novel embedding by exchanging column vectors of two text embeddings for generating a new combinatorial image through a cutting-edge diffusion model. Furthermore, we design an acceptable region by managing suitable CLIP distances between the new image and the original concept generations, increasing the likelihood of accepting the new image with a high-quality combination. This region allows us to efficiently sample a small subset from a new image pool generated by using randomly exchanging column vectors. Lastly, we employ a segmentation method to compare CLIP distances among the segmented components, ultimately selecting the most promising image from the sampled subset. Our experiments focus on text pairs of objects from ImageNet, and our results demonstrate that our approach outperforms recent methods such as Stable-Diffusion2, DALLE2, ERNIE-ViLG2 and Bing in generating novel and surprising object images, even when the associated concepts appear to be implausible, such as lionfish-abacus and kangaroo-pears. Furthermore, during the sampling process, our approach without training and human preference is also comparable to PickScore and HPSv2 trained using human preference datasets.

Primary Area: representation learning for computer vision, audio, language, and other modalities

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: 2396