Reproducibility of "ITI-GEN: Inclusive Text-to-Image Generation"
Abstract: A major limitation of current text-to-image generation models is their inherent tendency to incorporate biases, thereby not demonstrating inclusivity in certain attributes. An approach to enhance the inclusiveness is suggested by \cite{zhang2023itigen}: Inclusive Text-to-Image Generation (ITI-GEN). The authors state that ITI-GEN leverages reference images to improve the inclusiveness of text-to-image generation by learning inclusive prompt embeddings for targeted attributes. In this paper, the reproducibility of ITI-GEN is investigated in an attempt to validate the main claims presented by the authors. Moreover, additional experiments are conducted to provide further evidence supporting their assertions and research their limitations. This concerns the research on inclusive prompt embeddings, the inclusivity of untargeted attributes and the influence of the reference images. The results from the reproducibility study mainly show support for their claims. The additional experiments reveal that ITI-GEN only guarantees inclusivity for the specified targeted attributes. To address this shortcoming, we present a possible solution, namely ensuring a balanced reference dataset.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Changyou_Chen1
Submission Number: 2169
Loading