One-step Image-function Generation via Consistency Training

Ke Liu, Feng Liu, Jingjun Gu, Shuyi Zhang, wangzhihua, Jiajun Bu, Bo Han, Haishuai Wang

Keywords: Image generation; Diffusion models; Consistency Models

TL;DR: We use Transformer-based generator to generate image functions via consistency training, and the image functions can be rendered as any-resolution images.

Abstract: Consistency models aim to deliver a U-Net generator to map noise to images directly and enable swift inference with minimal steps, even trained in isolation with consistency training mode. However, the U-Net generator requires heavy feature extraction layers for multi-level resolutions and learning convolution kernels with specific receptive fields, resulting in the challenge that consistency models suffer from heavy training resources and fail to generate images with any user-specific resolutions. In this paper, we first validate that training the original consistency model with a small batch size via consistency training mode is pretty unstable, which motivates us to investigate efficient and flexible consistency models. To this end, we propose to use a novel Transformer-based generator to generate continuous image functions, which can then be differentially rendered as images with arbitrary resolutions. We adopt implicit neural representations (INRs) to form such continuous functions, which help to decouple the resolution of generated images and the total amount of the parameters generated from the neural network. Extensive experiments on one-step image generation demonstrate that our method greatly improves the performance of consistency models with low training resources and also provides an efficient any-resolution image sampling process.

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: 9675