One-Step Image-Function Generation via Consistency Training

View at publisher Summarise with VIRAT
Authors: Kai Liu, Feng Liu, Jingjun Gu, Shimin Zhang, Zhihua Wang, Jiajun Bu, Bo Han, Haishuai Wang
Cited by: 0
References: 62

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Diffusion models achieve remarkable image generation quality but suffer from slow inference due to hundreds of denoising steps. Consistency models alleviate this and enable one-step sampling by directly mapping noise to images with a U-Net generator, but they remain unstable and resource-hungry under small-batch training and can only generate images at fixed resolutions. These limitations motivate us to explore more efficient and flexible architectures. To this end, we propose a novel framework for efficient one-step flexible-resolution image generation. First, we represent images as continuous image functions based on implicit neural representations (INRs), which decouple image resolution from the total amount of the generated parameters and can be differentiably rendered at arbitrary resolutions. Then, we design a Transformer-based generator that directly outputs such continuous functions, providing an end-to-end architecture tailored for consistency training and stabilizing training under limited resources. To further improve convergence, we introduce an image reconstruction warm-up task that strengthens the function prediction module. Extensive experiments on one-step image generation demonstrate that our method significantly improves the stability and performance of consistency models under low-resource settings, while also enabling efficient any-resolution sampling.

Fields of Study & Subjects

Generative Adversarial Networks and Image Synthesis 3D Shape Modeling and Analysis Advanced Neural Network Applications

Publication details

Journal / source
IEEE Transactions on Multimedia
Volume
28
Pages
8273–8284
Publication date
1 January 2026
Language
EN