Learning Visual Generative Priors without Text

TL;DR

This paper introduces Lumos, a self-supervised vision-based framework for image-to-image generation that learns visual priors without relying on expensive text-image pairs, outperforming some text-to-image models on certain tasks.

Contribution

Lumos demonstrates a scalable, pure vision-based training method for I2I models that serve as strong visual priors, reducing dependence on text-image data and outperforming T2I models in some tasks.

Findings

Lumos can learn I2I models from in-the-wild images in a self-supervised manner.

I2I models serve as better visual priors than T2I models for certain tasks.

I2I priors outperform T2I priors on text-irrelevant tasks like image-to-3D and image-to-video.

Abstract

Although text-to-image (T2I) models have recently thrived as visual generative priors, their reliance on high-quality text-image pairs makes scaling up expensive. We argue that grasping the cross-modality alignment is not a necessity for a sound visual generative prior, whose focus should be on texture modeling. Such a philosophy inspires us to study image-to-image (I2I) generation, where models can learn from in-the-wild images in a self-supervised manner. We first develop a pure vision-based training framework, Lumos, and confirm the feasibility and the scalability of learning I2I models. We then find that, as an upstream task of T2I, our I2I model serves as a more foundational visual prior and achieves on-par or better performance than existing T2I models using only 1/10 text-image pairs for fine-tuning. We further demonstrate the superiority of I2I priors over T2I priors on some text-irrelevant visual generative tasks, like image-to-3D and image-to-video. Our project page is available at https://ant-research.github.io/lumos.