Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Diffuse and Disperse: Image Generation with Representation Regularization

Created by
  • Haebom

Author

Runqian Wang, Kaiming He

Outline

In this paper, we propose a new regularization technique, 'Dispersive Loss', to improve the performance of diffusion-based generative models. Existing diffusion models rely on regression-based objective functions and lack explicit regularization. Dispersive Loss solves this problem by inducing the dispersion of internal representations in the hidden space. It is similar to contrastive learning, but it does not require positive sample pairs, and therefore does not interfere with the sampling process used in regression. Compared to the existing representation alignment (REPA) technique, it has the advantage of not requiring pre-training, additional parameters, or external data, and it outperforms the existing best-performing models through experiments on various models on the ImageNet dataset. It is expected to contribute to bridging the gap between generative modeling and representation learning.

Takeaways, Limitations

•
Takeaways:
◦
We present Dispersive Loss, a simple and effective regularization technique that improves the performance of diffusion-based generative models.
◦
Unlike existing methods, it achieves performance improvement without pre-training, additional parameters, or external data.
◦
Suggesting a potential contribution to the convergence between generative modeling and representation learning.
◦
The superiority of Dispersive Loss is verified through ImageNet experimental results.
•
Limitations:
◦
The effect of Dispersive Loss may be limited to the ImageNet dataset. Further research is needed on generalization performance on other datasets.
◦
There is a lack of theoretical analysis on the operating principle of Dispersive Loss. A more in-depth analysis is needed to clearly identify the mechanism of performance improvement.
◦
Extensive experiments on different diffusion model architectures are still lacking. Performance and generalization performance on different architectures should be verified.
👍
Made with Slashpage