Scaling Diffusion Models for Downstream Prediction

Rahul Ravishankar, Zeeshan Patel, Jathushan Rajasegaran, Jitendra MalikSubmission to ICLR 2025 · Withdrawn

Its authors withdrew this paper, so there is no full text to read or summarize.

Abstract

In this paper, we argue that iterative computation, as exemplified by diffusion models, offers a powerful paradigm for not only image generation but also for visual perception tasks. First, we unify few of the mid-level vision tasks as image to image translations tasks ranging from depth estimation to optical flow to segmentation. Then, through extensive experiments across these tasks, we demonstrate how diffusion models scale with increased compute during both training and inference. Notably, we train various dense and Mixture of Expert models up to 2.8 billion parameters, and we utilize increased sampling steps, use various ensembling methods to increase compute at test time. Our work provides compelling evidence for the benefits of scaling compute at train and test time for diffusion models for visual perception, and by studying the scaling properties carefully, we were able to archive same performance of the state-of-the-art models with less compute.
BibTeX
@inproceedings{ravishankar2025scaling,
  author = {Rahul Ravishankar and Zeeshan Patel and Jathushan Rajasegaran and Jitendra Malik},
  title = {{Scaling Diffusion Models for Downstream Prediction}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.01208},
}
Figure