VR-Sampling: Accelerating Flow Generative Model Training with Variance Reduction Sampling
Keywords: Flow Generative Models, Training Acceleration, Diffusion Models
Abstract: Recent advancements in text-to-image and text-to-video models, such as Stable Diffusion 3 (SD3), Flux and OpenSora, have adopted rectified flow over traditional diffusion models to enhance training and inference efficiency. SD3 notes increased difficulty in learning at intermediate timesteps but does not clarify the underlying cause. In this paper, we theoretically identify the root cause as a higher variance in the loss gradient estimates at these timesteps, which hinders training efficiency. Furthermore, this high-variance region is significantly influenced by the noise schedulers (i.e., how we add noises to clean images) and data (or latent space) dimensions. Building on this theoretical insight, we propose a Variance-Reduction Sampling (VR-sampling) strategy that samples the timesteps in high-variance region more frequently to enhance training efficiency in flow models. VR-sampling constructs sampling distributions based on Monte Carlo estimates of the loss gradient variance, allowing it to easily extend to different noise schedulers and data dimensions. Experiments demonstrate that VR sampling accelerates training by up to 33\% on ImageNet 256 and 50\% on ImageNet 512 datasets in rectified flow models. Furthermore, VR-sampling could simplify the hyperparameter tuning of logit-normal sampling introduced in SD3.
The code is available anonymously in~\url{https://github.com/AnonymousProjects/VR_sampling.git}.
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: 1096
Loading