EbbFlow improves image generation by adjusting temporal training samples
EbbFlow: PINN-Inspired Progressive Curriculum for MeanFlow Training
Machine Learning
Summary
Some AI systems generate images by learning how things change over time, but their training mixes different kinds of timing information all at once. The authors of this paper propose a new way called EbbFlow that starts training with one type of timing data and gradually adds others, carefully changing how the timing is sampled. This approach helps the AI learn better, producing clearer images on a test using ImageNet data. EbbFlow shows that how you sample time during training can make a noticeable difference in AI image generation.
What this means in practice
- •For image generation engineers: Use scheduled temporal sampling to improve model training efficiency and output quality in generative image networks.
- •For machine learning pipeline developers: Incorporate progressive sampling schedules to better control training for models predicting temporal dynamics.
Authors
Mengfan Yuan, Xunheng Kuang, Jianbin Chen, Lei He
Abstract
MeanFlow learns interval-average velocity for few-step generation, but its fixed temporal sampler mixes diagonal and interior supervision throughout training. We interpret temporal coordinates as collocation points: diagonal samples provide a stochastic boundary target, whereas finite intervals supply a differential-consistency target. The diagonal fraction controls how often the consistency component appears, while interval length enters its gradient through an explicit factor. Motivated by this distinction, EbbFlow uses Temporal Supervision Allocation (TSA) to schedule the diagonal fraction and interval geometry. It starts with diagonal supervision, then admits interior samples and expands their maximum interval; a left-anchored interval law preserves coverage across time. The per-sample loss, target, backbone, and inference procedure remain unchanged. In matched class-conditional ImageNet $256\times 256$ runs without classifier-free guidance (DiT-B/2, one training seed per method, 400k updates), EbbFlow achieves 50k-sample FID of 39.65 at one network evaluation and 36.02 at two-evaluation ODE sampling, versus 44.82 and 39.86 for MeanFlow. Gradient probes show a mid-training difference in component geometry. Ablations support allocation scheduling and left-anchored intervals, although a fixed full-horizon variant performs better at one evaluation. These results suggest that temporal sampling is a useful training control whose benefits depend on the inference protocol.