Fitting Image Diffusion Models on Video Datasets
This work addresses the limitation of static image training for generative models in capturing temporal dynamics, offering an incremental improvement for video-based applications.
The paper tackles the problem of training image diffusion models on video datasets by proposing a training strategy that leverages temporal inductive bias, resulting in over 2x faster convergence and lower FID scores on both training and validation distributions.
Image diffusion models are trained on independently sampled static images. While this is the bedrock task protocol in generative modeling, capturing the temporal world through the lens of static snapshots is information-deficient by design. This limitation leads to slower convergence, limited distributional coverage, and reduced generalization. In this work, we propose a simple and effective training strategy that leverages the temporal inductive bias present in continuous video frames to improve diffusion training. Notably, the proposed method requires no architectural modification and can be seamlessly integrated into standard diffusion training pipelines. We evaluate our method on the HandCo dataset, where hand-object interactions exhibit dense temporal coherence and subtle variations in finger articulation often result in semantically distinct motions. Empirically, our method accelerates convergence by over 2$\text{x}$ faster and achieves lower FID on both training and validation distributions. It also improves generative diversity by encouraging the model to capture meaningful temporal variations. We further provide an optimization analysis showing that our regularization reduces the gradient variance, which contributes to faster convergence.