Yuval Domb

Dor Shmilovich, Tony Wu, Aviad Dahan et al.

Diffusion Transformers, particularly for video generation, achieve remarkable quality but suffer from quadratic attention complexity, leading to prohibitive latency. Existing acceleration methods face a fundamental trade-off: dynamically estimating sparse attention patterns at each denoising step incurs high computational overhead and estimation errors, while static sparsity patterns remain fixed and often suboptimal throughout denoising. We identify a key structural property of diffusion attention, namely, its sparsity patterns exhibit strong temporal coherence across denoising steps. Tiles deemed non-essential at step $t$ typically remain so at step $t+δ$. Leveraging this observation, we introduce LiteAttention, a method that exploits temporal coherence to enable evolutionary computation skips across the denoising sequence. By marking non-essential tiles early and propagating skip decisions forward, LiteAttention eliminates redundant attention computations without repeated profiling overheads, combining the adaptivity of dynamic methods with the efficiency of static ones. We implement a highly optimized LiteAttention kernel on top of FlashAttention and demonstrate substantial speedups on production video diffusion models, with no degradation in quality. The code and implementation details will be publicly released.

Yuval Domb

We develop a geometric and information-theoretic framework for encoder-decoder learning built on the Information Bottleneck (IB) principle. Recasting IB as a rate-distortion problem with Kullback-Leibler (KL) divergence as distortion, we show that the optimal representation at any distortion level is a soft clustering of the \emph{predictive manifold} $\mathcal{M}=\{p(Y|x):x\in\mathcal{X}\}$ inside the probability simplex, admitting a linear decoder in the canonical parameterization. We derive a chain of exact transformations, from flat Dirichlet to exponential to isotropic Gaussian, connecting the maximum entropy prior on the simplex to Euclidean space, with quantified entropy overhead at each step, and show that Sketched Isotropic Gaussian Regularization (SIGReg) implements a Gaussian relaxation of this principle whose overhead affects rate accounting but not achievable prediction. This relaxation provides a principled distributional regularizer for learning with limited or no supervision. Using the Conditional Entropy Bottleneck (CEB) decomposition, we derive concrete encoder losses for supervised and semi-supervised settings, estimated via minibatch marginals without variational bounds. In the self-supervised setting, the CEB conditional rate is replaced by a view-prediction proxy. SIGReg serves as the distributional regularizer for both the semi-supervised and self-supervised settings. Experiments on toy problems and FashionMNIST confirm the predicted rate-distortion trade-offs and show that the non-parametric estimator is competitive with the standard variational approach.