A video matting extension that matches high-resolution quality at an estimated 7.4x speedup, cutting peak GPU memory to 7.6 percent at 4K
AspireAI community · arXiv (Shuyun Wang, Shangchen Zhou, Peiqing Yang et al.)
High-resolution video matting requires preserving fine details such as hair strands and soft transparency while maintaining consistency across frames. Memory-based methods get their quality by matching current-frame features against information stored from earlier frames, and that matching becomes increasingly costly as resolution grows. MatAnyone-E, an efficient extension of MatAnyone 2, concentrates computation where it is most needed: it first performs low-resolution propagation to identify uncertain regions and the memory candidates most relevant to them, then performs high-resolution matching only for those regions using the selected memory while reusing low-resolution matching results elsewhere. A lightweight decoder designed around measured runtime bottlenecks further reduces the cost of generating high-resolution mattes. The result maintains matting quality comparable to MatAnyone 2 while achieving an estimated 7.4x inference speedup and reducing peak GPU memory usage to 7.6% of MatAnyone 2 at 4K. With compilation-enabled inference on a single NVIDIA H200 GPU it reaches 73.07 FPS at 1080p and 28.62 FPS at 4K. The limitation: the speedup is described by the authors as estimated rather than a measured end-to-end wall-clock figure, and the headline quality claim is comparability with the original rather than an improvement on it.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.