Private image editing on a phone got 3.27 times faster by streaming reusable attention state under a bandwidth budget
AspireAI community · arXiv preprint
Diffusion-based image editing on a phone is a privacy problem and a performance problem at once: keeping the request local protects the user's photo and prompt, but the reusable attention state for common templates is too large to store on the device, and fetching it from the cloud can be slower than just generating locally.
SAKURA treats that transfer as a bandwidth problem to be scheduled. It compresses the reusable key-value state with a fine-grained token-importance-aware scheme, picks the cache representation for each upcoming shard based on current network bandwidth, and overlaps cache transfer and reconstruction with on-device execution in a pipelined runtime.
Across three editing tasks and two diffusion models on a Snapdragon 8 Elite Gen 5 phone, it cuts latency by 1.82x against CacheGen adapted to diffusion transformers and by 3.27x against fully on-device generation, while reducing cache storage 3.06x versus an uncompressed cache and preserving image quality. The larger point is architectural: for template-heavy workloads, the cloud can send reusable state while the request itself never leaves the phone. Caveat: three tasks on one phone SKU, and the image-quality claim is the authors' own.
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.