How does SAKURA keep image editing private and fast on phones?

Diffusion transformers power many image-editing services today. To keep personal photos off remote servers, current systems split the work: the cloud holds reusable template data, the phone runs the request-specific part.

That split creates a bottleneck. The reusable key-value state grows with every template — face swap, virtual try-on, style transfer — and quickly exceeds what a phone can store. Pushing the cache to the cloud solves capacity but puts every transfer on the critical path. On a fluctuating mobile link the download can take longer than the computation itself.

SAKURA attacks the problem from three angles. First, a fine-grained token-importance-aware compressor shrinks the cache before it leaves the server. Second, the runtime picks a representation for each upcoming shard based on live bandwidth measurements. Third, a pipelined scheduler overlaps transfer and reconstruction with on-device execution so nothing sits idle.

The compressor evaluates every token instead of treating the whole cache uniformly. Tokens that matter more for visual fidelity keep more bits. Background or redundant tokens get aggressive quantization. The paper reports a 3.06× reduction versus uncompressed cache while preserving image quality.

At runtime the system monitors the wireless link. It decides shard-by-shard whether to send a compact representation, a medium one, or the full precision version. This adaptive selection means a sudden dip in throughput does not stall the pipeline — the next shard simply ships lighter. The pipeline runs three stages in parallel: the network pulls the next shard, the decompressor rebuilds it in memory, and the diffusion model consumes the previous shard. Overlap hides transfer latency behind compute.

Latency drops 1.82× versus CacheGen adapted to DiT and 3.27× versus pure on-device generation.

Three editing tasks and two diffusion models were used in the evaluation. The device was a single Snapdragon 8 Elite Gen 5 handset. How the gains scale across other chipsets or with larger template libraries is not reported.

The architecture assumes the cloud already holds precomputed KV state for every template. Building and updating that library remains an offline process outside SAKURA's scope. The paper does not discuss energy consumption or thermal impact during sustained sessions.

For a phone owner the practical shift is clear: the same private editing flow stays local, but the template catalogue can grow without filling storage, and the wait after tapping "apply" shrinks even on a shaky connection.