InteractiveAvatar: Real-Time Streaming Video Generation for Consistent and Intent-Aware Avatars

In plain English

A streaming avatar video generator built on a distilled diffusion transformer that keeps a two-tier visual memory — a short FIFO window of recent frames plus a fixed-size buffer of semantically distinct key frames — so appearance does not drift over a long session, and routes user input through an LLM that emits a spoken reply plus a separate action prompt. It reports 26.68 FPS against 21.94 for the next-fastest baseline and the best scores on its own object and temporal consistency metrics. Practitioners should note what is missing: no end-to-end latency figure, no lead on image-quality metrics such as FID or FVD, and an evaluation designed by the authors.