InteractiveAvatar
Real-Time Streaming Video Generation for Consistent and Intent-Aware Avatars (ECCV 2026)
InteractiveAvatar is a real-time streaming video generator for intent-aware, visually consistent digital humans.
The two bottlenecks it tackles:
- Understanding-interaction gap — avatars react to surface cues but miss user intent.
- Long-term semantic drift — quality and identity degrade as streams grow infinitely long.
We design a long-short term token memory that anchors identity, scene, and intent across arbitrary horizons, enabling stable streaming generation.
ECCV 2026.