Researchers have introduced Avatar-Forever, a novel framework designed for high-quality, real-time infinite interactive avatars. This system decouples the training of generation efficiency and long-horizon robustness into parallel branches, addressing limitations in existing sequential training pipelines. Avatar-Forever utilizes a full-parameter distillation for visual quality and a lightweight adapter trained with Recovery-oriented Rollout Training (RRT) for improved generation robustness. Additionally, it incorporates ForeverCache, a chunk-wise feature caching mechanism to optimize streaming inference by reducing redundant computations. Built on a 22B video foundation model, Avatar-Forever achieves unbounded audio-driven avatar generation with consistent identity and motion, reaching a throughput of 27.2 FPS at 768x512 resolution on a single H100 GPU. AI
IMPACT This framework could enable more stable and realistic digital humans for interactive applications.
RANK_REASON The cluster contains a research paper detailing a new technical framework for avatar generation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
