Avatar-Forever: Real-Time Infinite LTX 2.3-Based Talking Avatars

Overview

Most streaming-video systems train few-step generators through sequential, distillation-centric pipelines, where failures from earlier stages compound and distillation objectives degrade quality over long autoregressive rollouts. Avatar-Forever instead treats generation efficiency and long-horizon robustness as two independent capabilities trained in parallel:

• A full-parameter distillation branch trains an efficient generator with high visual quality
• A lightweight long-horizon adapter is trained with Recovery-oriented Rollout Training (RRT) to stay robust under long-horizon inference

Avatar-Forever decoupled parallel training framework (source: leeruibin/avatarforever)

ForeverCache Streaming

During streaming inference, Avatar-Forever uses ForeverCache, a chunk-wise feature caching mechanism that reduces redundant history computation, keeping memory and compute bounded while generating unbounded-length video. Combined with the LTX 2.3 22B foundation model, it maintains identity consistency, motion coherence, and visual fidelity across long conversations, songs, and interactive sessions.

Availability

The project has released the paper, inference code, and model checkpoints; training code and data are still pending.

• Weights: avatarforever-ltx-2.3-22b.safetensors on Hugging Face
• Text encoder: gated Gemma 3 12B (accept the license before downloading)
• Inference: Python CLI in the GitHub repository — no ComfyUI integration yet

Links

• Project Page
• GitHub Repository
• Hugging Face Model
• Paper (arXiv:2608.12107)