LiveFoley generates real-time video-to-audio with 156ms latency on one GPU
victor_explore · x · 2026-11-08
LiveFoley (arXiv 2611.05988) 是一种流式视频转音频 (V2A) 生成模型,面向游戏、AR/VR 和实时视频等沉浸式交互环境,在单个 GPU 上实现了 156ms 的端到端延迟。
该论文解决了先前流式 V2A 方法的两个瓶颈:
- 视觉前端:现有的截断感受野(truncated receptive fields)会破坏视听对齐。LiveFoley 引入了一种尾随滑动窗口(Trailing Sliding Window),无需重新训练即可将非因果片段级编码器转换为流式模块,并通过可配置的跳数(hop size)将前瞻延迟保持在人类感知阈值以下。
- 音频后端:它使用带有更宽去噪窗口的局部扩散 transformer(LocDiT)代替了逐块扩散去噪——历史 LocDiT 添加了过去上下文而无需额外延迟,并且还提出了一个 Future-Warmup 变体,以及一个支持不同延迟/计算预算的 Unified Varhop Framework。
论文、项目页面和演示均可用。
More from Multimodal
- RelayVSR, a ComfyUI single-file video super-resolution model, trends on Hugging Face — Comfy-Org · 2026-11-08
- Midjourney style code --sref 787130812 produces painterly surfing artwork — sergeantsref · 2026-11-08
- cinecode: open-source tool turns Claude Code into a free film studio that renders full MP4s from one sentence — solyarisoftware · 2026-11-08
- 131 agents, 23,354 model calls, 26 hours: creator builds most complex AI music video yet — tetsuoai · 2026-11-08
- AI 'Rural Modular' Video: Three Elderly Women Patching Doepfer Synths in a Village — love1008 · 2026-11-08
- AI-generated Lost Star Wars Audition Tapes parody nails 80s nostalgia — TinfoilTricorn · 2026-11-08