OPEN_SOURCE ↗
YT · YOUTUBE// 7h agoMODEL RELEASE
LPM 1.0 enables real-time character performances
The Large Performance Model (LPM 1.0) is a 17-billion parameter Diffusion Transformer system designed for high-fidelity character lip-sync and reactive dialogue behaviors in real-time applications. It addresses the "performance trilemma" by balancing high expressiveness, real-time inference, and identity stability, allowing digital humans to listen, react, and speak with professional-grade consistency.
// ANALYSIS
LPM 1.0 is a breakthrough for digital humans, moving beyond simple talking heads to "full-duplex" interactive actors that can listen and react in real-time.
- –Addresses the "performance trilemma" by balancing high expressiveness, real-time inference, and long-horizon identity stability.
- –Novel interleaved audio injection processes speaking and listening audio in separate layers to distinguish lip movements from micro-expressions.
- –Base 17B model is distilled into a causal streaming generator ("Online LPM") for low-latency, infinite-horizon interaction.
- –The system supports multimodal control, unifying text prompts, audio signals, and reference images for professional-grade identity preservation.
- –Includes LPM-Bench, a first-of-its-kind benchmark for standardizing the evaluation of interactive character performance quality.
// TAGS
lpm-1.0llmmultimodalvideo-genspeechagentbenchmark
DISCOVERED
7h ago
2026-04-12
PUBLISHED
7h ago
2026-04-12
RELEVANCE
7/ 10
AUTHOR
AI Search