LPM 1.0

LPM 1.0 stands for Large Performance Model—a 17B-parameter Diffusion Transformer for real-time AI video character performance generation. It creates characters that speak, listen, sing, and react with human-like nuance for conversational agents, game NPCs, and streaming applications.

LPM 1.0 uses a co-designed data pipeline, 17B Diffusion Transformer architecture, and DMD distillation into an online causal streaming generator. This compresses diffusion into 2 steps, enabling 0.35s latency while generating 480P/720P video at 24fps with identity consistency.

Full-duplex conversation means characters generate video while simultaneously listening and reacting to user input. LPM 1.0 creates listening behaviors—nods, gaze shifts, micro-expressions—during user speech, enabling natural dialogue flow without artificial turn-taking pauses.

LPM 1.0 achieves 0.35s latency vs LiveAvatar's 1+ seconds, supports full-duplex conversation (others don't), maintains infinite-length generation vs LiveAvatar's ~2 minute limit, and offers zero-shot generalization across character styles without fine-tuning.

LPM 1.0 supports photorealistic humans, 2D anime, 3D game characters, and non-humanoid creatures through zero-shot generalization. The model accepts any character style as input and generates expressive performance video without domain-specific training.

LPM 1.0 maintains identity-consistent character video generation for 45+ minutes without drift. Its multi-granularity identity conditioning system preserves appearance details across extended sessions, enabling long-form applications.

LPM 1.0 appears to be available through a commercial platform with pricing tiers and API access. The website offers showcases, technical documentation, and pricing plans but doesn't indicate open-source availability.

LPM 1.0 was developed by 20+ researchers at Anuttacon, an AI research organization. The model is trending on Hugging Face Papers and represents state-of-the-art advancement in real-time character performance generation.

Conversational AI agents, game NPCs, live streaming characters, educational tools, and accessibility solutions all benefit from LPM 1.0's real-time, identity-consistent performance generation with full-duplex conversation capabilities.

Multi-granularity identity conditioning, 17B Diffusion Transformer architecture, DMD distillation for real-time generation, multimodal control unification, and full-duplex conversation support collectively enable LPM 1.0's breakthrough capabilities.