What is LPM?
Large Performance Model Explained

LPM (Large Performance Model) is a video-based AI model developed by miHoYo (the studio behind Genshin Impact and Honkai: Star Rail) that generates realistic character performance videos. The first version, LPM 1.0, was released as a research paper in April 2026 and represents a major step forward in making virtual characters behave like expressive, socially aware performers.

The Problem LPM Solves: The Performance Trilemma

Existing AI video models struggle to achieve three things simultaneously — a challenge that the LPM researchers call the "performance trilemma":

• High Expressiveness — rich facial expressions, natural gestures, and emotional depth
• Real-Time Inference — fast enough for live interaction
• Long-Horizon Identity Stability — the character stays consistent over extended interactions

Previous models could achieve one or two of these, but not all three at once. LPM 1.0 claims to resolve this trilemma through a systems-level engineering approach that combines a large base model with an optimized streaming pipeline.

LPM 1.0 Architecture

LPM 1.0 consists of two main components working together:

The training pipeline involves building a multimodal human-centric dataset through strict filtering, speaking-listening audio-video pairing, performance understanding annotation, and identity-aware multi-reference extraction.

What Can LPM 1.0 Do?

LPM 1.0 focuses on single-person full-duplex audio-visual conversational performance. In practical terms, this means:

• Speaking videos — given synthesized audio, the character speaks with natural lip sync, facial expressions, and body gestures
• Listening videos — given user audio, the character listens with realistic reactions — nodding, eye contact, subtle expressions
• Text-controlled motion — text prompts can guide the character's gestures and movement style
• Identity-stable generation — the character maintains visual consistency across an entire conversation, theoretically infinite in length
• Real-time speed — Online LPM runs fast enough for interactive applications

Target Applications

The LPM paper describes three primary use cases:

• Conversational AI agents — visual embodiment for chatbots and virtual assistants
• Live streaming characters — AI-driven VTubers or digital humans that interact with audiences in real time
• Game NPCs — non-player characters that respond to players with natural, expressive behavior

LPM-Bench: The Evaluation Benchmark

Alongside the model, the team introduced LPM-Bench, described as the first benchmark specifically designed for evaluating interactive character performance. This benchmark evaluates models on expressiveness, identity consistency, real-time capability, and conversational naturalness — dimensions that existing video generation benchmarks don't adequately cover.

Limitations and Availability

Despite its impressive capabilities on paper, LPM 1.0 has significant limitations for practical use:

• Not publicly available — no model weights, no source code, no API, no demo
• Non-commercial academic use only — even within the research community, access is restricted
• No timeline for release — miHoYo has stated no plans for any form of public access
• Audio-driven only — requires audio input; cannot generate videos from text prompts alone
• Single-person only — does not support multi-character scenes or group conversations

How to Generate Character Performance Videos Today

Since LPM 1.0 is not available to the public, anyone looking to create character performance videos today needs an alternative solution. CPMV AI (Character Performance Model Video) provides an accessible online tool for generating expressive character videos from text prompts, powered by Veo 3.1.

While CPMV takes a different technical approach (text-to-video rather than audio-driven generation), it delivers on the core promise of character performance — creating characters that appear expressive, natural, and alive.