Anuttacon · 2026-04-09 · major
LPM 1.0 — Large Performance Model for video-based character performance
A 17B Diffusion Transformer that turns a single character image plus live audio into a real-time conversational video of that character listening, speaking, reacting and emoting — with identity-stable, infinite-length streaming output.
A 17B video model that makes a still character image act out a full conversation in real time — listening, speaking, reacting, in character.
Key specs
| Base params | 17B |
|---|---|
| Ar xiv | 2604.07823 |
What is it?
LPM 1.0 is the first release from Anuttacon, the AI lab founded by miHoYo co-founder Cai Haoyu. Given a character image plus audio, it generates a full-duplex conversational video of that character — not just a talking head, but listening behavior, micro-expressions, turn-taking, and motion — while staying identity-consistent over arbitrarily long interactions.
How does it work?
The authors frame the problem as a 'performance trilemma' between expressiveness, real-time inference, and long-horizon identity stability. A 17B-parameter Diffusion Transformer (Base LPM) is trained on a filtered multimodal dataset of paired speaking–listening clips with identity-aware reference extraction, then distilled into a causal streaming generator (Online LPM) for low-latency, infinite-length playback. Motion is steered by text prompts alongside the audio.
Why does it matter?
Every other avatar model either lip-syncs well on short clips or runs in real time at the cost of identity drift. LPM 1.0 is the first to argue — and benchmark, via the new LPM-Bench — that you can get all three at once with enough systems-level co-design. If the results hold, it becomes the visual engine under conversational agents, live streaming avatars, and game NPCs.
Who is it for?
Video generation researchers, game/NPC devs, avatar and live-streaming platforms.