A real-time AI avatar that talks and listens for 45 minutes

Most AI voice companions today still feel emotionally hollow. You are either staring at a pulsating soundwave graphic or watching a stiff digital avatar whose lips flap mechanically over a frozen, glassy-eyed face. Talk to them for more than two minutes, and their facial geometry usually melts into nightmare fuel.

Researchers have just unveiled LPM 1.0, a real-time conversational avatar model designed to fix exactly that. Powered by a single reference photo and an incoming audio stream, it generates an interactive video figure on the fly—and keeps the character's visual identity intact for up to 45 minutes without warping.

Real gestures instead of cheap puppetry

Traditional digital humans are basically just lip-sync filters. LPM 1.0 works differently by recognizing three distinct communicative states: speaking, listening, and idling.

When you talk, the digital character does not freeze. It listens, reacting naturally to incoming voice cues with subtle nods, idle posture shifts, and wandering eye movements. When responding, it doesn't just match syllables. It mimics hesitation, smiles, reaches out to virtually high-five the camera, and even reaches up to adjust its desk microphone.

Under the hood, the system relies on what researchers call multi-granularity identity conditioning. In addition to the base photo, it takes multi-angle and expressive references so the AI never has to hallucinate fine details like teeth, emotional laugh lines, or side-profile angles.

The system works out of the box across realistic portraits, anime characters, and 3D game models without extra fine-tuning. According to project manager Ailing Zeng, it can plug directly into audio models like ChatGPT or Doubao for live video calls, or run offline to turn podcast scripts into multi-person video panels.

Don't expect a download link

If you want to play with it today, you are out of luck. The team has made it clear that LPM 1.0 is strictly an internal research initiative. There are no plans to release the model weights, open-source the code, or launch a public web demo.

The researchers openly note that the output still exhibits visible rendering artifacts, with an objective quality gap compared to authentic camera footage. Crucially, all faces displayed in the demonstrations were entirely synthetic figures rather than real human likenesses.

The primary roadblock is security. A system that can turn any single snapshot into a photorealistic, 45-minute live interactive avatar is essentially off-the-shelf infrastructure for real-time deepfakes. Dropping this into the wild without strict verification would make video-based impersonation and financial fraud trivial. The team noted they will only consider access once robust guardrails and safety frameworks exist.

Why it matters

We think the real breakthrough in LPM 1.0 is not raw video generation length, but how it solves the listening phase of conversation.

Real human dialogue relies heavily on non-verbal cues. When you pause to think or ramble, a slight nod or an eye shift from the listener keeps you engaged. Most avatars fail because their idle state feels dead. By giving an AI character natural micro-reactions while it waits for its turn to speak, LPM 1.0 crosses the uncanny valley in a way traditional video renderers have failed to do.

For game studios building living NPCs and companies designing virtual assistants, this architecture points to a viable future where characters converse with zero pre-rendered motion capture.

Yet the safety dilemma is stark. When a single photo and an audio prompt are all it takes to generate an unblinking, continuous visual lie in real time, the barrier to identity theft collapses. Keeping LPM 1.0 locked in the lab is the only sensible move right now, and the rest of the industry will need to figure out real-time detection before tools like this inevitably slip out.