SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI videoLPM 1.0

LPM 1.0

LPM 1.0 (Large Performance Model) is a real-time AI video character performance generation model developed by the Anuttacon team. It adopts a 17B-parameter Diffusion Transformer architecture and can transform a single image into a virtual character that can speak, listen, sing, and react in real time, with an end-to-end latency of only 0.35 seconds. The model supports full-duplex conversational video generation, and while speaking, the character can generate listening response behaviors such as nodding, gaze shifts, and micro-expressions. Through multi-granularity identity conditioning, LPM 1.0 can maintain character identity consistency for more than 10 minutes of continuous generation without drift. It supports unified control across three modalities: text, audio, and images, and has zero-shot character generalization capability, able to handle any style including photorealistic humans, 2D anime, 3D game characters, and non-human creatures. The output specifications are 480P/720P resolution, 24fps real-time video. LPM 1.0 uses DMD distillation technology to compress the diffusion process to 2 steps, achieving online causal streaming inference. The model is suitable for scenarios such as conversational AI agents, game NPCs, virtual livestreaming characters, and accessible education. Users can use it directly online without installing a GPU or Python environment, and paid plans start at $9.9 per month.

Open website

About this tool

Purpose and Positioning

LPM 1.0 (Large Performance Model) is a large model for video character performance, built by more than 20 researchers at Anuttacon. Its core goal is to provide a visual engine for conversational agents, virtual livestreaming characters, and game NPCs, turning a single image into a virtual character that can speak, listen, sing, and react in real time. The official website points out that current virtual avatar systems generally suffer from problems such as animation loops, frozen expressions, and an inability to truly listen or react, and LPM 1.0 aims to change this situation.

Main Capabilities

Full-Duplex Conversational Video Generation

LPM 1.0 is the first model to achieve full-duplex conversational video generation. While speaking, the character has precise lip synchronization and body rhythm, and while the user is speaking, it can generate reactive listening behaviors, including nodding, gaze shifts, and micro-expressions.

Real-Time Streaming Generation

LPM 1.0 achieves 0.35-second end-to-end latency through DMD (Distribution Matching Distillation) distillation. The online LPM causal streaming generator compresses the diffusion process to only 2 generation steps, achieving real-time output while maintaining the full quality of the 17B model.

Identity Consistency

LPM 1.0 uses multi-granularity identity conditioning, including global appearance references, multi-view body images, and facial expression samples. The official website states that it can maintain identity consistency for more than 10 minutes of continuous generation and eliminate hallucinated details such as teeth, expression wrinkles, and profile geometry.

Multimodal Control

LPM 1.0 unifies three natural control signals—text, audio, and images—within a single generation process: text prompts guide motion and behavior, audio drives lip synchronization and emotion, and reference images define character identity.

Zero-Shot Character Generalization

LPM 1.0 accepts any character style as input, including photorealistic humans, 2D anime, 3D game characters, and non-human creatures, and can generate expressive video performances without fine-tuning or domain-specific training.

Output Specifications

The LPM 1.0 online streaming generator produces 480P/720P resolution, 24fps video in real time.

Technical Pipeline

Multimodal Dataset Construction

LPM 1.0 is based on a multimodal human-centered dataset, constructed through strict filtering, speech-listening audio-video pairing, performance understanding, and identity-aware multi-reference extraction.

Base LPM Training

The base LPM is a 17B-parameter Diffusion Transformer, trained with multimodal conditional control, simultaneously processes character images (including identity-aware references), audio signals, and text prompts.

DMD Distillation and Online Inference

The base LPM is distilled via DMD into an online LPM causal streaming generator. At inference time, LPM 1.0 generates character videos in three dialogue states: listening (producing reactive nodding and gaze shifts based on user audio), speaking (producing lip-synced performance based on synthesized audio), and silence (producing natural idle behavior based on text conditions).

How to Use

Users can use LPM 1.0 directly online, without GPU, Python, or animation pipeline setup, to preview character performance videos in a browser. The official website provides curated demos, and users can compare pricing before generating their own videos. Paid plans start at $9.9 per month.

Use Cases

  • Conversational AI agents: Transform text-based chatbots into visual conversational agents with human-like character performances. The official website states that LPM 1.0 is plug-and-play compatible with A2A models such as ChatGPT and Doubao.
  • Game NPCs: Game developers can create NPCs with real-time, identity-consistent performances, replacing preset animations.
  • Virtual livestreaming characters: Supports real-time livestreaming characters with full-duplex conversation, singing, and emotional expression.
  • Accessibility and education: Generate expressive virtual tutors and companion characters that can naturally respond to user interaction.

Limitations and Comparisons

According to information from the official website, comparisons between LPM 1.0 and similar systems include:

  • Compared with LiveAvatar: LPM 1.0 has a latency of 0.35 seconds, while LiveAvatar exceeds 1 second; LPM 1.0 supports full-duplex conversation while LiveAvatar does not; LPM 1.0 supports unlimited-length generation, while LiveAvatar is limited to about 2 minutes.
  • Compared with OmniHuman: OmniHuman runs offline and has a fixed output length, while LPM 1.0 generates in real time and supports unlimited length and full-duplex conversation.
  • Compared with Kling-Avatar-2: Kling-Avatar-2 has a latency of about 0.8 seconds, no full-duplex support, and a maximum of 5 minutes; LPM 1.0 has a latency of 0.35 seconds and supports unlimited-length generation.

The LPM-Bench benchmark mentioned on the official website shows that LPM 1.0 achieves state-of-the-art performance across all evaluation dimensions. Specific complete pricing plans, free quotas, supported input format details, and other information are not yet available.

Related tools

Browse category