LongCat-Video-Avatar 1.5

Overview

LongCat-Video-Avatar 1.5 is a production-grade open-source audio-driven digital human model, released by Meituan's LongCat team in May 2026. It upgrades the original LongCat-Video-Avatar by replacing the Wav2Vec2 audio encoder with Whisper Large for finer phoneme, rhythm, and multilingual prosody capture, and by implementing GRPO (Group Relative Policy Optimization) alignment for human-preference training and DMD step distillation to reduce inference from 50 to 8 NFE, achieving approximately 15x speedup. The publisher reports that Video-Avatar 1.5 generates a 10-second video in approximately one minute and achieves competitive or leading results against HeyGen, OmniHuman 1.5, and Kling Avatar 2.0 on the EvalTalker benchmark (65.9%, 61.1%, and 54.3% win rates respectively). It supports open-domain scenes including real humans, anime characters, virtual idols, and animals, with multi-person interaction enabling natural speaker-listener role distinction.

Architecture

The model uses a DiT (Diffusion Transformer) backbone shared with LongCat-Video, but retrained and adapted for the avatar-specific task. The audio processing pipeline was upgraded from Wav2Vec2 to Whisper Large to provide more detailed phonetic, intonational, and rhythmic features as input conditioning. A LoRA adapter layer enables efficient task-switching between single-person and multi-person modes while sharing the same base model — avoiding the need for separate model deployments. For long-video stability, a Cross-Chunk Latent Stitching mechanism maintains temporal consistency across extended sequences by blending latent representations at chunk boundaries. Frame-level GRPO applies human preference alignment at fine granularity, specifically targeting historically difficult cases: hand deformation, structural collapse in complex poses, and expression-speech mismatch.

Training

Training uses a multi-stage data pipeline combining offline annotation, online quality validation, and three targeted datasets: multi-person (active speaker detection), silent (natural micro-expressions without spurious lip motion), and emotion (frame-level emotional transitions). GRPO alignment applies frame-level reward signals instead of video-level, allowing precise optimization of specific visual quality dimensions. The DMD distillation process compresses the 50-step diffusion process to 8 NFE while maintaining visual fidelity — crucial for production deployment where inference time directly impacts user experience.

Deployment

Model weights are available on HuggingFace (meituan-longcat/LongCat-Video-Avatar-1.5), GitHub (meituan-longcat/LongCat-Video), and ModelScope. Deployment requires a GPU with sufficient VRAM for the DiT backbone plus the Whisper Large audio encoder. The shared base model + LoRA adapter architecture enables a single deployment to serve both single-person and multi-person modes. A project page with documentation and a technical report (arXiv:2605.26486) provide detailed configuration guidance.

Capabilities

Described in the cited primary documentation. Not an independent performance evaluation.

• Audio-driven digital human with accurate lip-synchronization.
• Whisper Large audio encoder.
• Full-body temporal stability and identity consistency.
• 8 NFE step distillation for production inference.
• Generalizes to anime and animals via RLHF training.
• Multi-person interaction and object handling support.

Access & License

Open source. Verify repository license before commercial use.

Sources

• LongCat-Video-Avatar 1.5 (arXiv:2605.26486)

  Primary source · Published 2026-05-26 · Accessed 2026-07-27

  Whisper Large audio encoder, RLHF, 8 NFE distillation. 500+ case benchmark: competitive with HeyGen, OmniHuman 1.5, Kling Avatar 2.0.

• LongCat-Video GitHub (includes Avatar)

  Primary source · Accessed 2026-07-27

  Built on LongCat-Video DiT foundation.

Related comparisons

• LongCat-Video-Avatar 1.5 and Kling Avatar 2.0: Open-Weight vs Closed Digital Human Compared

  A source-based comparison of LongCat-Video-Avatar 1.5 (open-source audio-driven digital human, Whisper Large, 8 NFE, self-hostable) and Kling Avatar 2.0 (Kuaishou's closed digital human, 5-minute clips, 1080p/48fps, platform-only). Covers architecture, duration, resolution, pricing, licensing, and the EvalTalker benchmark.

• LongCat-Video-Avatar 1.5 and HeyGen: Open-Weight vs Platform Avatar Compared

  A source-based comparison of LongCat-Video-Avatar 1.5 (open-source audio-driven digital human, Whisper Large, 8 NFE, self-hostable) and HeyGen (commercial avatar platform, 175+ languages, 15-second avatar creation, video translation). Covers capabilities, languages, pricing, licensing, and the EvalTalker benchmark.

• LongCat-Video-Avatar 1.5 and OmniHuman 1.5: Open-Weight vs Closed Cognitive Avatar Compared

  A source-based comparison of LongCat-Video-Avatar 1.5 (open-source audio-driven digital human, Whisper Large, 8 NFE, self-hostable) and OmniHuman 1.5 (ByteDance's cognitive avatar, System 1+2 dual architecture, up to 60s output). Covers architecture, duration, multi-person, pricing, licensing, and the EvalTalker benchmark.

FAQ

Where can the primary source for LongCat-Video-Avatar 1.5 be found?

The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Video-Avatar 1.5.

Is LongCat-Video-Avatar 1.5 available through the LongCat API?

LongCat-Video-Avatar 1.5 is documented in the meituan-longcat GitHub organization. Refer to the publisher API documentation at longcat.chat for current API availability.

Is this page an official LongCat or Meituan page?

No. This page is published by an independent third-party site. It is not affiliated with, endorsed by, or sponsored by LongCat or Meituan.

How often is this page updated?

This page was last verified on 2026-07-27. Content is reviewed when new publisher documentation or model releases become available.