# LongCat-Video-Avatar 1.6 Setup: The Flags That Matter

The open avatar model can turn speech and a portrait into video, but its example settings and script defaults do not match.

By Paras Madan | 2026-11-16


Meituan released LongCat-Video-Avatar 1.6 on November 15, with downloadable weights and code for audio-driven character video. For developers trying it, the first useful detail is surprisingly easy to miss: downloading the new checkpoint does not automatically select the new model behaviour.

The single-person script still defaults to `avatar-v1.0`. Its mixing defaults also differ from the published 1.6 example. This is a setup guide based on the official documentation and source, not a hands-on quality or speed test.

## What changed in Avatar 1.6

The [release README](https://github.com/meituan-longcat/LongCat-Video/blob/main/README.md) describes more natural mouth movement and less accumulated error during long generation. The team says the model can produce videos longer than ten minutes. Those are developer-reported capabilities, not a guarantee that an arbitrary ten-minute recording will produce a usable result.

The practical audience is a team that wants control over an avatar pipeline: inputs, generation settings and infrastructure. An open checkpoint is useful for that, but it still leaves you responsible for installation, compute, output review and permission to use the person’s likeness and voice.

## Prepare the environment and inputs

Follow the repository installation sequence before running the command below. It specifies Python 3.10, a CUDA-compatible PyTorch installation, FlashAttention and separate avatar requirements, including audio-processing dependencies. Download both the base LongCat-Video weights and the Avatar 1.6 package as documented. Keep the downloaded folder names aligned with the checkpoint path you pass to the script.

For a controlled first attempt, use a short recording and a portrait you own or have permission to animate. Avoid starting with multiple speakers, background music and a long script together. If the output fails, too many moving parts make diagnosis unnecessarily difficult.

Create `my-avatar.json` using the structure in the [official input example](https://github.com/meituan-longcat/LongCat-Video/blob/main/assets/avatar/single_example_1.json):

```json
{
  "prompt": "A presenter faces the camera and speaks calmly.",
  "cond_image": "inputs/presenter.png",
  "cond_audio": {
    "person1": "inputs/speech.wav"
  }
}
```

These paths are illustrative. Replace them with files that exist in your checkout. Keeping the portrait, recording and prompt fixed gives you a useful baseline when comparing settings later.

## Pass the version and mixing settings explicitly

This adapts the official two-GPU image-and-audio example:

```bash
torchrun --nproc_per_node=2 run_demo_avatar_single_audio_to_video.py \
  --context_parallel_size=2 \
  --checkpoint_dir=./weights/LongCat-Video-Avatar-1.6 \
  --stage_1=ai2v \
  --input_json=my-avatar.json \
  --model_type=avatar-v1.6 \
  --use_distill \
  --lora_path=loraB/lora.safetensors \
  --mix_steps=4 \
  --mix_alpha=1.0
```

**Two GPUs here describes the published example, not a verified minimum hardware requirement.** Match the process and context-parallel counts to your deployment rather than assuming this command will fit any pair of GPUs.

The [model card](https://huggingface.co/meituan-longcat/LongCat-Video-Avatar-1.6) identifies `loraA` as DMD and `loraB` as OPSD. With the explicit mixing values above, DMD handles the first four denoising steps and OPSD the remaining four. LoRAs are small adapters that modify the base model’s behaviour; mixing them changes the generation process, not merely the output filename.

## Why omitting flags can change your result

The [single-person script](https://github.com/meituan-longcat/LongCat-Video/blob/main/run_demo_avatar_single_audio_to_video.py) currently defaults to `mix_steps=3` and `mix_alpha=0.5`, while the model card’s example uses four and one. It also resets mixing for the older model versions. Explicit arguments make the intended configuration visible and avoid depending on whichever defaults your checkout contains.

The parser rejects simultaneous FP8 and INT8, and rejects active 1.6 mixing without distillation. It restricts mixing steps to zero through eight and the mixing weight to zero through one. Before investigating a poor render, save the exact command and check these settings. A reproducible configuration is more useful than a screenshot of the final frame.

## Measure warm performance and review the whole clip

The model card warns that compilation adds overhead to the first runs and that new input shapes can trigger recompilation. Compare warmed runs with the same shapes before deciding an optimisation made inference slower. Its FP8 option requires Transformer Engine; enable one optimisation at a time so quality or dependency failures remain traceable.

For output review, inspect the beginning, middle and end for mouth movement, identity drift and hand consistency. Then extend the recording gradually. **A longer successful render is not automatically a better production asset.** The useful metric is how much acceptable footage survives review per unit of compute and editing time.

The [project page](https://meigen-ai.github.io/LongCat-Video-Avatar-1.6-Page/) offers matched-input comparisons with 1.5 and commercial systems, but its demonstration material is marked academic-only. Do not treat permission to use the MIT-licensed model weights as permission to reuse those demo identities or clips in an advertisement.

For a team building avatar content, the first milestone should be a repeatable short clip with known settings and clear usage rights. Once that works, longer generation becomes an engineering question you can measure instead of a promise borrowed from a launch page.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn complex releases and practical workflows into stories developers can understand and use.

[Book a 30-minute call](https://cal.com/paras-madan/30min).
