LongCat Video Avatar 1.5 Tested — Open-Source Talking Avatar AI (Honest Verdict)

LongCat Video Avatar 1.5 is Meituan's new open-source talking-avatar model. Feed it one image plus a voice clip and it generates a lip-synced video of that face speaking. The base model dropped a couple of months ago; this 1.5 release is the upgraded framework. I ran it end-to-end on a cloud H100 and gave it the honest test.

Verdict: it works, but it isn't there yet. A single generation took ~20 minutes on an 80GB H100 (RunPod, ~$3.33/hr) and the mouth movement came out overdone. For real work I'd reach for a faster tool — but as an open-source learning exercise, the full setup is worth seeing.

What's new in 1.5

• Whisper-large audio encoder — replaces the old back-to-back encoder for better lip sync.
• 8-step inference via DMD2 distillation — fast generation instead of 50+ diffusion steps.
• 8-bit quantization — lower VRAM footprint, though it still wants roughly 80GB of GPU.
• Tasks supported: audio+text→video (AT2V), audio+text+image→video, and video continuation.
• License: MIT, fully open source. Weights are ~74GB on Hugging Face.

How to run it

The weights are 74GB and it needs ~80GB VRAM, so a cloud GPU is the practical route. Steps from the video:

1. Rent the GPU. On RunPod, deploy an H100 80GB pod, bump container disk to ~250–300GB, open a couple of extra ports, launch JupyterLab.
2. Update + clone. apt-get update && apt-get upgrade -y, then git clone the LongCat-Video repo and cd in.
3. Install Miniconda, then create the env: bash conda create -n longcatvideo python=3.10 -y conda activate longcatvideo
4. Install dependencies — torch first, then the repo requirements. FlashAttention-2 is the one that fights back (see the pro tip).
5. Audio libs: conda install -c conda-forge librosa ffmpeg. If inference complains, also pip install pyloudnorm audio-separator.
6. Download the weights with the Hugging Face CLI + hf_transfer — both the Avatar 1.5 model and its base model.
7. Configure the prompt JSON — set a text prompt, your prompt_face.jpg, and a prompt_voice.wav voice reference inside assets/avatar/.
8. Run inference with torchrun (--nproc_per_node 1, context-parallel size 1) pointing at your JSON. Output lands in the output folder.

Pro tip — the FlashAttention fix

The install that breaks everyone is flash-attn. Installing it normally throws No module named torch because it tries to build in an isolated environment without torch present. The fix the video lands on: install it without build isolation so it sees the torch you already have:

pip install flash-attn --no-build-isolation

Do that after torch is in the env, and the rest of requirements.txt installs clean.

Should you use it?

Honestly — not for production yet. ~20 minutes per clip on an H100 for output with too much mouth motion is a hard sell when faster avatar tools exist. But it's MIT-licensed and open, the pipeline above is a clean template for spinning up any heavy model on RunPod, and with a better source image the results should improve. Try it to learn the workflow.

👉 GPU I used: RunPod · Hostinger (coupon PROMPT for 10% off) 🔗 Model: meituan-longcat/LongCat-Video-Avatar-1.5

For the full walkthrough — every error and fix in real time — watch the video above and subscribe on the channel.