wtf is liveavatar-app-nz?

LiveAvatar for Cog packages an existing model called LiveAvatar so it can run as an on-demand, pay-per-use service on a hosting platform called app.nz, using a deployment format called Cog. It does not build the animation model itself, it wraps the upstream LiveAvatar project, which turns a single reference image, a driving audio clip, and a short scene description into a video of that character talking or singing. The underlying model can animate photorealistic people, stylized or anime characters, or animals, adding facial expression and body motion guided by a text prompt alongside the driving audio. It generates video in short repeated chunks rather than one continuous stream, which the author says suits scale-to-zero request and response serving, not an always running live stream. Running it yourself needs serious hardware. The README recommends an H100 or H200 card with 80 gigabytes of memory, since the full precision model needs that much on one card, a lower precision mode is reported by the upstream project to run on 48 gigabyte cards, though this packaging has not yet confirmed that smaller setup works end to end. You also need at least 64 gigabytes of disk space, since the model weights add up to roughly 47 gigabytes and download automatically the first time the service starts, from a persistent folder, from app.nz's own model mirror, or from Hugging Face. A single prediction takes a character image, up to 300 seconds of audio, and a text prompt describing the scene, plus settings for output quality, how many clips to generate, and how many generation steps to run, returning an MP4 with the original audio built in. Environment variables let you tune memory further, such as moving parts of the model to the CPU between steps or enabling compilation for faster repeated runs at the cost of a slower first run. The project is pinned to one specific upstream commit rather than tracking ongoing changes. Automated tests run without touching a GPU or downloading weights, while a separate manual checklist checks lip sync quality and consistency on real hardware before a release ships. The README asks that this tool only be used to animate people or characters you have permission to use, with synthetic video clearly labeled. The adapter and everything it depends on upstream are released under the Apache-2.0 license.