VRhino
Run AI video models locally with a shared native runtime — no per-model Python environment.
English | 简体中文
Install
Linux x86_64 — install v0.9.4-alpha Distribution v2 with one command (no sudo):
curl -fsSL https://raw.githubusercontent.com/pixelrhino-ai/vrhino/v0.9.4-alpha/install.sh | shThe installer downloads and verifies the core and its exact private NVIDIA runtime. The first install downloads both packages; later core upgrades reuse the same runtime when its identity is unchanged. Follow the final PATH instruction for the current terminal, then run:
vrhino doctor
Manual download / installation options · Windows v0.8 CUDA package
A compatible NVIDIA GPU/driver and sufficient memory are required. Model weights are separate; native execution needs no Python or CUDA Toolkit installation.
Measured Wan2.1 improvements
On the tested RTX 4090 D / Wan2.1 T2V 1.3B task, controlled source A/B generation time fell 7.72% and device-total sampled peak fell 30.93%. Checkpoints and video outputs matched on the tested inputs. These are scoped measurements, not claims for every model/GPU or cross-framework quality. Configuration, raw timing rows and limitations.
Available models
| Model | Task | Entry |
|---|---|---|
| Wan2.2 T2V A14B | Text to video | pull / run · vrhino/wan2.2-t2v-a14b:1.0.0 |
| Wan2.1 T2V 1.3B | Text to video | pull / run · vrhino/wan2.1-t2v-1.3b:1.0.1 |
| LTX-Video 0.9.1 | Text to video | pull / run · vrhino/ltx-video-v0.9.1:1.1.1 |
| Mochi 1 Preview | Text to video | pull / run · vrhino/mochi-1-preview:1.0.1 |
| MuseTalk 1.5 | Lip sync | pull / run · vrhino/musetalk-v1.5:1.0.1 |
| LatentSync 1.6 | Lip sync | pull / run · vrhino/latentsync-1.6:1.0.1 |
Hardware requirements and tested platforms vary by model; see the installation and model notes. VRhino is an alpha project.
Generate a video
vrhino pull vrhino/wan2.2-t2v-a14b:1.0.0
vrhino run vrhino/wan2.2-t2v-a14b:1.0.0 \
--prompt "a red panda running through fresh snow" --output video.mp4The first pull downloads, verifies and converts the model into a local package.
To use a larger model/cache disk, set VRHINO_HOME before pulling.
Wan2.2 uses the same Product command as other text-to-video models. Its first pull is large because VRhino verifies and converts the fixed upstream weights. See the Wan2.2 notes for storage, hardware and alpha qualification details.
Lip sync
vrhino pull vrhino/musetalk-v1.5:1.0.1 vrhino run vrhino/musetalk-v1.5:1.0.1 \ --video input.mp4 --audio driving.wav --output lipsync.mp4
LatentSync uses the same video/audio command shape with its package ID above.
More
- Installation and cache configuration
- Architecture · Build from source
- Local Native API — v0.6.0-alpha contract (
vrhino serve) - v0.9.4-alpha release and checksums · Windows v0.8 scope
- Contributing
Project-owned source: Apache-2.0. Distributed binaries retain their binary license; dependencies and models retain their own terms. See third-party notices and model license notes.