OpenSpline: video generator — what it does and what you need to run it

OpenSpline is an open source video generator that creates an animated talking avatar in real time. The project is built around pairing with voice AI: the avatar syncs lip movements to speech and runs locally. As described by the author, five lines of code are enough to get it running.

What it does

OpenSpline gives voice AI a "face." The generator creates video of an avatar that talks in real time, with lip sync. The project is designed as a local solution — processing happens on the user's machine, with no data sent outside.

The repository tags list integrations with ElevenLabs, Gemini, OpenAI and LiveKit. Pipecat and WebRTC are also mentioned, which points to working alongside voice pipelines and streaming video over the network. The authors do not disclose in the fact sheet any details about supported output formats or how voice models are connected.

What you need to run it

The code is written in Python and distributed under the MIT license. The license permits use, modification and distribution of the code, including as part of other products, provided the license text is retained.

The platform is CUDA, meaning an NVIDIA graphics card is required. The authors note that PyTorch uses CUDA 12.8 for the Blackwell architecture, including RTX 50 series graphics cards, and CUDA 12.6 for older NVIDIA GPUs. Cards older than the Ampere architecture run in FP32 mode, which requires more VRAM and may be slower. AMD and Apple GPU support is not implemented yet. The authors specifically note that running on Blackwell and on cards older than Ampere still needs hardware verification.

Who it suits

OpenSpline will be useful to those building voice assistants, chatbots with a face, or interactive systems based on voice AI who want to add a visual interlocutor. The local mode of operation will suit those who do not want to send their video stream to third-party services. The project is also of interest to developers looking for an open foundation for real-time avatar experiments.

It will not suit those without an NVIDIA graphics card that supports CUDA: owners of AMD and Apple GPUs are left out of the current version. Users with older NVIDIA cards from before Ampere should keep in mind the higher VRAM requirements and a possible drop in speed. Those looking for a ready-made commercial product with an interface rather than code for integration are also unlikely to find OpenSpline a fit — it is a library, not an application.

OpenSpline is a young project: the repository was created in late September 2026, and the last code change was in October of the same year. This is worth keeping in mind when planning adoption: the project is actively developing, but some of the stated configurations are still being verified.