Openspline: open-source realtime talking avatar hits 90+ FPS on an RTX 4090
Mikasa0xdev · reddit · 2026-10-18
Developer Mikasa0xdev released Openspline, an open-source avatar layer for existing voice AI agents. Pick an image and it generates a realtime lip-synced talking avatar locally on your GPU; any existing voice agent can be integrated in 5 lines of code. The lite model runs at 90+ FPS on an NVIDIA RTX 4090. The project is at a very early stage but results look promising — code at github.com/openspline/openspline with a demo at openspline.dev (screenshot shows a live video call with a Nikola Tesla avatar).
More from Multimodal
- MOSS-Transcribe-Diarize beats VibeVoice-Large at ~10% the size, early tester reports — SharkWipf · 2026-10-18
- Gemini 4 one-shots a Gojo vs Sukuna 3D fight scene as a single HTML file — Ryoiki-Tokuiten · 2026-10-18
- Redditor shows 'Final Boss' anime scene built entirely with Opus 5.5 — mahamara · 2026-10-18
- Voice demo built with Jev draws "holy crap, that's the missing thing to get rid of keyboards" reaction — schwentker · 2026-10-18
- Video-Index 1.1: agentic video understanding beats feeding 512 fixed frames — thoma_gu · 2026-10-18
- Pixel-art cat curled around a planet, generated by Opus 5.5 — liminal_bardo · 2026-10-18