EdenLive weights on Hugging Face: 210GB across 14B/5B few-step and bidirectional DiT variants
AdinaYakup · x · 2026-10-18
The EdenLive model card on Hugging Face (VAST-AI/EdenLive) details the real-time generative rendering system from VAST: it renders a proxy/reference video stream, target first frame, and text prompt into refined or stylized video, with references from 2D/3D scenes, engines, or recordings. Control is injected through a post-trained DINOv3 encoder embedding each proxy frame to condition the diffusion transformer; upstream code maintains interaction state while the model handles rendering. Demonstrated at 24+ FPS with minute-scale generation and 480P/720P output; the 5B variant deploys on a single GPU at 720P. Weights total about 210GB across dit-14b, dit-14b-bidir, dit-5b and dit-5b-bidir (few-step chunk-AR and bidirectional DiT), loaded fully from a local directory.
Related event: VAST Open-Sources EdenLive Real-Time Generative Renderer(3 posts)→
More from Embodied
- China installs 9x more industrial robots than US, and DC is worried — binarybits · 2026-10-18
- Giving the rabbit r1 a second life with llama.cpp and on-device s2s apps — mervenoyann · 2026-10-18
- Boston Dynamics' brutal lesson: Handle out, Stretch in as DHL signs for 1,000+ robots — TinfoilTricorn · 2026-10-18
- Stanford AA 274A Principles of Robot Autonomy lectures by Marco Pavone are freely available — caglar_ee · 2026-10-18
- VAST releases EdenLive: 5B real-time video model rendering at 24+ FPS on a single GPU — AdinaYakup · 2026-10-18
- Meta glasses trick: AI running route plus Google Maps Ask AI equals instant route map — IanAndrewsDC · 2026-10-18