V-RAE: Rethinking Video Latent Spaces via Vision Foundation Models
sainingxie · x · 2026-08-27
This paper introduces V-RAE, a video representation autoencoder that builds generative latents on top of frozen vision foundation model representations. It uses a lightweight temporal pooling module to remove redundancy while preserving semantic structure. V-RAE achieves 2.13 rFID on K600 and converges up to 6x faster. The authors also introduce tFVD, a temporal-coherence diagnostic for downstream generation quality.
More from Multimodal
- Magnific One generates fake Zhangjiajie tourist photos indistinguishable from real shots — techhalla · 2026-10-15
- Midjourney showcase: dark glass-cube abstract renders — ciguleva · 2026-10-15
- ByteDance's OmniHuman 2.0 ships: 5-minute single-take avatars at 40% lower price — DavidmComfort · 2026-10-15
- Topview open-sources 3D Builder: one prompt gives an editable 3D set for AI filmmaking — alex_verem · 2026-10-15
- Gothic Church 3D Gaussian Splat Coming to Vision Pro, Hailed as Sharpest Ever — willeastcott · 2026-10-15
- Neon Noir prompt template goes viral: cinematic bar shots with one Flux prompt — LudovicCreator · 2026-10-15