U-JiT does pixel-space video generation without a VAE, hitting SOTA on Kinetics-600
LucaAmb · x · 2026-11-05
U-JiT is a video generation architecture that skips the video autoencoder entirely: it generates directly in pixel space and is trained end-to-end, achieving state-of-the-art results on Kinetics-600. The key insight is rethinking compression through the architecture itself, avoiding the compression-reconstruction tradeoff of pre-trained VAE + latent DiT pipelines.
More from Multimodal
- Sketch to printable STL in 4.5 minutes: an open-source Kimi K3.1 Pro CAD workflow — Arindam_1729 · 2026-11-05
- ReSeed: 0.002% extra params in first denoising step restore diversity lost to distillation — kwangmoo_yi · 2026-11-05
- ReSeed: tiny first-step biases restore diversity lost in diffusion distillation — kwangmoo_yi · 2026-11-05
- ChatGPT adds GIF generation with Images 2.6 model as OpenAI hits 4B images weekly — The Decoder · 2026-11-05
- Viggle Remake launches with open weights: recast any viral clip with your AI character at $0.96/10s — ZabihullahAtal · 2026-11-05
- Dev builds a handwriting engine that mimics penmanship styles from 1522 to today — floguo · 2026-11-05