ResearchAug 29, 2026 · 12:374 sources

LeVJEPA Matches V-JEPA 2 With 20x Less Video Pretraining Compute

From huggingnews.com

Image: huggingnews.com
AI;DR

LeVJEPA is a new pretraining method that achieves performance equivalent to V-JEPA 2 while reducing compute costs by 5.6 to 20.8 times. The open source method replaces target encoders and reconstruction machinery with a single encoder trained using invariance and SIGReg. It maintains appearance and motion understanding even when 95% of video tokens are randomly dropped.

Read the sourceSign in to suggest

Sources

+ 3 more
  1. SUPPORT@randall_balestr“open source + reproducible”x.com
  2. SUPPORT@askalphaxiv“randomly dropping 95% of video tokens, preventing representation collapse”x.com
  3. SUPPORT@ylecun“A new Pareto frontier in video pretraining”x.com
#open-source#inference#videoMarkdown
Content history (1)
  • 2026-10-04 · Summary · vi · Wording fix
    Phương pháp pretraining mới này thay thế các bộ mã hóa mục tiêu (target encoders) và cơ c…
    Phương pháp pretraining mới này thay thế các bộ mã hóa mục tiêu và cơ chế tái tạo bằng mộ…

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today