DiReCT : Disentangled Regularization of Contrastive Trajectories for Physics-Refined Video Generation

Abolfazl Meyarian †, Amin Karimi Monsefi †, Rajiv Ramnath, Ser-Nam Lim · 2026

Read DiReCT : Disentangled Regularization of Contrastive Trajectories for Physics-Refined Video Generation by Abolfazl Meyarian †, Amin Karimi Monsefi †, Rajiv Ramnath, Ser-Nam Lim (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-25931.txt.

Description

Meyarian and co-authors present DiReCT, a post-training method that makes generated videos obey physics better. Flow-matching generators often break basic physics while looking smooth. Contrastive training could help, but text prompts tangle scene content with physical behavior. That tangle makes the contrastive signal fight the main objective. DiReCT splits the signal into two scales. One uses far-off negatives for broad separation. The other builds hard negatives that share the scene but change one physical trait. Applied to Wan 2.1, physical commonsense scores rose by 16.7 percent.

Catalog record

CREATORAbolfazl Meyarian †, Amin Karimi Monsefi †, Rajiv Ramnath, Ser-Nam Lim
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60429
LICENSEhttp://creativecommons.org/publicdomain/zero/1.0/
RIGHTSCC0
SOURCEarXiv

Source: https://arxiv.org/abs/2603.25931

Browse the full library · About the Tower of Babel