LatSearch: Latent Reward-Guided Search for Faster Inference-Time Scaling in Video Diffusion

Zengqun Zhao 1 , Ziquan Liu 1 , Yu Cao 1 , Shaogang Gong 1 , Zhensong Zhang 3 , Jifei Song 3 , Jiankang Deng 2 , Ioannis Patras 1 · 2026

Read LatSearch: Latent Reward-Guided Search for Faster Inference-Time Scaling in Video Diffusion by Zengqun Zhao 1 , Ziquan Liu 1 , Yu Cao 1 , Shaogang Gong 1 , Zhensong Zhang 3 , Jifei Song 3 , Jiankang Deng 2 , Ioannis Patras 1 (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-2603-14526.txt.

Description

Zhao and co-authors speed up inference-time search for video diffusion models. Hunting for golden starting noise can raise video quality. Past searches score only final decoded videos, so feedback comes late and costs much. Their latent reward model scores partly denoised frames at any step. It rates visual quality, motion quality and text match. LatSearch uses those scores to guide the search. The team reports better control and sample efficiency at lower cost than final-video search methods.

Catalog record

CREATORZengqun Zhao 1 , Ziquan Liu 1 , Yu Cao 1 , Shaogang Gong 1 , Zhensong Zhang 3 , Jifei Song 3 , Jiankang Deng 2 , Ioannis Patras 1
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65641
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.14526

Browse the full library · About the Tower of Babel