Rethinking Visual Intelligence: Insights from Video Pretraining

TL;DR

This paper explores how pretraining on video data with diffusion models enhances visual understanding and adaptability, showing superior data efficiency over language models in various tasks.

Contribution

It introduces the use of video diffusion models for pretraining, demonstrating their potential as visual foundation models with strong inductive biases.

Findings

VDMs outperform LLMs in data efficiency across multiple benchmarks.

Video pretraining provides beneficial inductive biases for visual tasks.

VDMs show promise for broad visual problem-solving capabilities.

Abstract

Large language models (LLMs) have demonstrated that large-scale pretraining enables systems to adapt rapidly to new problems with little supervision in the language domain. This success, however, has not translated as effectively to the visual domain, where models, including LLMs, continue to struggle with compositional understanding, sample efficiency, and general-purpose problem-solving. We investigate Video Diffusion Models (VDMs) as a promising direction for bridging this gap. Pretraining on spatiotemporal data endows these models with strong inductive biases for structure and dynamics, which we hypothesize can support broad task adaptability. To test this, we design a controlled evaluation in which both a pretrained LLM and a pretrained VDM are equipped with lightweight adapters and presented with tasks in their natural modalities. Across benchmarks including ARC-AGI, ConceptARC, visual games, route planning, and cellular automata, VDMs demonstrate higher data efficiency than their language counterparts. Taken together, our results indicate that video pretraining offers inductive biases that support progress toward visual foundation models.