EvoWorld enables causal evolution in generative video models
Original title: EvoWorld: Towards Out-of-Sight World Evolution in Generative Video Models
EvoWorld uses a vision-language model as world brain and autoregressive renderer for out-of-sight causal evolution
Existing diffusion-based video models lack an explicit mechanism for maintaining world state and reasoning over event-conditioned consequences. We propose the Out-of-Sight Causal Evolution problem requiring models to render events, infer consequences for world entities, and reveal causally plausible outcomes. EvoWorld framework has a vision-language model observe generated segments, maintain an evolving world account, and reason about plausible state transitions while a segment-autoregressive video renderer realizes them visually. Experiments show substantial improvements in event-conditioned causal evolution over recent models while preserving visual quality and consistency.
- EvoWorld-Data and EvoWorld-Bench introduced for training and evaluation
- Vision-language model serves as the world brain
- Substantially improves event-conditioned causal evolution
- Preserves visual quality and consistency
From the original article
1 min read at huggingface.co
Generative video models are increasingly used as visual world simulators, yet their primary strength lies in rendering plausible pixels rather than reasoning about how a world should evolve. Coherent world evolution requires more than visual fidelity: when an event occurs, a model must infer its consequences for the entities in the world and then render the resulting state.
By Yequan Bie, Dengyang Jiang, Zhixuan Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.