Causal Initiation
Suppresses motion before contact or interaction evidence appears.
ECCV 2026
Event-Driven Video Generation
EVD turns frame-first video generation into event-grounded state transitions, improving causal interactions without sacrificing appearance.
Core Idea
Events should decide when state is allowed to change.
Event Realization
Concentrates updates where the interaction should actually happen.
Stable Postconditions
Reduces late drift after objects settle, stop, close, or land.
Contact Stability
Improves support, placement, and physically plausible interaction outcomes.
Curated Comparisons
EVD vs. strong video generation baselines.
Each row uses the same prompt across Sora, Movie Gen, DiT, and EVD. EVD is highlighted to make the method comparison easy to scan.
More Samples
More Event-Driven Samples.
Browse EVD generations grouped by the EVD-Bench taxonomy: state persistence, spatial accuracy, support relations, and contact stability. These samples focus on the final EVD output so the breadth of event-grounded behavior is easy to scan.
Inspect all 4-way comparison clips
Method Overview
A lightweight event pathway gates video DiT updates.
EVD predicts token-aligned event activity, forms a stable event gate, and applies that gate to the denoising update so only event-supported regions are allowed to change state.
Citation
BibTeX
@inproceedings{maduabuchi2026eventdriven,
title = {Event-Driven Video Generation},
author = {Maduabuchi, Chika and Wang, Jindong},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026}
}