Evaluation metrics & methods
Fine-grained metrics for visual quality, prompt alignment, temporal consistency, physical realism, and human preferences, including interpretable automated evaluators.
2nd Workshop Proposed for CVPR 2027 / June 20–25, 2027
Video Generation and World Models:Benchmarks and Evaluation
From video synthesis to interactive world simulation.
Advancing reliable, reproducible, and human-aligned evaluation.
Overview
VGBE 2027 focuses on benchmarks and evaluation for video generation and world models. As generative systems evolve from content synthesis toward interactive world simulation, the workshop aims to advance reliable metrics, standardized benchmarks, and reproducible protocols for assessing both model capabilities and practical utility.
Topics span perceptual quality, prompt alignment, physical realism, temporal dynamics, and interactive behavior, alongside task-specific evaluation for robotics, autonomous driving, and embodied AI. A central question is how to connect general-purpose assessments with performance in real-world applications.
The workshop also explores how human-aligned evaluators can provide reliable feedback for reward modeling, reinforcement learning, and iterative model improvement. We aim to bring together researchers from academia and industry to identify critical failure modes and develop rigorous evaluation practices for video generation and world models.
Topics
Datasets & reproducible benchmarks
Diverse evaluation datasets, standardized benchmarks, and reproducible protocols that enable systematic comparison across video generation and world modeling systems.
Trustworthy evaluation & reward feedback
Human-aligned evaluation of safety, bias, fairness, and explainability, with reliable feedback for reward modeling, reinforcement learning, and post-training.
Video generation models
Text-to-video, image-to-video, video-to-video, and video editing, alongside autoregressive and streaming generation, controllability, and long-form temporal coherence.
World models & interactive simulation
Action-conditioned prediction, spatial and physical consistency, long-horizon world dynamics, interactive simulation, and closed-loop agent performance.
Applications & downstream utility
Task-specific assessment in robotics, autonomous driving, embodied AI, and low-level vision, with applications in science, medicine, engineering, and education.
Workshop program
Schedule
We are currently planning VGBE 2027 as a full-day workshop, featuring invited talks, paper and poster presentations, and community discussions. The full-day format is provisional and subject to workshop acceptance and final scheduling.
Invited talks
Research perspectives on the next generation of video models and their evaluation. Talk titles: To Be Announced (TBA).
Paper & poster sessions
A forum for discussing new evaluation methods, benchmarks, and empirical findings.
Community discussions
Discussions on reproducibility, shared evaluation protocols, and future research directions.
The detailed timetable, session times, and talk titles will be announced once finalized.
Research perspectives
Invited Speakers
Alan Bovik
University of Colorado Boulder
Tentative
Ming-Hsuan Yang
UC Merced · Google DeepMind
Tentative
Jiajun Wu
Stanford University
Tentative
Wenhu Chen
Meta Superintelligence Labs
Tentative
Balu Adsumilli
YouTube / Google
Tentative-
Jun Gao
University of Michigan · NVIDIA
Tentative -
Yi Jiang
ByteDance Seed
Tentative 
Yan Wang
NVIDIA Research
Tentative
Competition
TBD
Challenge details will be announced here.
Planning ahead
Important Dates
- Call for papers
- To Be Announced (TBA)
- Paper submission deadline
- To Be Announced (TBA)
- Author notification
- To Be Announced (TBA)
- Camera-ready deadline (accepted archival papers)
- (AoE)
- Challenge dates
- To Be Announced (TBA)
- Workshop date & venue
- To Be Announced (TBA)
Call for Contributions
We invite contributions on the evaluation and benchmarking of video generation and world models, including new metrics, reproducible benchmarks, human-aligned assessment, and application-driven studies. We welcome both preliminary findings and original research in the following formats.
Up to 4 pages
Preliminary research, work in progress, and technical reports.
Excluding references
Up to 8 pages
Original research contributions with developed methods and empirical evaluation.
Excluding references
Papers should use the official CVPR 2027 format and author guidelines and be anonymized for double-blind review. Deadlines and publication details will be announced here.
Submissions are not yet open. The workshop's OpenReview submission link is TBD.
We plan to give one Best Paper Award and one Best Paper Runner-Up Award, along with several oral presentation slots.
The VGBE community
Previous Editions
Explore the first workshop and its archived program.
Benchmarks and Evaluation
Stay connected
Questions about the proposal?
Reach out to the VGBE team.
Organizers
To Be Announced (TBA)