MMEval: Evaluating Video Generation Models for Motion Quality
Keywords: image-to-video, evaluation
Abstract: Recent advancements in video generation, especially with diffusion models, have led to new challenges in evaluating the generated outputs, highlighting the need for well-curated evaluation metrics and benchmarks. While prior work has focused on assessing text-to-video models for overall video quality, such as temporal coherence and prompt consistency, they overlook a crucial aspect: motion modeling abilities of generative models. To address this gap, we propose a structured approach to evaluate image-to-video generation models, with a focus on their motion modeling abilities. For example, we assess how accurately models generate motions like "circular movement for a rotating ferris wheel" or "oscillatory motion for a pendulum". We categorize videos into linear, circular, and oscillatory motion-types and formulate metrics to capture key motion properties for each category. Our benchmark, MMEval, along with the code and image-prompt-video sets, will be publicly released.
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5679
Loading