Search papers, labs, and topics across Lattice.
This paper introduces VGI-bench, a comprehensive evaluation framework designed to assess the visual reasoning capabilities of video generation models through 27 tasks and 810 instances. The study reveals that while current generative systems, such as Seedance 2.0, can tackle some visually grounded reasoning tasks, they still fall short of reliability, achieving only 51.0% accuracy under the proposed evaluation criteria. Additionally, the analysis uncovers critical insights into output failure modes and the limitations of self-correction in these models, highlighting areas for future improvement.
Current video generation models struggle with visual reasoning, achieving only 51% accuracy on a new benchmark designed to probe their capabilities.
Recent studies suggest that video generation models can exhibit certain forms of zero-shot visual reasoning through generated frames. Yet reliable evaluation remains challenging: benchmarks should adopt inputs aligned with the visual priors of current video models, require valid evolving processes rather than only plausible final states, and calibrate task difficulty to remain challenging yet partly feasible. To this end, we introduce VGI-bench, containing 27 tasks and 810 instances, organized by a two-level taxonomy of task domains and skill tags for fine-grained evaluation of visual reasoning capabilities of video generation models. Our evaluations show that current generative systems can solve a subset of visually grounded reasoning tasks, but remain far from reliable, with even the strongest model, Seedance~2.0, achieving only 51.0% under our evaluation criteria. Our analysis further explore the output failure modes, input condition sensitivity, performance transfer boundary from synthetic fine-tuning, and internal denoising perspective revealing limited self-correction, where later steps mainly refine early hypotheses rather than correct reasoning errors. We hope VGI-bench will help stimulate the development of next-generation video generation models. We will release our code and data.
Microsoft Research