PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding

Daniel Bolya (Meta) · Po-Yao Huang (Meta) · Peize Sun (Meta) · Jang Hyun Cho (Meta Superintelligence Labs (MSL)) · Andrea Madotto (Meta Superintelligence Labs (MSL)) · Tengyu Ma (Meta) · Hanoona Bangalath (Mohamed Bin Zayed University of Artificial Intelligence) · Nikhila Ravi (Facebook) · Piotr Dollar (Facebook AI Research) · Christoph Feichtenhofer (Meta) · Kristen Grauman (University of Texas at Austin) · Salman Khan (MBZ University of AI) · Tammy Stark (Facebook) · Babak Damavandi (Meta) · Seungwhan Moon (Facebook) · Effrosyni Mavroudi (Facebook) · Triantafyllos Afouras (FAIR, Meta) · Tushar Nagarajan (Meta AI) · Muhammad Maaz (Mohamed Bin Zayed University of Artificial Intelligence) · Yale Song (Google) · Shuming Hu (Facebook) · Suyog Jain (Facebook) · Miguel Martin (University of Adelaide) · Huiyu Wang (FAIR, Meta) · Shashank Jain (State University of New York, Stony Brook) · Vivian Lee (Facebook) · Andrew Westbury (Facebook) · Philipp Kraehenbuehl · Lorenzo Torresani (Northeastern University)
black-box modelsdata gapsdistillationevaluation suitefine-grained question-answer pairshuman-labeled instancesimage understandingperception language modelplm–videobenchreasoning tasksreproducible frameworkspatio-temporally grounded captionssynthetic datatraining recipesvideo understandingvision-language models

Vision-language models are integral to computer vision research, yet many high-performing models remain closed-source, obscuring their data, design and training recipe. The research community has responded by using distillation from black-box models to label training data, achieving strong benchmark results, at the cost of measurable scientific progress. However, without knowing the details of the teacher model and its data sources, scientific progress remains difficult to measure. In this paper, we study building a Perception Language Model (PLM) in a fully open and reproducible framework for transparent research in image and video understanding. We analyze standard training pipelines without distillation from proprietary models and explore large-scale synthetic data to identify critical data gaps, particularly in detailed video understanding. To bridge these gaps, we release 2.8M human-labeled instances of fine-grained video question-answer pairs and spatio-temporally grounded video captions. Additionally, we introduce PLM–VideoBench, a suite for evaluating challenging video understanding tasks focusing on the ability to reason about ''what'', ''where'', ''when'', and ''how'' of a video. We make our work fully reproducible by providing data, training recipes, code & models.