CORTEXA
← Browse
arxivcs.CVcs.LG2026-07-02

Object-centric LeJEPA

Jakob Geusen, Ender Konukoglu

Image encoders trained with LeJEPA can deliver strong features for downstream tasks, but, like other image-level self-supervised methods, typically require large training datasets. Aligning representations at the level of objects rather than whole scenes promises greater data efficiency, but doing this in a completely self-supervised way, effectively jointly partitioning a scene and representing its objects, is unstable: the two are locked in a cyclic dependency, partitioning requires meaningful representations, while meaningful representations require consistent partitioning. We sidestep this instability by taking object masks as given during training, using cheap, off-the-shelf SAM proposals. We extend LeJEPA - whose distributional anti-collapse objective ports naturally from whole images to variable-sized sets of objects - to align object-centric representations rather than whole images. An additional instance-separating loss, which treats other objects in the same scene as negatives, further boosts downstream performance. Across two model scales and 10-100% of COCO, object-level LeJEPA outperforms image-level LeJEPA on tracking (DAVIS), classification (ImageNet-1k), segmentation (ADE20k), and re-identification (NAVI).

View free PDFSource page

Related papers

arxivcs.GRcs.CVcs.LG2026-06-29

RenderFormer++: Scalable and Physically Grounded Feed-Forward Neural Rendering

Huangsheng Du, Haoran Zhu, Youcheng Cai, Jinyang Meng, Ligang Liu

We present RenderFormer++, a scalable and physically grounded feed-forward neural rendering framework for global illumination in mesh scenes. Existing Transformer-based neural rendering methods such as RenderFormer achieve promising cross-scene generalization, but suffer from lim…

View free PDFSource page
arxivcs.CVcs.LG2026-07-01

Information-Regularized Attention for Visual-Centric Reasoning

Guohao Sun, Xiaofang Wang, Yash Patel, Mengchen Liu, Zhiqiang Tao, Praveen Krishnan

Vision-language models (VLMs) have become a paradigm for multimodal learning, yet remain unstable due to object hallucination, weak visual grounding, and catastrophic forgetting after full-parameter instruction tuning. We claim these failures result from a lack of explicit contro…

View free PDFSource page
arxivcs.LGcs.AIcs.CV2026-07-16

Multi-Axis Max@K Reinforcement Learning for Representative Diversity in Text-to-Image Generation

Ku Onoda, Paavo Parmas, Hiroki Furuta, Soichiro Nishimori, Yuta Oshima, Shohei Taniguchi, et al.

Text-to-image (T2I) models can synthesize realistic, prompt-aligned images, yet samples generated for the same prompt often cover only a small subset of visually distinct modes. This limits the diversity of images, and for person-centric prompts, can reflect or amplify demographi…

View free PDFSource page