Research map / Video / Action recognition
V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning
Lorenzo Mur Labadia, Matthew Muckley, Amir Bar, Mido Assran, Koustuv Sinha, Michael Rabbat, Yann LeCun, Nicolas Ballas, Adrien Bardes
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add weighted context loss on visible tokens (ablation) | V-JEPA 2 (ViT-L) · ADE20K | mIoU: 22.2 -> 33.9 | Improved |
| add weighted context loss on visible tokens (ablation) | V-JEPA 2 (ViT-L) · NYUv2 | RMSE: 0.682 -> 0.473 | Improved |
| add weighted context loss on visible tokens (ablation) | V-JEPA 2 (ViT-L) · SSv2 | top-1 accuracy: 72.8 -> 62.5 | Hurt |
| add weighted context loss on visible tokens (ablation) | V-JEPA 2 (ViT-L) · ImageNet-1K | top-1 accuracy: 82.2 -> 72.6 | Hurt |
| add deep self-supervision at intermediate encoder layers (ablation) | context-loss model · SSv2 | top-1 accuracy: 62.5 -> 72.0 | Improved |
| add deep self-supervision at intermediate encoder layers (ablation) | context-loss model · ADE20K | mIoU: 33.9 -> 38.6 | Improved |
| scale pretraining data to VisionMix163M (ablation) | V-JEPA 2.1 ViT-L · ImageNet-1K / ADE20K | accuracy / mIoU: 80.8 -> 81.6 on ImageNet, 38.6 -> 40.8 on ADE20K | Improved |
| add multi-modal tokenizer (2D conv for images) (ablation) | V-JEPA 2.1 ViT-L · ADE20K | mIoU: 40.8 to 41.4 | Improved |
| scale encoder from ViT-L to ViT-G (ablation) | V-JEPA 2.1 · ADE20K | mIoU: 41.4 -> 47.1 | Improved |
| add high-resolution cool-down phase (ablation) | V-JEPA 2.1 ViT-G · NYUv2 | RMSE: 0.365 -> 0.307 | Improved |
| distill ViT-G into ViT-L (ablation) | V-JEPA 2.1 ViT-L · SSv2 | top-1 accuracy: 74.2% to 76.5% | Improved |
| distill ViT-G into ViT-L (ablation) | V-JEPA 2.1 ViT-L · ADE20K | mIoU: 42.0 mIoU to 46.7 mIoU | Improved |
Closest papers by meaning
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.87)
- Visual Representation Learning with Stochastic Frame Prediction (ICML 2024 · similarity 0.86)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.86)
- seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models (NeurIPS 2025 · similarity 0.86)
- How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks (NeurIPS 2024 · similarity 0.86)
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
