Human-JEPA: A Human-Centric Vision Model that Perceives and Anticipates

Human-JEPA is a human-centric video model designed to recognize the present and anticipate near-future motion, extending V-JEPA 2.1 with anchored forecasting rather than changing its architecture. Its dense context targets are tied to a frozen copy of the original initialization, preventing silent degradation during continued pretraining, while an image branch trained on LUPerson-T crops preserves appearance and identity features. The model also replaces spatial block masking with a strict past-to-future split, forcing latent prediction to represent scene dynamics instead of copying visible content. Trained on 536,699 Kinetics-700 clips plus 958k LUPerson-T crops, the 0.3B-parameter Human-JEPA reaches 0.620 AP for pose and 0.4635 mAP on Market-1501 person re-identification, outperforming the larger Sapiens2 specialist on both measures while conceding high-resolution parsing. Its predictor achieves 0.873 cosine agreement with the actual future on held-out clips, and unlike the released V-JEPA 2.1 head, it does not reduce early-action accuracy. Ablations show why the design matters: block masking causes a 17-point re-identification collapse and roughly a five-point action penalty, whereas forecasting preserves action performance near the frozen base. The paper also tests nine person-level prediction objectives and finds no reliable partner modeling, concluding that scene-level future prediction is more effective than masking or forecasting individual people.