v-jepa

A tiny, fully-reproducible JEPA world model that learns the physics of a bouncing DVD logo in representation space, dreams its future, and detects anomalies. Trains on a CPU in ~10s. Interactive browser demo. CA: 0x42bef487C250dd054035d5E8d69C12549d1F7Ba3

Single-cell fate from live-cell imaging: where identical cells take two fates, and when it becomes visible (S-BIAD2515)

A phone video of a person becomes a Unitree G1 skill: SAM 3.1, SAM 3D Body, Muse Spark and V-JEPA 2 in, a trained humanoid policy out.

A quaternion-enhanced Video Joint-Embedding Predictive Architecture with continuous spectral autoencoders, Fourier-domain topology, and a 2D torus latent space for world model learning.

Frozen V-JEPA 2 (300M params) + PPO policy head achieves +763 mean reward on CarRacing-v3 — 2.4x over CNN baseline, single RTX 3080

A researcher map of open-source world models - what exists, what each line of work claims, and where its authors say it breaks.

Educational implementation of V-JEPA (Video Joint Embedding Predictive Architecture) in PyTorch. Includes Conv3D tubelet embeddings, Video ViT, EMA target encoder, spatiotemporal masking, linear probing, retrieval evaluation, and effective-rank analysis on Something-Something V2.

Crop segmentation from Sentinel-2 time series (PASTIS) by transferring video transformers: TimeSformer and V-JEPA 2.1 with LoRA

Event detection and zero-shot spatial localization with V-JEPA 2.1, PEFT, Integrated Grandients and VTCD

See what a JEPA video encoder sees: a local latent inspector for Meta's V-JEPA 2, with PCA feature overlays and a latent surprise timeline.

Do frozen video encoders and a generative video tokenizer see hidden object physics? Linear probes on V-JEPA 2, DINOv2 and the Cosmos tokenizer over MuJoCo push clips with hidden mass and friction.

Action-conditioned latent world model and experiment code for RoboCasa365.

Open-source ecosystem for egocentric robot data: raw video in, a V-JEPA 2.1 model predicts every step, a review app lets people approve or fix them, and approved steps become training data that retrains the model. 92.5% frame accuracy on held-out video; review time per clip 40-50min → 5min