vjepa

High-throughput, resume-safe vision and video embedding pipelines for VLM datasets—26 pinned encoders, streaming Hugging Face ingestion, and Safetensors shard commits.

Open-source vision stack with stereo camera hardware, GPU processing, and AI agent for training video classifiers.

vjepa / vjepa2 / vjepa2.1 PCA visualization utility for dense features and world model inspection.

A physics-based video search engine using Meta's V-JEPA 2 world model to find videos with similar motion dynamics.

Can the V-JEPA2 model be used as a world model?

Where V-JEPA 2.1's dense video features hold up under noise, blur, frame drops, occlusion and low light, across all four model sizes. Code and data for the Poisson Labs post.

Masked Multi-Component Gated Decomposition Architecture

Assess Data Quality Before Annotation or Labelled Data Quality after Annotation (Txt files/Yolo Format). Visualise the patterns covered by each class/activity.

V-JEPA for Gray-Scott dynamics. Initial work produced during the 24-hour Hack the World(s) hackathon. 1st place 🏆

Patch-level predictive surprise from video foundation model embeddings. The embedding delta is the attention signal.

From-scratch JEPA world-model engine in PyTorch (FSDP, bf16, SIGReg/VICReg): data curation, distributed training and inference optimisation, evaluated on CAMELS cosmology against control floors with seed noise bands.

Badminton rally auto-clipping: V-JEPA 2 probe, zero-training cross-venue P/R 0.85/0.84 | 羽毛球整场视频→回合合集+精彩集锦

SCOUT: frozen-encoder embedding-space prediction for sim-to-real text-based person retrieval. ECCV 2026 Workshop (AI City Challenge Track 4), accepted as a poster.

Locally-Hosted Media Gallery App with AI Similarity Search

AC-VJEPA: adversarial-critic extension of V-JEPA; three preregistered experiments, honest negative results.