This is a HuggingFace Transformers conversion of Meta's original V-JEPA 2.1 ViT-B checkpoint. Original weights belong to Meta AI; the license follows Meta's original terms. This checkpoint has been uploaded here pending transfer to Meta's official Hub organization.

Original repository: https://github.com/facebookresearch/vjepa2
Paper: V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

Model details

Parameter	Value
Architecture	ViT-B encoder
Crop size	384 px
Frames per clip	64
Distilled from	ViT-G teacher

Usage

python

import torch
from transformers import AutoVideoProcessor, VJEPA2Model

model_id = "davevanveen/vjepa2.1-vitb-fpc64-384"

processor = AutoVideoProcessor.from_pretrained(model_id)
model = VJEPA2Model.from_pretrained(model_id)
model.eval()

# video: list of PIL images or a (T, H, W, C) numpy array
# Here we use a random tensor as a placeholder
video_frames = torch.randint(0, 256, (64, 384, 384, 3), dtype=torch.uint8).numpy()

inputs = processor(videos=[video_frames], return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)

# Patch-level features: (batch, num_patches, hidden_dim)
features = outputs.last_hidden_state
print(features.shape)

Attribution

Original V-JEPA 2 work by Meta AI Research. If you use this model, please cite the original work:

bibtex

@article{vjepa2_2025,
  title   = {V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning},
  author  = {Assran, Mahmoud and others},
  year    = {2025},
  url     = {https://github.com/facebookresearch/vjepa2}
}