LeVJEPA 1.5 Large

11/05 UPDATE: Due to an error we uploaded weights from a different pretraining run which is now fixed!

LeVJEPA 1.5 Large is a self-supervised video encoder. It reaches a mean score of 78.76% across 22 image and video datasets, while training on ~5,300 unique video hours and 151 exaflops of compute.

For comparison, V-JEPA 1 (79.77%) used 25.7× more data and 5.6× more compute, and V-JEPA 2 (81.63%) used 321× more data and ~28.5× more compute.

LeVJEPA trains a single encoder with an invariance loss over global and local views of a clip, regularized by SIGReg (LeJEPA's Sketched Isotropic Gaussian Regularizer). It uses no EMA target encoder, stop-gradient, predictor or pixel decoder. Code: MLO-lab/LeVJEPA · Paper: arXiv 2608.27395 · Project page: levjepa.github.io

Training budget

LeVJEPA 1.5 Large V-JEPA 1 V-JEPA 2
Unique video hours ~5,300 ~136,000 ~1,700,000
Data relative to LeVJEPA 1× 25.7× 321×
Training compute (EF) 151 844 ~4,300*
Compute relative to LeVJEPA 1× 5.6× ~28.5×
Mean score, 22 datasets 78.76 79.77 81.63

* V-JEPA 2 training compute is an estimate.

Results

All three encoders are evaluated frozen with attentive probing. Bold marks the highest score in each row.

Image · 15 datasets

Dataset LeVJEPA 1.5 Large V-JEPA 1 V-JEPA 2
ImageNet-1K 75.20 79.45 82.68
STL-10 90.30 93.49 99.05
Flowers102 88.23 92.28 95.37
Aircraft 61.75 55.30 67.78
Oxford Pets 82.47 82.31 90.92
DTD 65.85 67.02 72.55
EuroSAT 98.06 97.17 96.39
DermaMNIST 88.98 83.44 83.49
BloodMNIST 99.09 98.48 98.19
BreastMNIST 87.82 88.46 84.62
RetinaMNIST 62.25 64.00 69.25
PneumoniaMNIST 91.03 94.23 89.58
OrganCMNIST 88.22 87.03 85.15
OrganSMNIST 77.61 74.10 70.21
OrganAMNIST 93.31 88.66 88.27
Mean 83.34 83.03 84.90

Video · 7 datasets

Dataset LeVJEPA 1.5 Large V-JEPA 1 V-JEPA 2
SSv2 60.10 69.55 70.00
Kinetics-710 68.23 75.27 77.01
Diving48 27.43 34.54 39.46
KTH 78.24 87.04 89.81
DynTex++ 99.03 98.33 98.33
YUP++ 75.28 79.63 82.50
COVID-BLUES 74.24 65.15 65.15
Mean 68.94 72.79 74.61

Overall · 22 datasets

LeVJEPA 1.5 Large V-JEPA 1 V-JEPA 2
Mean 78.76 79.77 81.63

Usage

The model takes ImageNet-normalized clips of 16 frames at 224×224, shaped (B, C, T, H, W), sampled at ~7.5 fps (16 frames ≈ 2 s). It returns the [cls] token followed by the patch tokens of every frame.

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("galilai-group/LeVJEPA-1.5-Large", trust_remote_code=True).eval()

video = torch.randn(1, 3, 16, 224, 224)   # (B, C, T, H, W), ImageNet-normalized
with torch.no_grad():
    out = model(pixel_values=video)

out.last_hidden_state   # (1, 3137, 1024): [cls] + 16 * 14 * 14 patch tokens, in (frame, row, column) order
out["pooler_output"]    # (1, 1024): the [cls] token

trust_remote_code=True is required because the modeling code (RoPE and block-causal attention) ships with the weights. Normalize with the ImageNet statistics: mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225].

  • Images: repeat a single image along the time axis, image.unsqueeze(2).repeat(1, 1, 16, 1, 1). This is the protocol used for the image results above.
  • Attention mode: leave config.attn_mode at "block_causal". The weights were trained causal; full attention runs without an error but returns worse features.

Without transformers, the LeVJEPA repository loads the same weights with its own code:

from legacy.encoder import load_legacy_encoder

encoder = load_legacy_encoder("galilai-group/LeVJEPA-1.5-Large", name="vit_large", num_frames=16)
cls, patches, _ = encoder(clip)   # clip (B, T, C, H, W) -> (B, 1024), (B, T, 196, 1024)

The repository also contains the V-JEPA 2 attentive-probe evaluation (python -m eval.eval model=levjepa_v1_vitl checkpoint=galilai-group/LeVJEPA-1.5-Large) and a notebook that visualizes the patch features. Make sure to utilize the newest repository version (v2) and follow the hyperparameter guidelines for best probing results.

Model details

  • Architecture: ViT-L/16 video encoder: 24 layers, width 1024, 16 heads.

    • One token per 16×16 patch per frame (tubelet size 1), with 3-D RoPE positions.
    • Block-causal attention: bidirectional within a frame, causal across frames. Each frame's tokens depend only on the current and earlier frames.
    • A [cls] token reads the whole clip.
  • Parameters: 303M.

  • Input: 16 frames at 224×224 (14×14 patches per frame), sampled at ~7.5 fps, ImageNet-normalized. RoPE also accepts other resolutions and clip lengths.

  • Evaluation protocol: frozen encoder with the V-JEPA 2 attentive probe (4-layers) trained on its patch tokens. All three models are evaluated with the same protocol tuned for them.

  • Intended use: frozen feature extraction (attentive or linear probing, retrieval, or a backbone for downstream heads). The model has no classification head.

  • License: CC BY-NC 4.0 (non-commercial). The encoder code is adapted from Meta's V-JEPA.

Citation

@misc{kuhn2026levjepaefficientscalable,
      title={LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics},
      author={Lukas Kuhn and Lucas Maes and Giuseppe Serra and Quentin Le Lidec and Yann LeCun and Randall Balestriero and Florian Buettner},
      year={2026},
      eprint={2608.27395},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.27395},
}
Downloads last month
340
Safetensors
Model size
0.3B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for galilai-group/LeVJEPA-1.5-Large