# LeVJEPA

> Source: https://aiwiki.ai/wiki/levjepa
> Updated: 2026-08-28
> Fact-checked: 2026-08-28
> Categories: AI Research, Computer Vision, Machine Learning
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "LeVJEPA." aiwiki.ai, 28 Aug 2026. https://aiwiki.ai/wiki/levjepa
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**LeVJEPA** is a [self-supervised learning](https://aiwiki.ai/wiki/self-supervised_learning) method for training video encoders. It applies the LeJEPA objective to video, combining an invariance loss with the Sketched Isotropic Gaussian Regularizer (SIGReg) in a shared encoder. Unlike V-JEPA-style training, the method does not use a predictor, a stop-gradient operation, or a second encoder as a training target.[1][2]

Lukas Kuhn, Lucas Maes, Giuseppe Serra, Quentin Le Lidec, Yann LeCun, Randall Balestriero, and Florian Buettner introduced LeVJEPA in an arXiv preprint submitted on August 27, 2026. The authors released training code and a pretrained checkpoint alongside the paper. As of August 28, 2026, its evaluation results had not been independently reproduced or peer reviewed.[1][3][4]

LeVJEPA is distinct from LeVLJEPA, a separate vision-language pretraining method by some of the same researchers. LeVJEPA trains representations from video alone.[1][8]

## Method

Training begins with a 16-frame video clip. The method constructs one global view at full spatial resolution and several local views made with spatial crops and photometric augmentation. All views cover the same temporal window. One shared [Vision Transformer](https://aiwiki.ai/wiki/vision_transformer) processes them, and a learned `[cls]` token supplies a clip-level representation to a small projector.[1][2]

The objective has two terms. A mean-squared-error invariance term draws each local-view embedding toward the global-view embedding. Gradients pass through both sides of this comparison, so the invariance term alone permits the trivial solution in which every input receives the same embedding. SIGReg counters that outcome by testing random one-dimensional projections of a batch of embeddings against a standard Gaussian distribution. The authors use the LeJEPA default weight of 0.02 for SIGReg in every reported experiment.[1][5]

The theoretical collapse guarantee belongs to the LeJEPA objective under its stated assumptions. It does not guarantee that every implementation, optimization run, or downstream application will perform correctly. LeVJEPA also retains an exponential moving average of encoder weights for evaluation, but that copy does not provide training targets and does not participate in the loss.[1][5]

### Tokenization and attention

The default encoder divides each frame into 16 by 16 spatial patches without combining adjacent frames at the input. During pretraining, it uniformly discards 95 percent of patch tokens after patch embedding. The remaining tokens are the encoder's only view of the clip; token dropping is not used during downstream evaluation or feature extraction.[1][3]

Patch tokens attend bidirectionally within one frame and causally across frames. A frame representation can therefore depend on the current and preceding frames, but not on later frames. The `[cls]` token attends to the full clip. In the paper's ImageNet-1K frozen-probe ablation, block-causal attention scored 51.2 percent top-1 accuracy and fully bidirectional attention scored 50.7 percent under the same listed configuration. This author-reported comparison supports no measurable loss in that experiment, not a universal equivalence across tasks.[1]

## Pretraining data and evaluation

The controlled comparisons use a class-balanced 20 percent subset of K710, the union of the Kinetics-400, Kinetics-600, and Kinetics-700 training sets with validation overlap removed. The authors retrained comparison methods from their official implementations for 240 epochs at an effective batch size of 3,072. They evaluated frozen encoders with attentive probes on [ImageNet](https://aiwiki.ai/wiki/imagenet) and Something-Something-v2, and with a mean-pooled linear probe on Kinetics-400.[1][7]

Those probes test whether a frozen representation supports classification after a small readout is trained. They do not measure an end-to-end deployed video system. The Kinetics-400 result also uses a different, weaker readout than the attentive probes used for the other two reported tasks.[1]

### Matched-epoch comparison

In the matched-epoch comparison, LeVJEPA and [V-JEPA 2](https://aiwiki.ai/wiki/v_jepa_2) see the same data for the same 240 epochs and use the same effective batch size, but they do not consume equal compute. Across ViT-S, ViT-B, and ViT-L, the authors report that LeVJEPA produces comparable or higher ImageNet-1K attentive-probe accuracy with 5.6 to 20.8 times less total pretraining compute. At ViT-B, the methods differ by less than one accuracy point while LeVJEPA uses 4.8 ExaFLOPs and V-JEPA 2 uses 36.4 ExaFLOPs. At ViT-L, LeVJEPA scores 1.9 points higher while using 5.6 times less compute.[1][6]

These figures compare pretraining FLOPs under the paper's implementation and accounting protocol. They are not measurements of inference latency, energy use, monetary cost, or performance on arbitrary hardware.

### Matched-FLOP comparison

The matched-FLOP experiment holds total pretraining compute constant for ViT-B encoders on the same K710 subset. Because each LeVJEPA training sample costs fewer FLOPs, its equal-compute run lasts 1,085 epochs and uses 10 local views, rather than the 240-epoch, four-view setup used in the matched-epoch comparison.[1]

| Method | ImageNet-1K attentive probe | Something-Something-v2 attentive probe | Kinetics-400 linear probe |
|---|---:|---:|---:|
| VideoMAEv2 | 53.4 | 43.6 | 37.4 |
| V-JEPA 2 | 51.6 | 42.5 | 40.7 |
| LeVJEPA | 61.0 | 40.4 | 44.6 |

Under this protocol, the authors report that LeVJEPA leads the strongest listed video baseline on ImageNet-1K by 7.6 points and has the highest Kinetics-400 result. On Something-Something-v2, it trails VideoMAEv2 by 3.2 points. The table is an author-run preprint comparison, not an independent benchmark.[1]

The paper also compares LeVJEPA with [DINOv2](https://aiwiki.ai/wiki/dinov2) at equal total FLOPs using frames from the same source videos. DINOv2 scores 53.8 against LeVJEPA's 50.7 on ImageNet-1K, while LeVJEPA scores 30.4 against DINOv2's 16.9 on Something-Something-v2. The authors interpret the latter as evidence that video pretraining retains more motion information under their protocol.[1]

## Ablations and scaling

The paper reports that ImageNet-1K attentive-probe accuracy rises from 33.9 percent with no token dropping to 47.6 percent at 95 percent dropping in one ViT-B sweep. Motion-focused results behave differently: high dropping ratios reduce Something-Something-v2 accuracy under short schedules, although longer training recovers much of the difference in the reported experiments. Uniform random dropping also scores 50.7 on ImageNet-1K against 39.6 for tube dropping in a separate comparison.[1]

At a matched token budget, per-frame tokenization scores 50.7 on ImageNet-1K and 30.4 on Something-Something-v2, compared with 47.4 and 28.8 for input patches that aggregate two frames. These results support the authors' default for the tested tasks, but do not show that temporal aggregation is unnecessary for every video application.[1]

For a larger-data experiment, the authors train a ViT-L/16 for 100 epochs on K710, Something-Something-v2, Walking Tours, and the PE Video Dataset. The arXiv paper reports 69.5 percent ImageNet-1K and 55.0 percent Something-Something-v2 top-1 accuracy under frozen attentive probing. This run changes both the dataset and schedule, so it is evidence that the reported model benefits from more data, not a controlled estimate of the contribution of any one source.[1]

The authors also report a smaller experiment in which a ViT-Tiny trained for 12 hours on one 16 GB RTX 5080 using eight Walking Tours videos improves from 8.9 to 25.2 percent ImageNet-1K frozen accuracy. That result demonstrates one reported training configuration; it does not establish a minimum hardware requirement or expected performance for other datasets.[1][3]

## Availability

The official GitHub repository contains the training implementation, configurations, dataset preparation scripts, and notebooks. Most repository code is MIT licensed, while `module.py`, which adapts Meta's V-JEPA implementation, remains under CC BY-NC 4.0.[3]

The authors also released `galilai-group/LeVJEPA-VideoMix-Large`, a 303-million-parameter ViT-L/16 checkpoint trained on what its model card calls 1.8 million VideoMix clips. The checkpoint is a feature encoder without a classification head and is licensed CC BY-NC 4.0. Its Hugging Face loading example requires `trust_remote_code=True`, which executes modeling code supplied with the repository and should be reviewed under the user's normal software-security process.[3][4]

Public code and weights make the method inspectable and support reproduction attempts. Their availability is not evidence that the published results have already been independently replicated.

## Limitations

LeVJEPA was evaluated in a first-version preprint on a restricted controlled corpus and at model sizes no larger than ViT-L. The paper identifies behavior at larger data, model, and batch scales as unresolved. It also does not evaluate dense prediction tasks such as segmentation or tracking, despite presenting qualitative visualizations of organized patch features.[1]

Aggressive token dropping has a task-dependent tradeoff, with weaker short-schedule motion results. The block-causal ablation covers one frozen ImageNet probe rather than streaming latency, online adaptation, action prediction, or robotics control. None of the reported classification probes establishes production reliability, real-time performance, or safety.[1]

## References

1. Lukas Kuhn, Lucas Maes, Giuseppe Serra, Quentin Le Lidec, Yann LeCun, Randall Balestriero, and Florian Buettner. "LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics." arXiv:2608.27395v1, August 27, 2026. https://arxiv.org/abs/2608.27395
2. LeVJEPA authors. "LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics." Project page, accessed August 28, 2026. https://levjepa.github.io/
3. Galilai Group. "galilai-group/levjepa." Official GitHub repository, accessed August 28, 2026. https://github.com/galilai-group/levjepa
4. Galilai Group. "LeVJEPA-VideoMix-Large." Hugging Face model repository, accessed August 28, 2026. https://huggingface.co/galilai-group/LeVJEPA-VideoMix-Large
5. Randall Balestriero and Yann LeCun. "LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics." arXiv:2511.08544, November 2025. https://arxiv.org/abs/2511.08544
6. Adrien Bardes et al. "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning." arXiv:2506.09985, June 2025. https://arxiv.org/abs/2506.09985
7. Will Kay et al. "The Kinetics Human Action Video Dataset." arXiv:1705.06950, May 2017. https://arxiv.org/abs/1705.06950
8. Lukas Kuhn, Giuseppe Serra, Randall Balestriero, and Florian Buettner. "LeVLJEPA: End-to-End Vision-Language Pretraining Without Negatives." arXiv:2607.00784, July 2026. https://arxiv.org/abs/2607.00784