Can video-based learning do without massive resources?

LeVJEPA 1.5 Large reaches 78.76% on 22 image and video benchmarks, with 321 times less data and 28.5 times less compute than V-JEPA 2.

Training models capable of understanding images and videos often relies on considerable amounts of data and compute. A new version of LeVJEPA, developed by a research team including Lukas Kuhn, Yann LeCun, and Randall Balestriero, suggests that a lighter approach can achieve results close to models trained with far greater resources. Across 22 visual evaluations, LeVJEPA 1.5 Large reaches an average score of 78.76%, compared with 79.77% for V-JEPA 1 and 81.63% for V-JEPA 2.

The difference lies mainly in the resources used. Where V-JEPA 2 uses about 1.7 million hours of unique video, this new version was trained on roughly 5,300 hours, or 321 times less. Its compute budget reaches 151 cumulative exaFLOPs, compared with about 4,300 for V-JEPA 2, an estimate that represents a reduction factor of 28.5. The results are detailed in the LeVJEPA 1.5 Large technical report, as an extension of the scientific work published on arXiv. Learning to understand a video without reconstructing it

The method is based on self-supervised learning, which leverages videos without requiring human annotations for each scene. Rather than asking the system to reconstruct missing images or explicitly predict representations produced by a second network, the researchers use a single encoder tasked with learning features common to different views of the same sequence.

Each video clip is observed through a global view and several local crops. Training encourages their numerical representations to remain consistent, despite differences in framing and appearance. A regularization technique called SIGReg prevents the system from producing identical representations for all inputs, a phenomenon known as representational collapse.

This architecture avoids several components traditionally used in JEPA-type methods, notably the target encoder and the predictor network. The researchers also showed that it was possible to randomly drop 95% of visual tokens during training, greatly reducing the operations required while preserving the information useful for learning.

The model also uses block-causal attention, which allows each frame to use information from previous frames without looking at subsequent ones. This setup is useful for processing continuous video streams and for future systems capable of progressively building a representation of their environment. Close results on images, farther behind on video

The scale-up introduced with version 1.5 Large is based on a ViT-L encoder with 303 million parameters, trained from 1.8 million video clips from sources including Kinetics-710, Something-Something v2, Walking Tours, and PE-Video. The evaluations use an encoder whose parameters remain frozen, paired with a lightweight module tasked with leveraging the learned features.

Across the fifteen image benchmarks, LeVJEPA 1.5 Large achieves an average of 83.34%, slightly higher than V-JEPA 1’s 83.03% and relatively close to V-JEPA 2’s 84.90%. Several results surpass those of the two reference models, especially on EuroSAT, with 98.06%, and BloodMNIST, with 99.09%.

The comparison is less favorable on the seven video evaluations. The new model reaches 68.94% on average, compared with 72.79% for V-JEPA 1 and 74.61% for V-JEPA 2. On Something-Something v2, focused on action and interaction understanding, it scores 60.10%, while V-JEPA 2 reaches 70%. The gap remains significant on Kinetics-710 as well, with 68.23% versus 77.01%.

These results show that resource savings do not come with complete parity in capabilities. Version 1.5 Large appears particularly competitive on image recognition tasks, whereas the V-JEPA models retain a clearer advantage on several temporal understanding tasks. A reduction in the cost of video learning

The value of this approach goes beyond score comparisons. The experiments described by the team show that, with identical data volume and number of epochs, the method can match or exceed V-JEPA 2 depending on encoder size, with compute consumption reduced by a factor ranging from 5.6 to 20.8.

In another experiment, a small ViT-Tiny model was trained for twelve hours on a single RTX 5080 with 16 GB of memory, using eight unannotated urban walking videos. The researchers report an increase from 8.9% to 25.2% on ImageNet-1K, according to their evaluation protocol. This result illustrates the possibility of experimenting with video learning on less expensive hardware, even if the performance of such a model remains far from that of larger architectures.

The optimization also affects memory. By retaining only a small proportion of visual information at each training step, the researchers reduce the load on the encoder, without applying this pruning when it is used at inference time. The system then exploits the full information in each sequence.

These technical choices open a path for applications that need to extract visual features from videos, such as action recognition, content retrieval, or scene analysis. LeVJEPA remains, however, a representation encoder, not a video generator or an autonomous physical simulator. Accessible weights, but reserved for non-commercial use

The LeVJEPA 1.5 Large weights are available on Hugging Face, while the project’s GitHub repository provides the code and training instructions. The model accepts sequences of sixteen frames at a resolution of 224 × 224, sampled at about 7.5 frames per second, or a little over two seconds of video.

Access to the model does not, however, mean freedom for commercial use. The weights are distributed under the CC BY-NC 4.0 license, which permits reuse in a non-commercial setting subject to compliance with attribution conditions. The code repository also includes separate license terms depending on the components.

With this version, the researchers document an improvement in the efficiency of video learning without claiming first place across all evaluations. The reduction in required resources is their main result, while the persistent gaps on temporal tasks make