LibreLeVJEPAl-embed

NON-COMMERCIAL WEIGHTS: This checkpoint is licensed under CC BY-NC 4.0. It requires attribution and may not be used for commercial purposes.

LibreLeVJEPAl-embed

LeVJEPA ViT-L/16 video encoder for 16-frame clip and patch embeddings, converted for LibreYOLO.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreLeVJEPAl-embed.pt")
result = model.predict("clip.mp4")[0]
clip_embedding = result.embeddings.data       # (1, 1024), L2-normalized
patch_embeddings = model.embed_tokens("clip.mp4")  # (1, 16, 14, 14, 1024)

Source

Converted from galilai-group/LeVJEPA-VideoMix-Large at revision e831a0347737fcaa660b39c57d41c109de399845.

The upstream model card identifies the checkpoint as CC BY-NC 4.0 and attributes LeVJEPA to Lukas Kuhn, Lucas Maes, Giuseppe Serra, Quentin Le Lidec, Yann LeCun, Randall Balestriero, and Florian Buettner.

Modifications

The learned tensors are unchanged. LibreYOLO wraps the original state dict with v1.0 checkpoint metadata describing the family, task, preprocessing, clip geometry, attention mode, and CLS pooling contract.

The native LibreYOLO inference graph was independently implemented from the paper and permissively licensed transformer primitives. Strict full-tensor parity against the pinned official checkpoint is exact on the released FP32 model (max_abs_diff == 0.0). No upstream remote Python code is included.

Embedding contract

The checkpoint consumes 16 RGB frames sampled at approximately 7.5 FPS, resized and center-cropped to 224 pixels, then normalized with ImageNet mean and standard deviation. predict() returns the final CLS token normalized to unit length. embed_tokens() returns the unnormalized time-major patch grid.

This is a feature encoder. It does not contain a detector, tracker, classification head, captioner, or LibreYOLO pretraining workflow.

License

The learned weights are Creative Commons Attribution-NonCommercial 4.0 International. Redistribution is permitted under that license; use is limited to non-commercial purposes and attribution is required. See LICENSE for the official legal code and NOTICE for source attribution and modification disclosure.

LibreYOLO's source code is separately licensed under MIT. The MIT license does not override the checkpoint's CC BY-NC 4.0 terms.