LibreVJEPA2l256-embed

LibreVJEPA2l256-embed

V-JEPA 2.0 l256 video encoder (clip embedding), converted for LibreYOLO.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreVJEPA2l256-embed.pt")
result = model.predict("clip.mp4")

Source

Converted from facebook/vjepa2-vitl-fpc64-256 at revision b3c1679b7c34d3255ef3547f27c7b226aefab26f.

Modifications

The upstream checkpoint is remapped into LibreYOLO's native V-JEPA 2 module and wrapped with LibreYOLO v1.0 checkpoint metadata (family, size, task, dataset variant, clip geometry, preprocessing and the pooling rule). Tensor values are unchanged: the conversion is a key remap, loaded strictly, and the self-supervised predictor tower is dropped as a named, asserted set rather than by substring match.

Parity against unmodified transformers==5.1.0 on float32 CPU is exact (max_abs_diff == 0.0) for both the full final token tensor and the mean-pooled, L2-normalized vector.

Embedding contract

The public embedding is a LibreYOLO pooling contract: the arithmetic mean of the final encoder tokens, L2-normalized. Upstream designates no global retrieval vector, and no retrieval benchmark is claimed for it. The native spatiotemporal token grid is available separately via model.embed_tokens(...).

V-JEPA 2 is trained on video. An image is accepted as a single-frame input, which is a static appearance representation, not a motion one.

License

These weights are MIT, inherited from the source checkpoint above. The full text is in LICENSE, with attribution in NOTICE. Licences in this family differ per artifact, so do not assume a family-wide licence.