lucid-dl/vjepa2-ac-vitg

V-JEPA 2-AC ViT-g/16

https://arxiv.org/abs/2506.09985

Lucid port of https://dl.fbaipublicfiles.com/vjepa2/vjepa2-ac-vitg.pt, converted to Lucid-native safetensors.

Available weights

Tag	Params	GFLOPs	Size	Source
OFFICIAL (default)	1317.4M	—	5025.54 MB	vjepa2/vjepa2-ac-vitg.pt

Usage

import lucid
import lucid.models as models
from lucid.models.weights import VJEPA2ACWeights

# default tag
model = models.vjepa2_ac_vit_giant(pretrained=True)

# explicit tag (enum or string)
model = models.vjepa2_ac_vit_giant(weights=VJEPA2ACWeights.OFFICIAL)
model = models.vjepa2_ac_vit_giant(pretrained="OFFICIAL")

# preprocessing travels with the weights
weights = VJEPA2ACWeights.OFFICIAL
preprocess = weights.transforms()
# a clip, plus one action and one state per transition
video = lucid.rand(1, 64, 3, 256, 256)
actions = lucid.rand(1, 63, 7)
states = lucid.rand(1, 63, 7)
out = model(video, actions, states)
latents = out.prediction  # (B, steps * tokens_per_step, dim)

Conversion

Converted from https://dl.fbaipublicfiles.com/vjepa2/vjepa2-ac-vitg.pt via python -m tools.convert_weights vjepa2_ac_vit_giant --tag OFFICIAL. Key set, tensor shapes and a strict load verified against a freshly built Lucid model.

License

mit — inherited from the original weights.

Citation

Assran et al., "V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning," arXiv:2506.09985, 2025.