vjepa2_ac_vit_giant(pretrained: bool | str = False, weights: VJEPA2ACWeights | None = None, overrides: object = {})Construct the official V-JEPA 2-AC ViT-g action model.
Model Size
Parameters
pretrainedbool or str= FalseTrue loads the Lucid-hosted official action checkpoint; a string
selects a declared weight tag explicitly.Explicit weight enum member; takes precedence over
pretrained.**overridesobject= {}Optional
VJEPA2ACConfig field overrides.Returns
VJEPA2ACModelThe ViT-g/16 latent encoder and action-conditioned predictor.
Notes
Reference: Assran et al., arXiv:2506.09985, 2025. The release publishes
one action-conditioned ViT-g checkpoint; smaller variants are
intentionally not registered. One action-model step is one frame — the
released loop repeats each frame to fill the encoder's tubelet — so a
clip of T frames takes T - 1 conditioning rows. Checked against
the official implementation with that checkpoint: encoder 8.5e-5,
predictor 4.5e-5 relative.
Examples
>>> from lucid.models import AutoConfig
>>> config = AutoConfig.from_pretrained("vjepa2_ac_vit_giant")
>>> config.encoder_dim, config.encoder_depth, config.encoder_heads
(1408, 40, 22)
>>> config.predictor_dim, config.predictor_depth, config.predictor_heads
(1024, 24, 16)
>>> config.tokens_per_step
256