clip_vit_base_16(pretrained: bool = False, overrides: object = {})Construct CLIP with a ViT-B/16 image tower.
Model Size
Parameters
pretrainedbool= FalseLoad the OpenAI WIT-400M checkpoint from the Lucid hub.
**overridesobject= {}Optional
CLIPConfig field overrides.Returns
CLIPModelBoth towers and the learned temperature.
Notes
Reference: Radford et al., ICML 2021 (arXiv:2103.00020), §2.4.
Identical to ViT-B/32 apart from the patch size, which alone quadruples the token count — the paper's evidence that resolution of the tokenisation, not width, is what the image tower was starved of.
Examples
>>> from lucid.models import clip_vit_base_16
>>> model = clip_vit_base_16()
>>> model.config.patch_size, model.config.vision_width
(16, 768)