stable_diffusion(pretrained: bool = False, overrides: object = {})Construct the v1 architecture — 512 pixels, CLIP ViT-L/14 width.
Model Size
Parameters
pretrainedbool= FalseLoad the released v1 checkpoint from the Lucid hub.
**overridesobject= {}Optional
StableDiffusionConfig field overrides.Returns
StableDiffusionModelAutoencoder, conditional U-Net and sampler.
Notes
Reference: Rombach et al., CVPR 2022 (arXiv:2112.10752).
cross_attention_dim is 768, which is exactly
lucid.models.clip_vit_large_14's text width — that tower is
what produced the conditioning these settings were trained against.
Examples
>>> from lucid.models import stable_diffusion
>>> model = stable_diffusion()
That call materialises 943M parameters — about three gigabytes — and
what it will build is a property of the configuration, so the shape
is read there instead:
>>> from lucid.models.generative.stable_diffusion import (
... StableDiffusionConfig)
>>> config = StableDiffusionConfig()
>>> config.latent_size, config.cross_attention_dim
(64, 768)