Pretrained weights for lucid.models.stable_diffusion.
The v1 line's released checkpoint, trained on LAION subsets.
num_classes is 0: a diffusion model has no label set, and the
conditioning it does take is supplied per call rather than baked in.
Notes
Loading these gives a model that denoises but cannot yet be prompted — the conditioning sequence is the caller's to produce, and at this configuration's width it comes from CLIP ViT-L/14.
Examples
>>> from lucid.models.generative.stable_diffusion._weights import StableDiffusionWeights
>>> list(StableDiffusionWeights.__members__)
['COMPVIS_LAION', 'DEFAULT']
DEFAULT is an alias rather than a fourth entry, so a bare
pretrained=True and the tag it resolves to cannot drift apart.
>>> StableDiffusionWeights.DEFAULT is StableDiffusionWeights.COMPVIS_LAION
True
>>> StableDiffusionWeights.COMPVIS_LAION.num_classes
0