SD1.4 nudity SAEs

TopK sparse autoencoders trained on activations of three cross-attention blocks in the Stable Diffusion v1.4 UNet (revision 133a221b8aa7292a167afc5127cb63fb5005638b), used for nudity removal.

Layer	Hook	d_in	Latents (32×)	k	Step	File
L1	`up_blocks.1.attentions.1`	1280	40960	256	37500	`sd14_sae_v2_layer1/sae_sd14_v2/sd14_v2_l1_topk_k256_5759626/step_37500/up_blocks.1.attentions.1/sae.safetensors`
L2	`up_blocks.2.attentions.1`	640	20480	128	150000	`sd14_sae_v2_layer2/sae_sd14_v2/sd14_v2_l2_topk_k128_5759627/step_150000/up_blocks.2.attentions.1/sae.safetensors`
L3	`up_blocks.3.attentions.1`	320	10240	64	600000	`sd14_sae_v2_layer3/sae_sd14_v2/sd14_v2_l3_topk_k64_5760499/step_600000/up_blocks.3.attentions.1/sae.safetensors`

Each hook directory also holds the cfg.json the SAE was saved with. Optimizer state and training activations are not included.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for LukiLenkiewicz/sd14-nudity-sae

Base model

CompVis/stable-diffusion-v1-4

Finetuned

(877)

this model