YAML Metadata
Warning:
empty or missing yaml metadata in repo card
(https://huggingface.co/docs/hub/model-cards#model-card-metadata)
Model zoo of Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models
Overview
This repository hosts the checkpoints for the RECE in the diffusers' format. These checkpoints can be directly loaded into your UNet if you are using diffusers.
Checkpoints
The checkpoints are named with the suffix "_ep*" indicating the epoch selected based on validation performance.
Erasure Details
Configuration Updates. Some settings have been updated from the current Arxiv version. For all concepts, the coefficients of Eq.3 are: $\lambda_1=0.1$ and $\lambda_2=0.1$. The regularization coefficients $\lambda$ and epochs are set as follows:
- Nudity and unsafe concepts(I2P concepts), $\lambda=1e-1$, with nudity for 3 epochs and unsafe concepts for 2 epochs.
- Artistic styles, $\lambda=1e-3$, 1 epoch.
- Difficult objects for UCE(e.g., church and garbage truck), $\lambda=1e-3$, 1 epoch.
- Easy objects for UCE(e.g., English Springer, golf ball and parachute), $\lambda=1e-1$, 1 epoch.
- For other objects where erasing accuracies reach 0 using UCE, RECE's further erasure is not applied.
The files' suffixes suggest the epochs we selected and the epoch sequence number starts from 0.