sungjuncho/fade-unlearncanvas-retain-seed2

FADE UnlearnCanvas -- Retain-only oracle (seed 2)

A retain-only oracle: Stable Diffusion fine-tuned on the 47/50 UnlearnCanvas styles that exclude Monet, Picasso, Van_Gogh (seed 2 of 3 independently-trained retain seeds). This is the reference model FADE compares an unlearned model against -- it represents 'what the model should look like if it had never seen the forget styles,' with no unlearning algorithm involved.

Files

File	Description
retain_seed2.safetensors	retain-only, seed 2 (47/50 styles, excludes Monet/Picasso/Van_Gogh)

About FADE

FADE (Functional Alignment for Distributional Equivalence) is a metric for evaluating machine unlearning in diffusion models: it generates images from an unlearned model and a retain-only oracle, then measures a variational upper bound on the KL divergence between the two models' output distributions on those images. See the project repo for the full method and evaluation code: the project's public GitHub repo (link forthcoming)

This checkpoint is a fine-tune of Stable Diffusion v1.5, built on the UnlearnCanvas dataset/benchmark (Zhang et al., 2024). It inherits Stable Diffusion's CreativeML OpenRAIL-M license -- see the license file for the specific use-based restrictions that apply.

Citation

@article{cho2025referencespecific,
  title={Reference-Specific Unlearning Metrics Can Hide the Truth: A Reality Check},
  author={Cho, Sungjun and Hwang, Dasol and Sala, Frederic and Hwang, Sangheum and Cho, Kyunghyun and Cha, Sungmin},
  journal={arXiv preprint arXiv:2510.12981},
  year={2025}
}

Loading

import torch
from safetensors.torch import load_file

state_dict = load_file("<downloaded_file>.safetensors")
# merge into a CompVis-format LDM model, e.g. via UnlearnCanvas's
# ldm.util.instantiate_from_config, then:
# model.load_state_dict(state_dict, strict=False)