Safe Latent Diffusion

Official Implementation of the Paper Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models

Interactive Demo

An interactive demonstration is available on huggingface and also in Colab.

Installation

You can either clone the repository and install it locally by running

git clone https://github.com/ml-research/safe-latent-diffusion.git
cd ./safe-latent-diffusion
pip install .

or install it directly from git

pip install git+https://github.com/ml-research/safe-latent-diffusion.git

Usage

This repository provides a new diffusion pipeline with SLD based on the diffusers library. The SLDPipeline extends the StableDiffusionPipeline and can therefore be loaded from a stable diffusion checkpoint like shown below.

from sld import SLDPipeline
device='cuda'

pipe = SLDPipeline.from_pretrained(
    "AIML-TUDA/stable-diffusion-safe",
).to(device)

Safety guidance is enabled by default, wherefore images generated with the pipeline above will always be guided according to the predefined safety concept.

The current safety concept can be checked and changed via

pipe.safety_concept

All sld parameters presented in the paper are available as arguments to the SLDPipeline and default to the medium hyperparameter configuration. An examplary use of the pipeline could look like this:

import torch
gen = torch.Generator(device)
prompt = 'portrait of Sickly diseased dying Samurai warrior, sun shining, photo realistic illustration by greg rutkowski, thomas kindkade, alphonse mucha, loish, norman rockwell.'
gen.manual_seed(2602096847)
out = pipe(prompt=prompt, generator=gen, guidance_scale=10,
           sld_warmup_steps=7,
           sld_guidance_scale=2000,
           sld_threshold=0.025,
           sld_momentum_scale=0.5,
           sld_mom_beta=0.7
           )
image = out.images[0]

To disable safe latent diffusion, i.e. generate the image as if using the original stable diffusion, simply set sld_guidance_scale=0.

Use in 🧨 Diffusers

Safe Latent Diffusion is fully integrated in 🧨diffusers. Loading the pipeline can simply be done like this:

from diffusers import DiffusionPipeline

device='cuda'

pipe = DiffusionPipeline.from_pretrained(
    "AIML-TUDA/stable-diffusion-safe",
).to(device)

Inappropriate Image Prompts (I2P)

The I2G benchmark presented in the paper is available on huggingface or can be loaded directly via the datasets library.

from datasets import load_dataset
data = load_dataset('AIML-TUDA/i2p', split='train')

Citation

If you like or use our work please cite us:

@inproceedings{schramowski2022safe,
      title={Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models}, 
      author={Patrick Schramowski and Manuel Brack and Björn Deiseroth and Kristian Kersting},
      year={2023},
      booktitle={Proceedings of the {IEEE} Conference on Computer Vision and Pattern Recognition ({CVPR})},
}

Above: Safe Latent diffusion applied on prompt

The most offensive image one can created with Artificial Intelligence, hate, harassment, violence, self-harm, sexual content, shocking images, illegal activity, [...]

Using Stable Diffusion v1.5 and Hyp-Strong of SLD