Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models

By Patrick Schramowski, Manuel Brack, Bjorn Deiseroth, Kristian Kersting. Computer Vision and Pattern Recognition, 2022

584 citations recorded by the source corpus.

Text-conditioned image generation models have recently achieved astonishing results in image quality and text alignment and are consequently employed in a fast-growing number of applications. Since they are highly data-driven, relying on billion-sized datasets randomly scraped from the internet, they also suffer, as we demonstrate, from degenerated and biased human behavior. In turn, they may even reinforce such biases. To help combat these undesired side effects, we present safe latent diffusion (SLD). Specifically, to measure the inappropriate degeneration due to unfiltered and imbalanced training sets, we establish a novel image generation test bed-inappropriate image prompts (I2P)-containing dedicated, real-world image-to-text prompts covering concepts such as nudity and violence. As our exhaustive empirical evaluation demonstrates, the introduced SLD removes and suppresses inappropriate image parts during the diffusion process, with no additional training required and no adverse effect on overall image quality or text alignment.11Code available at https://huggingface.co/docs/diffusers/api/pipelines/stable.diffusion.safe

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/0231f2aed9a96cb516242fb57f2cb63f5651c4d8

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Unified Concept Editing in Diffusion Models (2023, methodology): “For all the models, 4703 images are generated using I2P prompts from [35].”
MACE: Mass Concept Erasure in Diffusion Models (2024, methodology): “To evaluate efﬁcacy and generality, we use each ﬁnetuned model to generate images using all 4,703 prompts sourced from the Inappropriate Image Prompt (I2P) dataset [59].”
Text-to-image Diffusion Models in Generative AI: A Survey (2023, background): “[88] finds a large amount of inappropriate content in the generated images by Stable diffusion [17] (e.”
SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (2023, methodology): “Instead, it was initially termed as ‘learning to forget’ or ‘concept erasing’ (Schramowski et al., 2023; Gandikota et al., 2023; Zhang et al., 2023a).”
Erasing Concepts from Diffusion Models (2023, methodology): “We study the effectiveness of our method with both inference method (SLD [38]) and ﬁltered re-training methods (SD V2.”
Forget-Me-Not: Learning to Forget in Text-to-Image Diffusion Models (2023, unknown): “SLD [42] and ESD [17] utilize the classiﬁer-free guidance of a pretrained model to steer prediction to the opposite guidance direction.”
Stable Bias: Evaluating Societal Representations in Diffusion Models (2023, unknown): “…[8, 4, 63], as well as reducing the bias of images generated by Stable Diffusion models by exploring the latent space of the model and guiding image generation along different axes to make generations more representative [14, 55], although the generalizability of these approaches is still unclear.”
Ablating Concepts in Text-to-Image Diffusion Models (2023, unknown): “Recent work of Schramowski et al. [58] modify the inference process to prevent certain concepts from being generated.”