Abstract: "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."
Applicable Models:
- Stable Diffusion
Authors: Patrick Schramowski, Manuel Brack, Björn Deiseroth, Kristian Kersting
Considerations: What is considered appropriate and inappropriate varies strongly across cultures and is very context dependent
Datasets: https://huggingface.co/datasets/AIML-TUDA/i2p
Group: CulturalEvals
Hashtags: .nan
Link: 'Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models'
Modality: Image
Screenshots:
- Images/SLD1.png
- Images/SLD2.png
Suggested Evaluation: Evaluating text-to-image models for safety
Level: Output
URL: https://arxiv.org/pdf/2211.05105.pdf
What it is evaluating: Generating images for diverse set of prompts (novel I2P benchmark) and investigating how often e.g. violent/nude images will be generated. There is a distinction between implicit and explicit safety, i.e. unsafe results with “normal” prompts.
Metrics: .nan