Reproduce “How I Met Your Bias: Investigating Bias Amplification in Diffusion Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nathan Roos, Ekaterina Iakovleva, Ani Gjergji, Vito Paolo Pastore, Enzo Tartaglione
Research empirically investigated how diffusion model sampling algorithms and their hyperparameters influence the generation of biased images, demonstrating that bias is an adaptable behavior that can be reduced by adjusting parameters like guidance scale, integration steps, and stochasticity without model retraining. This offers a training-free strategy to mitigate bias, complementing existing model-centric interventions.
Key insights
- Bias amplification in diffusion models is not a fixed characteristic of the trained model but an adaptable behavior significantly influenced by sampling parameters.
- Adjustments to sampling parameters offer a training-free approach to bias mitigation, providing a computationally efficient alternative or complement to model retraining.
- A trade-off can exist where image quality plateaus at higher integration steps while bias amplification continues to increase, suggesting opportunities for debiasing without quality compromise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.