Reproduce “Enhancing AI Face Realism: Cost-Efficient Quality Improvement in Distilled Diffusion Models with a Fully Synthetic Dataset” 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
Jakub Wasala, Bartlomiej Wrzalski, Kornelia Noculak, Yuliia Tarasenko, Oliwer Krupa, Jan Kocon, Grzegorz Chodak
A method for cost-efficiently enhancing AI face realism in distilled diffusion models is introduced, leveraging a fully synthetic dataset to train an image-to-image enhancement layer. This approach reduced computational cost by 82% compared to full-scale models while achieving comparable perceptual quality.
Key insights
- The unsupervised non-pairwise ESA-CycleGAN approach outperformed the supervised U-Net architecture, effectively reducing artifacts and proving more robust to minor inconsistencies in the pairwise dataset.
- The I2I enhancement head introduced minimal computational overhead, ensuring the combined pipeline remained significantly faster than the full-scale, high-quality generative model.
- The novel `FID_diff` metric (Fréchet Inception Distance difference) accurately assessed the perceptual quality improvement by quantifying how much closer enhanced images were to high-quality targets compared to low-quality inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.