Reproduce “Fréchet Wavelet Distance: A Domain-Agnostic Metric for Image Generation” 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
Lokesh Veeramacheneni, Moritz Wolter, Hildegard Kuehne, Juergen Gall
Researchers at the University of Bonn and the University of Tuebingen introduce the Fréchet Wavelet Distance (FWD), a domain-agnostic metric for image generation quality that leverages the Wavelet Packet Transform. This new metric offers enhanced interpretability, is approximately 36 times more computationally efficient than FD-DINOv2, and demonstrates superior alignment with human perception compared to existing metrics, particularly on diverse and out-of-domain datasets.
Key insights
- Leveraging the Wavelet Packet Transform provides a domain-agnostic approach to image representation, bypassing the need for pre-trained, domain-specific deep neural networks and reducing inherent biases.
- Decomposing the image evaluation into distinct frequency packets offers fine-grained interpretability, allowing researchers to diagnose generative model strengths and weaknesses in specific spatial and textural details.
- A computationally inexpensive, direct mathematical transformation can yield an effective evaluation metric that is both robust to perturbations and aligns more consistently with human perception than complex, deep-learning-based alternatives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.