Reproduce “Latent Wavelet Diffusion For Ultra-High-Resolution Image Synthesis” 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
Luigi Sigillo, Shengfeng He, Danilo Comminiello
Latent Wavelet Diffusion (LWD) is a training framework enhancing ultra-high-resolution (2K-4K) image synthesis by integrating frequency-sensitive spatial supervision into latent diffusion models. It consistently improved image quality, for instance, reducing FID by up to 7% and LPIPS by up to 6% on 2K resolution benchmarks, while maintaining zero additional inference cost.
Key insights
- Integrating a scale-consistency loss during VAE fine-tuning is crucial for preparing a clean latent space, ensuring high-frequency activations correspond to meaningful structures rather than noise.
- Localized frequency analysis via wavelet transforms provides an interpretable, signal-driven method to identify spatial regions needing enhanced supervision, improving detail without architectural changes.
- Adaptive masking of the diffusion loss, based on local frequency energy, allows for efficient resource allocation during training, focusing on complex areas and accelerating convergence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.