Reproduce “Alias-Free Latent Diffusion Models: Improving Fractional Shift Equivariance of Diffusion Latent Space” 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
Yifan Zhou, Zeqi Xiao, Shuai Yang, Xingang Pan
Alias-Free Latent Diffusion Models (AF-LDM) introduce a signal-processing-inspired redesign of Latent Diffusion Models (LDMs) to achieve strong fractional shift-equivariance. This approach mitigates aliasing amplification and non-equivariant attention, leading to more consistent and stable outputs for applications like video editing and image-to-image translation, evidenced by a VAE decoder SPSNR increase to 33.88 and LDM latent SPSNR of 40.94.
Key insights
- Architectural alias-free designs alone are insufficient for LDMs; the iterative denoising and training processes can still amplify aliasing, necessitating additional regularization from an equivariance loss.
- The proposed equivariance loss is critical for effectively suppressing feature bandwidth and significantly enhancing consistency, working in synergy with the alias-free architectural modifications.
- Equivariant Attention, akin to Cross-Frame Attention, is essential for achieving shift-equivariance in LDMs and must be integrated into both the training and inference phases, not solely inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.