Reproduce “Hyperspherical Autoencoder for High-Fidelity Image Reconstruction and 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
Hun Chang, Byunghee Cha, Jong Chul Ye
HAE preserves local detail, aligns pretrained features by direction while allowing magnitudes to vary, and generates along spherical latent geometry. In a dog example, reconstruction stays similar across magnitude changes but is damaged by a direction perturbation. On ImageNet at 256 × 256, it reports rFID 0.78 and PSNR 25.20 dB; LightningDiT-XL reaches gFID 1.96 without guidance after 550 epochs, with reconstruction trade-offs.
Key insights
- Semantic and structural information within SSL-based latent features is predominantly encoded in their directional component, confirming the intrinsic hyperspherical geometry of these representations.
- Directional feature alignment effectively transfers semantic priors from VFMs while mitigating the optimization conflict between semantic preservation and pixel-level detail, which is common with rigid MSE objectives.
- Stochastic latent smoothing during adversarial adaptation serves as a crucial regularizer, bridging the gap between exact reconstruction and robust sampling for downstream generative tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.