Sphere Encoder 2

2026-10-01

AI Summary

This work addresses the blurriness and loss of fine details in images generated by conventional autoencoders, which stem from uneven latent space distributions and pixel-level reconstruction losses. To overcome these limitations, we propose an optimized autoencoding framework built upon a high-dimensional latent hypersphere. Methodologically, we refine the latent hypersphere sampling mechanism to eliminate generation gaps near the equatorial region. By integrating stochastic point decoding with an optimized single-step generation training strategy, our approach effectively mitigates feature averaging effects while preserving high-frequency details. This framework significantly enhances the clarity and quality of single-step image generation while retaining the computational efficiency and architectural simplicity inherent to autoencoders. The source code has been made publicly available.

Abstract

Sphere Encoder is an autoencoder that generates images by decoding random points from a high-dimensional latent sphere. We identify two limitations of the original formulation that reduce its generation quality. First, random points concentrate near the equator relative to the pole on an encoded latent, but the training rotation never reaches this region, leaving a gap that limits one-step generation. Second, training for generation with pixel-wise reconstruction loss encourages the decoder to average over plausible images, producing blurry images that lack high-frequency details. We present Sphere Encoder 2 to address both limitations, substantially improving image generation quality while maintaining the speed and simplicity of a autoencoder. Models are released at \href{https://github.com/kaiyuyue/sphere2}{github.com/kaiyuyue/sphere2}.

Problem

• Sphere Encoder
• image generation
• latent sphere
• pixel-wise reconstruction loss
• autoencoder

Innovation

• Sphere Encoder 2
• autoencoder
• latent sphere
• image generation
• reconstruction loss