Sphere Encoder 2 Fixes Latent-Space Gaps and Blur in Autoencoder Image Generation

Researchers have released Sphere Encoder 2, a follow-up to the original Sphere Encoder autoencoder that generates images by decoding random points sampled from a high-dimensional latent sphere. The paper identifies two core weaknesses in the prior model: random points tend to cluster near the equator of the latent sphere rather than near the poles, leaving a coverage gap that undermines one-step generation quality; and training with a pixel-wise reconstruction loss causes the decoder to average across plausible outputs, resulting in blurry images that lack fine detail.

Sphere Encoder 2 addresses both issues, yielding substantially improved image generation quality while preserving the speed and architectural simplicity of the autoencoder approach. Code and model weights have been made publicly available alongside the paper, which was submitted to arXiv in early October 2026 by researchers including Kaiyu Yue, Sean McLeish, Ruchit Rawal, Brian Bartoldson, Menglin Jia, and Tom Goldstein.