Sphere Encoder 2 hits pixel-space SOTA image quality with single-pass generation rivaling diffusion

tomgoldsteincs · x · 2026-11-06

Tom Goldstein's team released Sphere Encoder 2 (arXiv:2610.02208), a faster diffusion alternative that generates images in a single forward pass by mapping 100k-dimensional random latent vectors directly to natural images.

Key fixes over v1:

Random points cluster near the equator of the latent sphere, but training rotations never reach that region, leaving a gap that limited one-step generation;
Pixel-wise reconstruction loss pushed the decoder to average over plausible images, producing blurry, low-detail outputs.

The authors report SOTA image quality (FDr6) on ImageNet among all known pixel-space generators, including diffusion models, while keeping autoencoder speed and simplicity. Models are open-sourced.

Original post →

Sphere Encoder 2 hits pixel-space SOTA image quality with single-pass generation rivaling diffusion

More from Multimodal