Efficient Image Synthesis with Sphere Latent Encoder

2026-05-15

AI Summary

This work addresses key limitations in existing few-step image generation methods—namely, training instability, poor scalability, and computational inefficiency arising from frequent pixel-to-latent space conversions in Sphere Encoder, which also induces a conflict between reconstruction and generation objectives. To resolve these issues, the authors propose a decoupled framework that separates reconstruction and generation tasks by fixing a pretrained image encoder and training an independent denoising model entirely within a spherical latent space. This approach achieves, for the first time, complete decoupling of the two tasks while enabling efficient few-step generation without repeated transformations to pixel space. Experiments on AnimalFaces, Oxford-Flowers, and ImageNet-1K demonstrate that the method significantly outperforms Sphere Encoder in both generation quality and inference speed, while matching or exceeding strong baselines.

Abstract

Few-step image generation has seen rapid progress, with consistency and meanflow-based methods significantly reducing the number of sampling steps. Despite their low inference cost, these approaches often suffer from training instability and limited scalability. Sphere Encoder is a recent alternative that produces high-quality images in only a few steps; however, it requires repeated transitions between the pixel space and latent space during inference while jointly optimizing reconstruction and generation within a single architecture. This design leads to computational inefficiency and objective conflict between reconstruction and generation. To address these limitations, we decouple the framework into a fixed pretrained image encoder and a separate latent denoising model trained entirely in a spherical latent space. Our approach eliminates repeated pixel-space operations during training and inference, improving efficiency and allowing reconstruction and generation to specialize independently. On Animal-Faces, Oxford-Flowers and ImageNet-1K datasets, our method significantly outperforms Sphere Encoder in both generation quality and inference speed, while achieving competitive results against strong few-step and multi-step baselines.

Problem

• few-step image generation
• Sphere Encoder
• latent space
• reconstruction-generation conflict
• computational inefficiency

Innovation

• sphere latent space
• decoupled framework
• few-step image generation
• latent denoising model
• image encoder