Image Generation with a Sphere Encoder

2026/02/16 by Kaiyu Yue, Menglin Jia, Ji Hou +1 · 1 voice
Computer Science · #cs.CV

paper · pdf · doi:10.48550/arxiv.2602.15030

Abstract

We introduce the Sphere Encoder, an efficient generative framework capable of producing images in a single forward pass and competing with many-step diffusion models using fewer than five steps. Our approach works by learning an encoder that maps natural images uniformly onto a spherical latent space, and a decoder that maps random latent vectors back to the image space. Trained solely through image reconstruction losses, the model generates an image by simply decoding a random point on the sphere. Our architecture naturally supports conditional generation, and looping the encoder/decoder a few times can further enhance image quality. Across several datasets, the sphere encoder approach yields performance competitive with state of the art diffusions, but with a small fraction of the inference cost. Project page is available at https://sphere-encoder.github.io .

Citations

• PixelDiT: Pixel Diffusion Transformers for Image Generation
• Hyperspherical Latents Improve Continuous-Token Autoregressive Generation
• FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
• Mean Flows for One-step Generative Modeling
• Wan: Open and Advanced Large-Scale Video Generative Models
• Fractal Generative Models
• Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models
• LTX-Video: Realtime Video Latent Diffusion
• JetFormer: An Autoregressive Generative Model of Raw Images and Text
• Simpler Diffusion (SiD2): 1.5 FID on ImageNet512 with pixel-space diffusion
• One Step Diffusion via Shortcut Models
• Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models
• Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think
• Detecting, Explaining, and Mitigating Memorization in Diffusion Models
• Consistency Flow Matching: Defining Straight Flows with Velocity Consistency
• Consistency Models Made Easy
• Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction
• Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
• Rethinking FID: Towards a Better Evaluation Metric for Image Generation
• One-step Diffusion with Distribution Matching Distillation
• Adversarial Diffusion Distillation
• SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis
• Understanding and Mitigating Copying in Diffusion Models
• MEGABYTE: Predicting Million-byte Sequences with Multiscale Transformers
• Scaling up GANs for Text-to-Image Synthesis
• Consistency Models
• Simple diffusion: End-to-end diffusion for high resolution images
• Scalable Diffusion Models with Transformers
• Flow Matching for Generative Modeling
• Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
• Classifier-Free Diffusion Guidance
• General-purpose, long-context autoregressive modeling with Perceiver AR
• StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets
• High-Resolution Image Synthesis with Latent Diffusion Models
• High-Resolution Image Synthesis with Latent Diffusion Models
• Diffusion Models Beat GANs on Image Synthesis
• MLP-Mixer: An all-MLP Architecture for Vision
• RoFormer: Enhanced Transformer with Rotary Position Embedding
• Learning Transferable Visual Models From Natural Language Supervision
• Improved Denoising Diffusion Probabilistic Models
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• A Contrastive Learning Approach for Training Variational Autoencoder Priors
• Denoising Diffusion Implicit Models
• Denoising Diffusion Probabilistic Models
• The Power Spherical distribution
• Training Generative Adversarial Networks with Limited Data
• Latent Variables on Spheres for Autoencoders in High Dimensions
• StarGAN v2: Diverse Image Synthesis for Multiple Domains
• Root Mean Square Layer Normalization
• An Introduction to Variational Autoencoders
• Generating Long Sequences with Sparse Transformers
• Max-Sliced Wasserstein Distance and its use for GANs
• Sliced Wasserstein Generative Models
• Diagnosing and Enhancing VAE Models
• Orthogonal Estimation of Wasserstein Distances
• A Style-Based Generator Architecture for Generative Adversarial Networks
• A Style-Based Generator Architecture for Generative Adversarial Networks
• Taming VAEs
• Large Scale GAN Training for High Fidelity Natural Image Synthesis
• Spherical Latent Spaces for Stable Variational Autoencoders
• Glow: Generative Flow with Invertible 1x1 Convolutions
• Sliced-Wasserstein Autoencoder: An Embarrassingly Simple Generative Model
• The Unreasonable Effectiveness of Deep Features as a Perceptual Metric
• Wasserstein Auto-Encoders
• Progressive Growing of GANs for Improved Quality, Stability, and Variation
• FiLM: Visual Reasoning with a General Conditioning Layer
• Attention Is All You Need
• VAE with a VampPrior
• StackGAN: Text to Photo-realistic Image Synthesis with Stacked\n Generative Adversarial Networks
• Conditional Image Generation with PixelCNN Decoders
• Improved Techniques for Training GANs
• Perceptual Losses for Real-Time Style Transfer and Super-Resolution
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
• Adversarial Autoencoders
• Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks
• Fast R-CNN
• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• ImageNet Large Scale Visual Recognition Challenge
• ImageNet Large Scale Visual Recognition Challenge
• Soft-Input Soft-Output Single Tree-Search Sphere Decoding
• Auto-association by multilayer perceptrons and singular value decomposition

Discussions

• the project sphere-encoder.github.io was brought to my attention by Matt DesLauriers, I used Gemini to implement the paper and create a #WebGPU replacement for ONNX.js resulting in a lightweight infer [bsky, 0 points, 0 comments]

Related

• Diffusion Autoencoders: Toward a Meaningful and Decodable Representation
• DiffuseHigh: Training-free Progressive High-Resolution Image Synthesis through Structure Guidance
• Marigold: Affordable Adaptation of Diffusion-Based Image Generators for Image Analysis
• Sample what you cant compress
• Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation