mlsanityData to 14 Oct 2026

Hyperspherical Latents Improve Continuous-Token Autoregressive Generation

arXiv 2509.24335 ↗First posted 29 Sep 2025cs.CV cs.LGcited by 16

On the map: Diffusion models → Image generation, text-to-image and editing

Abstract (arXiv)

Autoregressive (AR) models are promising for image generation, yet continuous-token AR variants often trail latent diffusion and masked-generation models. The core issue is heterogeneous variance in VAE latents, which is amplified during AR decoding, especially under classifier-free guidance (CFG), and can cause variance collapse. We propose SphereAR to address this issue. Its core design is to constrain all AR inputs and outputs -- including after CFG -- to lie on a fixed-radius hypersphere (constant ℓ2 norm), leveraging hyperspherical VAEs. Our theoretical analysis shows that hyperspherical constraint removes the scale component (the primary cause of variance collapse), thereby stabilizing AR decoding. Empirically, on ImageNet generation, SphereAR-H (943M) sets a new state of the art for AR models, achieving FID 1.34. Even at smaller scales, SphereAR-L (479M) reaches FID 1.54 and SphereAR-B (208M) reaches 1.92, matching or surpassing much larger baselines such as MAR-H (943M, 1.55) and VAR-d30 (2B, 1.92). To our knowledge, this is the first time a pure next-token AR image generator with raster order surpasses diffusion and masked-generation models at comparable parameter scales.

Key facts. Cited by 16 papers in the map (data through 14 Oct 2026). Placed in Image generation, text-to-image and editing, one of 14 directions in Diffusion models. Builds on 52 of its 59 references that are in the map. First posted on arXiv 29 Sep 2025.

Builds on

52 of 59 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale14 Aug
  2. 2024Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction3 Apr

Other references · 50

  1. 2017Attention Is All You Needcited by 60,361
  2. 2014Adam: A Method for Stochastic Optimizationcited by 37,240
  3. 2020Language Models are Few-Shot Learnerscited by 32,419
  4. 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,863
  5. 2018The Unreasonable Effectiveness of Deep Features as a Perceptual Metriccited by 11,177
  6. 2021Diffusion Models Beat GANs on Image Synthesiscited by 8,255
  7. 2022Scalable Diffusion Models with Transformerscited by 6,054
  8. 2022Classifier-Free Diffusion Guidancecited by 5,847
  9. 2022Flow Matching for Generative Modelingcited by 5,702
  10. 2017Neural Discrete Representation Learningcited by 4,712
  11. 2021RoFormer: Enhanced Transformer with Rotary Position Embeddingcited by 4,703
  12. 2016Image-to-Image Translation with Conditional Adversarial Networkscited by 4,527
38 more
  1. 2022Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flowcited by 3,700
  2. 2022FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awarenesscited by 3,673
  3. 2020Taming Transformers for High-Resolution Image Synthesiscited by 3,253
  4. 2016Improved Techniques for Training GANscited by 2,970
  5. 2016Perceptual Losses for Real-Time Style Transfer and Super-Resolutioncited by 2,694
  6. 2020GLU Variants Improve Transformercited by 1,804
  7. 2019Root Mean Square Layer Normalizationcited by 1,480
  8. 2019Generating Diverse High-Fidelity Images with VQ-VAE-2cited by 1,387
  9. 2022MaskGIT: Masked Generative Image Transformercited by 1,099
  10. 2019Improved Precision and Recall Metric for Assessing Generative Modelscited by 944
  11. 2020On Layer Normalization in the Transformer Architecturecited by 922
  12. 2024Autoregressive Model Beats Diffusion: Llama for Scalable Image Generationcited by 777
  13. 2024SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformerscited by 737
  14. 2022Autoregressive Image Generation using Residual Quantizationcited by 722
  15. 2021Vector-quantized Image Modeling with Improved VQGANcited by 658
  16. 2024Autoregressive Image Generation without Vector Quantizationcited by 651
  17. 2023Language Model Beats Diffusion -- Tokenizer is Key to Visual Generationcited by 640
  18. 2025Mean Flows for One-step Generative Modelingcited by 561
  19. 2025Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Modelscited by 393
  20. 2024One Step Diffusion via Shortcut Modelscited by 391
  21. 2022BEiT v2: Masked Image Modeling with Vector-Quantized Visual Tokenizerscited by 303
  22. 2025MAGI-1: Autoregressive Video Generation at Scalecited by 276
  23. 2023Flow Matching on General Geometriescited by 265
  24. 2024TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generationcited by 202
  25. 2018Hyperspherical Variational Auto-Encoderscited by 189
  26. 2024Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokenscited by 147
  27. 2024Randomized Autoregressive Visual Generationcited by 127
  28. 2024Image and Video Tokenization with Binary Spherical Quantizationcited by 112
  29. 2024ImageFolder: Autoregressive Image Generation with Folded Tokenscited by 94
  30. 2024RandAR: Decoder-only Autoregressive Visual Generation in Random Orderscited by 83
  31. 2023GIVT: Generative Infinite-Vocabulary Transformerscited by 80
  32. 2024Autoregressive Speech Synthesis without Vector Quantizationcited by 65
  33. 2024Multimodal Latent Language Modeling with Next-Token Diffusioncited by 48
  34. 2025Fractal Generative Modelscited by 46
  35. 2025Diffusion Autoencoders are Scalable Image Tokenizerscited by 28
  36. 2020The Power Spherical distributioncited by 27
  37. 2025Latent Denoising Makes Good Tokenizerscited by 20
  38. 2025Unified Cross-Scale 3D Generation and Understanding via Autoregressive Modelingcited by 8

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 3; 2026: 13.

16 papers.

Sort:

By area:

  1. 2025FARMER: Flow AutoRegressive Transformer over Pixelscited by 18
  2. 2026LLaMo: Scaling Pretrained Language Models for Unified Motion Understanding and Generation with Continuous Autoregressive Tokenscited by 11
  3. 2026Image Generation with a Sphere Encodercited by 8
  4. 2025SimFlow: Simplified and End-to-End Training of Latent Normalizing Flowscited by 7
  5. 2026BitDance: Scaling Autoregressive Generative Models with Binary Tokenscited by 4
  6. 2025Visual Generation Tuningcited by 2
  7. 2026Geometric Autoencoder for Diffusion Modelscited by 2
  8. 2026Aligning Latent Geometry for Spherical Flow Matching in Image Generationcited by 2
  9. 2026Semantic-Aware Prefix Learning for Token-Efficient Image Generationcited by 1
  10. 2026Search at the Cost of Sampling: Nearly-Instant Latent Space Bayesian Optimizationcited by 1
  11. 2026Parallel Rollout Approximation for Pixel-Space Autoregressive Image Generationcited by 0
  12. 2026ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detectioncited by 0
4 more
  1. 2026Stable Autoregressive Speech Generation with Low-Frame-Rate High-Dimensional Continuous Tokenscited by 0
  2. 2026Learnable Classifier-Free Guidance Null Embeddings for Enhanced Controllable Speech Synthesiscited by 0
  3. 2026Learning Holographic Reduced Representations with Clifford Variational Autoencoderscited by 0
  4. 2026Sphere Encoder 2cited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Fast Autoregressive Models for Continuous Latent Generation
  2. 2024LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding
  3. 2024Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective
  4. 2025From Prediction to Perfection: Introducing Refinement to Autoregressive Image Generation
  5. 2025MixAR: Mixture Autoregressive Image Generation
  6. 2025Frequency Autoregressive Image Generation with Continuous Tokens
  7. 2025DC-AR: Efficient Masked Autoregressive Image Generation with Deep Compression Hybrid Tokenizer
  8. 2025ScalingAR: Scaling Confidence for Autoregressive Image Generation