Research map / Image generation / Autoregressive and latent generation
Hyperspherical Latents Improve Continuous-Token Autoregressive Generation
Guolin Ke, HUI XUE
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add normalization of AR inputs and outputs (after decoder-input normalization) (ablation) | SphereAR-L with Gaussian VAE · ImageNet 256x256 (50-epoch ablation) | FID: 2.89 -> 2.68 | Improved |
| add normalization only on VAE decoder input (ablation) | SphereAR-L with Gaussian VAE · ImageNet 256x256 (50-epoch ablation) | FID: 2.97 -> 2.89 | Improved |
| use spherical (Power Spherical) posterior instead of Gaussian with normalization (ablation) | SphereAR-L · ImageNet 256x256 (50-epoch ablation) | FID: 2.68 -> 2.52 | Improved |
| use S-VAE instead of MAR's VAE (ablation) | SphereAR-L · ImageNet 256x256 (50-epoch ablation) | FID: 4.54 -> 2.52 | Improved |
| fix the posterior variance (sigma-VAE) instead of learned diagonal Gaussian (ablation) | AR with VAE latents · ImageNet 256x256 (50-epoch ablation) | FID vs CFG | No clear effect |
| increase KL weight of diagonal-Gaussian VAE (ablation) | AR with Gaussian VAE · ImageNet 256x256 (50-epoch ablation) | FID vs CFG | Mixed |
| apply post-hoc normalization to diagonal-Gaussian latents (ablation) | AR with Gaussian VAE · ImageNet 256x256 (50-epoch ablation) | FID vs CFG | Improved |
| use hybrid CNN+ViT VAE backbone instead of pure CNN (ablation) | S-VAE · ImageNet 256x256 (50-epoch ablation) | training speed and FID: speed 2.25 -> 5.81 it/s; FID 2.63 -> 2.52 | Improved |
| use pure ViT VAE backbone instead of hybrid (ablation) | S-VAE · ImageNet 256x256 (50-epoch ablation) | FID: 2.52 -> 2.81 | Hurt |
| reduce diffusion-head sampling steps (100 to 10 to 1) (ablation) | SphereAR · inference speed on A100 | inference speed | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GIVT (1.67B) — GIVT: Generative Infinite-Vocabulary Transformers (ECCV 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.88)
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.88)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.88)
- ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 2025 · similarity 0.88)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 2025 · similarity 0.88)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
