Research map / Image generation / Autoregressive and latent generation
SpectralAR: Spectral Autoregressive Visual Generation
Yuanhui Huang, Weiliang Chen, Wenzhao Zheng, Yueqi Duan, Jie Zhou, Jiwen Lu
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 spectral (sub-image) supervision to the 1D tokenizer (ablation) | SpectralAR-d16 (TiTok baseline) · ImageNet-1K 256x256 gFID | gFID: 6.30 -> 5.64 | Improved |
| add spectral causal mask in encoder and decoder (ablation) | SpectralAR-d16 · ImageNet-1K 256x256 gFID | gFID: 5.64 -> 3.49 | Improved |
| add non-uniform token-frequency mapping (ablation) | SpectralAR-d16 · ImageNet-1K 256x256 gFID | gFID: 3.49 -> 3.02 | Improved |
| causal mask without non-uniform mapping vs spectral supervision only, effect on IS (ablation) | SpectralAR-d16 · ImageNet-1K 256x256 IS | IS: 255.1 -> 222.6 | Mixed |
| use spectral token sequence instead of raster-scan order | CIFAR-100 token sequences · CIFAR-100 val causality proxy | average R^2: 0.454 -> 0.962 | Improved |
| spectral sequence versus scale-wise sequence | CIFAR-100 token sequences · CIFAR-100 val causality proxy | average R^2: 0.928 -> 0.962 | Improved |
| truncate last 5 generated tokens (ablation) | SpectralAR-d16 · ImageNet-1K 256x256 | FID: 3.02 -> 3.34 | Hurt |
| truncate last 15 generated tokens (ablation) | SpectralAR-d16 · ImageNet-1K 256x256 | FID: 3.02 -> 27.69 | Hurt |
| refine strided sub-images with SpectralAR for super-resolution | upsampled 512x512 images · ImageNet super-resolution | FID: 3.09 -> 2.93 | Improved |
| use patch-based sub-images instead of strided for super-resolution (ablation) | SpectralAR super-resolution · ImageNet super-resolution | FID: 2.93 -> 14.76 | Hurt |
| replace TiTok 1D AR tokenizer with SpectralAR tokenization | 310M AR generator, 64 tokens · ImageNet-1K 256x256 | gFID: 6.30 -> 3.02 | Improved |
| nested spectral tokenization instead of plain 1D tokenization | tokenizer reconstruction · ImageNet-1K val | rFID: 1.70 -> 4.03 | Hurt |
Closest papers by meaning
- NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · similarity 0.90)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.89)
- Guiding Visual Autoregressive Models through Spectrum Weakening (NeurIPS 2026 · similarity 0.89)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.88)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · 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.
