Research map / Image generation / Autoregressive and latent generation
Spectral Image Tokenizer
Carlos Esteves, Mohammed Suhail, Ameesh Makadia
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 |
|---|---|---|---|
| use DWT spectral tokens (SIT-5) instead of ViT-VQGAN patch tokens | ViT-VQGAN · ImageNet 256x256 reconstruction | FID: 2.33 -> 0.97 | Improved |
| use SIT-6 instead of ViT-VQGAN at 512x512 | ViT-VQGAN · ImageNet 512x512 reconstruction | FID: 6.915 -> 1.74 | Improved |
| add scale-causal attention to encoder and decoder (SIT-SC-5) (ablation) | SIT-5 · ImageNet 256x256 reconstruction | FID: 0.97 -> 1.33 | Hurt |
| add scale-causal decoder attention (ablation) | SIT-4 · ImageNet 256x256 reconstruction | FID: 1.20 -> 1.45 | Hurt |
| add scale-causal attention on encoder and decoder (ablation) | SIT-4 · ImageNet 256x256 reconstruction | FID: 1.20 -> 1.97 | Hurt |
| use LeGall5/3 wavelet instead of Haar (ablation) | SIT-5 · ImageNet reconstruction | FID: 0.97 -> 1.10 | Hurt |
| use CDF9/7 wavelet instead of Haar (ablation) | SIT-5 · ImageNet reconstruction | FID: 0.97 -> 1.14 | Hurt |
| reduce vocabulary from 8192 to 4096 (ablation) | SIT-5 · ImageNet reconstruction | FID: 0.97 -> 1.33 | Hurt |
| increase vocabulary from 8192 to 16384 (ablation) | SIT-5 · ImageNet reconstruction | FID: 0.97 -> 0.91 | Improved |
| use SIT-2 with sequence length 2048 instead of SIT-4 with 1024 (ablation) | SIT · ImageNet reconstruction | FID: 1.20 -> 0.77 | Improved |
| use SIT-SC-5 instead of ViT-VQGAN at 64x64 input | ViT-VQGAN · ImageNet 64x64 reconstruction | FID: 3.53 -> 1.39 | Improved |
| use AR-SIT-SCD-4 instead of Parti350M at 256x256 | Parti350M · MS-COCO text-to-image | FID: 12.35 -> 12.57 | No clear effect |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.88)
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.88)
- ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025 · similarity 0.87)
- FlexTok: Resampling Images into 1D Token Sequences of Flexible Length (ICML 2025 · similarity 0.87)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · similarity 0.87)
- Learnings from Scaling Visual Tokenizers for Reconstruction and Generation (ICML 2025 · similarity 0.86)
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.
