Research map / Image generation / Autoregressive and latent generation
"Principal Components" Enable A New Language of Images
Xin Wen, Bingchen Zhao, Ismail Elezi, Jiankang Deng, Xiaojuan Qi
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 Semanticist DiT-L tokenizer instead of ImageFolder | Semanticist · ImageNet 256x256 reconstruction | rFID: 0.80 -> 0.72 (DiT-XL: 0.72) | Improved |
| scale diffusion decoder from DiT-L to DiT-XL (ablation) | Semanticist · ImageNet 256x256 reconstruction | rFID: 0.78 -> 0.72 | Improved |
| scale diffusion decoder from DiT-L to DiT-XL (ablation) | Semanticist · ImageNet 256x256 reconstruction | PSNR: 21.61 -> 21.43 | Hurt |
| scale diffusion decoder from DiT-L to DiT-XL (ablation) | Semanticist · ImageNet 256x256 reconstruction | SSIM: 0.626 -> 0.613 | Hurt |
| scale diffusion decoder from DiT-L to DiT-XL (ablation) | Semanticist · ImageNet 256x256 class-conditional generation | gFID: 2.57 -> 2.57 | No clear effect |
| scale diffusion decoder from DiT-L to DiT-XL (ablation) | Semanticist · ImageNet 256x256 class-conditional generation | IS: 260.9 -> 254.0 | Hurt |
| train autoregressive model on first 32 of 256 tokens only | Semanticist · ImageNet 256x256 class-conditional generation | gFID: 2.57 with 32 tokens/32 steps | Improved |
| add REPA regularizer to the DiT decoder (ablation) | Semanticist · ImageNet reconstruction (d64x64 tokens) | rFID | Improved |
| add REPA regularizer to the DiT decoder (ablation) | Semanticist · ImageNet reconstruction (d64x64 tokens) | rFID with all tokens | No clear effect |
| use lower-dimensional tokens (d16x256 vs d256x16) (ablation) | Semanticist · ImageNet reconstruction | rFID | Improved |
| increase CFG scale at inference (ablation) | Semanticist · ImageNet reconstruction (smaller model, fewer tokens) | reconstruction quality | Improved |
| use more tokens for linear probing (ablation) | Semanticist · ImageNet linear probing | top-1 accuracy: peak 63.5% at few tokens, declining as more tokens added | Hurt |
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.
- ImageFolder — ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025) · better on 2 of 2 shared results
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
- TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation (CVPR 2025 · similarity 0.89)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.89)
- UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation (ICLR 2026 · similarity 0.88)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.88)
- Image Understanding Makes for A Good Tokenizer for Image Generation (NeurIPS 2024 · similarity 0.88)
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.87)
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.
