Research map / Image generation / Autoregressive and latent generation
End-to-End Autoregressive Image Generation with 1D Semantic Tokenizer
Wenda Chu, Bingliang Zhang, Jiaqi Han, Yizhuo Li, Linjie Yang, Yisong Yue, Qiushan Guo
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 APR loss to end-to-end training (ablation) | EOSTok end-to-end training · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 8.01 -> 3.32 | Improved |
| add NTP supervision to tokenizer without APR (vanilla E2E) (ablation) | tokenizer · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | AR prediction accuracy: 10.8% -> 30.2% | Mixed |
| add decoder alignment to VFM (ablation) | 1D tokenizer baseline · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 12.27 -> 5.68 | Improved |
| add decoder alignment to VFM (ablation) | 1D tokenizer baseline · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | rFID: 1.75 -> 1.12 | Improved |
| use direct alignment of latent tokens to VFM (ablation) | 1D tokenizer with decoder alignment · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 5.68 -> 5.98 | Hurt |
| use implicit alignment instead of direct alignment (ablation) | 1D tokenizer · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 5.98 -> 3.32 | Improved |
| use implicit alignment instead of direct substitution (ablation) | 1D tokenizer · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 4.89 -> 3.32 | Improved |
| set APR weight to 1.0 instead of 0.0 (ablation) | EOSTok-L · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 4.09 -> 3.32 | Improved |
| increase APR weight from 1.0 to 4.0 (ablation) | EOSTok-L · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 3.32 -> 3.57 | Hurt |
| increase codebook size from 4096 to 16384 (ablation) | EOSTok-L · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 3.32 -> 4.08 | Hurt |
| increase codebook size from 4096 to 16384 (ablation) | EOSTok-L · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | rFID: 1.02 -> 0.96 | Improved |
| apply nested dropout with probability 0.5 (ablation) | EOSTok-L · ImageNet 256 (EOSTok-L, 50 epochs, w/o guidance) | gFID: 3.70 -> 3.32 | Improved |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.94)
- ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025 · similarity 0.92)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.92)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.91)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.91)
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.90)
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.
