Research map / Image generation / Autoregressive and latent generation
Holistic Tokenizer for Autoregressive Image Generation
Anlin Zheng, Haochen Wang, Yucheng Zhao, Weipeng Deng, Tiancai Wang, Xiangyu Zhang, 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 |
|---|---|---|---|
| add learnable queries (ablation) | Hita tokenizer · ImageNet 256 (Hita-B) | gFID: 9.37 -> 6.32 | Improved |
| add DINOv2 semantic injection (ablation) | Hita tokenizer · ImageNet 256 (Hita-B) | gFID: 6.32 -> 5.85 | Improved |
| add DINOv2 semantic injection (ablation) | Hita tokenizer · Holistic codebook usage | usage (%): 30.2 -> 100.0 | Improved |
| add DINOv2 semantic injection (ablation) | Hita tokenizer · ImageNet linear probing | top-1 accuracy: 28.2 -> 36.6 | Improved |
| use k>0 last holistic tokens in fusion (ablation) | Hita tokenizer · ImageNet 256 (Hita-L) | gFID: 6.80 -> 3.78 | Improved |
| use k=4 instead of k=1 (ablation) | Hita tokenizer · ImageNet 256 (Hita-L) | gFID: 3.78 -> 3.75 | Improved |
| use k=4 instead of k=2 (ablation) | Hita tokenizer · ImageNet 256 (Hita-L) | gFID: 4.14 -> 3.75 | Improved |
| use DINOv2 instead of CLIP for semantic injection (ablation) | Hita tokenizer · ImageNet 256 (Hita-B) | gFID: 6.42 -> 5.85 | Improved |
| inject ResNet features instead of nothing (ablation) | Hita tokenizer · ImageNet 256 (Hita-B) | gFID: 6.32 -> 6.81 | Hurt |
| use proposed fusion instead of Full variant (ablation) | Hita tokenizer · ImageNet 256 (Hita-B) | gFID: 11.64 -> 5.85 | Improved |
| add E_trans to the attention modules (ablation) | Hita tokenizer · ImageNet 256 (Hita-B) | gFID: 6.01 -> 5.85 | Improved |
| add E_causal to the attention modules (ablation) | Hita tokenizer · ImageNet 256 (Hita-B) | gFID: 6.12 -> 5.85 | Improved |
Closest papers by meaning
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.88)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.86)
- EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.86)
- OmniGen-AR: AutoRegressive Any-to-Image Generation (NeurIPS 2025 · similarity 0.86)
- ImageFolder: Autoregressive Image Generation with Folded Tokens (ICLR 2025 · similarity 0.86)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (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.
