Research map / Image generation / Autoregressive and latent generation
A Flat Vocabulary or a Rich Hierarchy? Re-introducing Intrinsic Structure Transforms the Autoregressive Image Generation
Landis He, Shikang Zheng
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 MASC hierarchical vocabulary reduction (k=8,192) | LlamaGen-XL · ImageNet 256x256 class-conditional | FID: 2.87 -> 2.58 | Improved |
| add k-means clustering prior instead of MASC (ablation) | LlamaGen-XL · ImageNet 256x256 class-conditional | FID: 2.87 -> 3.09 | Hurt |
| add k-means prior to LlamaGen-L (ablation) | LlamaGen-L · ImageNet 256x256 class-conditional | FID: 3.38 -> 3.81 | Hurt |
| add MASC to LlamaGen-L | LlamaGen-L · ImageNet 256x256 class-conditional | FID: 3.38 -> 2.92 | Improved |
| reduce MASC vocabulary from 8,192 to 4,096 (ablation) | LlamaGen-XL · ImageNet 256x256 class-conditional | FID: 2.58 -> 2.49 | Improved |
| reduce MASC vocabulary from 4,096 to 2,048 (ablation) | LlamaGen-XL · ImageNet 256x256 class-conditional | FID: 2.49 -> 2.65 | Hurt |
| use random clustering of the vocabulary (ablation) | VAR-d24 · ImageNet 256x256 class-conditional | FID: 2.09 -> 3.84 | Hurt |
| add MASC (2k) to VAR-d24 | VAR-d24 · ImageNet 256x256 class-conditional | FID: 2.09 -> 1.98 | Improved |
| add MASC (4k) to RAR-L with a large-vocabulary tokenizer | RAR-L + LlamaGen · ImageNet 256x256 class-conditional | convergence and FID: diverged -> 1.61 | Improved |
| use MASC target space instead of random or k-means at 4k (ablation) | RAR-L · ImageNet 256x256 class-conditional | FID: 4.24 (k-means) -> 1.61 | Improved |
| use manifold-aligned distance on top of density-driven construction (full MASC vs centroid) (ablation) | LlamaGen-XL · ImageNet 256x256 class-conditional | FID: 3.02 -> 2.58 | Improved |
| use density-driven bottom-up construction vs k-means++ (ablation) | LlamaGen-XL · ImageNet 256x256 class-conditional | FID: 3.17 -> 3.02 | Improved |
Closest papers by meaning
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.91)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.90)
- Improving Autoregressive Visual Generation with Cluster-Oriented Token Prediction (CVPR 2025 · similarity 0.90)
- Vision Foundation Models as Effective Visual Tokenizers for Autoregressive Generation (NeurIPS 2025 · similarity 0.89)
- Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.89)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · 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.
