Research map / Image generation / Autoregressive and latent generation
Locally Hierarchical Auto-Regressive Modeling for Image Generation
Tackgeun You, Saehoon Kim, Chiheon Kim, Doyup Lee, Bohyung Han
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 |
|---|---|---|---|
| PixelUnshuffle/shuffle resizing instead of average pooling (ablation) | HQ-VAE · ImageNet val | rFID: 3.67 -> 3.14 | Improved |
| train HQ-VAE for 50 instead of 15 epochs (ablation) | HQ-VAE · ImageNet val | rFID: 3.14 -> 2.61 | Improved |
| locally sequential instead of global conditioning (ablation) | HQ-Transformer · ImageNet | FID: 23.25 -> 11.33 | Improved |
| locally hierarchical instead of locally sequential decoding (ablation) | HQ-Transformer · ImageNet | FID: 11.33 -> 10.01 | Improved |
| soft labels with temperature tau=1 instead of one-hot (ablation) | HQ-Transformer · ImageNet | FID: 10.01 -> 9.36 | Improved |
| add third code level (4x4+8x8 to 4x4+8x8+16x16) (ablation) | HQ-TVAE · ImageNet val | FID: 37.59 -> 15.66 | Improved |
| add third code level (8x8+16x16 to 8x8+16x16+32x32) (ablation) | HQ-TVAE · ImageNet val | FID: 13.59 -> 11.10 | Improved |
| add third level 4x4 to 8x8+16x16 codes (ablation) | HQ-VAE · ImageNet val | rFID: 18.65 -> 2.48 | Improved |
| increase HQ-VAE resolution from 8x8+16x16 to 16x16+32x32 | HQ-VAE · ImageNet val | rFID: 2.61 -> 0.57 | Improved |
| scale HQ-TVAE from S to L | HQ-TVAE · ImageNet class-conditional | FID: 9.36 -> 7.15 | Improved |
| scale HQ-TVAE from S to L | HQ-TVAE · ImageNet class-conditional | throughput: 78.87 -> 49.44 | Hurt |
| two-level hierarchical codes instead of RQ-Transformer residual codes | HQ-TVAE (S) · CC3M | throughput: 46.47 -> 54.88 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MaskGIT — MaskGIT: Masked Generative Image Transformer (CVPR 2022) · better on 0 of 1 shared result
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
- Hierarchical Quantized Autoencoders (NeurIPS 2020 · similarity 0.91)
- Towards Accurate Image Coding: Improved Autoregressive Image Generation With Dynamic Vector Quantization (CVPR 2023 · similarity 0.89)
- Learned Image Transmission with Hierarchical Variational Autoencoder (AAAI 2025 · similarity 0.87)
- Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.86)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.86)
- Efficient Generative Modeling with Residual Vector Quantization-Based Tokens (ICML 2025 · similarity 0.85)
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.
