Research map / Image generation / Autoregressive and latent generation
FVAR: Next-Focus Prediction for Visual Autoregressive Modeling
Xiaofan Li, Chenming Wu, Yanpeng Sun, Jiaming Zhou, Delin Qu, Yansong Qu, Weihao Bo, Haibao Yu, Dingkang Liang
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 |
|---|---|---|---|
| replace uniform downsampling pyramid with progressive refocusing (defocus PSF) (ablation) | FVAR-d16 · ImageNet 256x256 | FID: 3.51 -> 2.89 | Improved |
| replace uniform downsampling with progressive refocusing (ablation) | FVAR-d16 · ImageNet 1024x1024 | FID: 8.15 -> 6.85 | Improved |
| use physics-based defocus PSF instead of Gaussian blur (ablation) | FVAR-d16 · ImageNet 256x256 | FID: 3.32 -> 2.89 | Improved |
| Gaussian blur instead of standard downsampling (ablation) | VAR-d16 · ImageNet 1024x1024 | FID: 8.15 -> 7.50 | Improved |
| add high-frequency residual teacher with distillation (ablation) | FVAR-d16 · ImageNet 256x256 | FID: 3.06 -> 2.89 | Improved |
| add high-frequency residual teacher (ablation) | FVAR-d16 · ImageNet 1024x1024 | FID: 7.20 -> 6.85 | Improved |
| use separate dual tokenizers for structure and alias (ablation) | FVAR-d16 · ImageNet 256x256 | FID: 3.14 -> 2.89 | Improved |
| use separate dual tokenizers (ablation) | FVAR-d16 · ImageNet 1024x1024 | FID: 7.40 -> 6.85 | Improved |
| FVAR on VAR-d12 versus VAR-d12 | VAR-d12 · ImageNet 256x256 | FID: 5.81 -> 3.95 | Improved |
| FVAR versus M-VAR at d24 (no added parameters) | VAR-d24 · ImageNet 256x256 | FID: 1.93 -> 1.75 | Improved |
| FVAR at 512x512 versus VAR | VAR-d36 · ImageNet 512x512 | FID: 2.63 -> 2.28 | Improved |
Closest papers by meaning
- ARVAR: Accelerating Visual Autoregressive Model via Attention Retrospect (ECCV 2026 · similarity 0.91)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.90)
- DVAR: Dynamic Visual Autoregressive Modeling for Image Super-Resolution (CVPR 2026 · similarity 0.90)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.89)
- Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · similarity 0.87)
- Edit the Bits, Diff the Codes: Bitwise Residual Editing for Visual Autoregressive Models (NeurIPS 2026 · 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.
