Research map / Image generation / Fast diffusion sampling
Visual Implicit Autoregressive Modeling
Pengfei Jiang, Jixiang Luo, Luxi Lin, Zhaohong Huang, Xuelong Li
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 deep explicit stack with implicit equilibrium layer (VIAR vs VAR-d30, cfg=2.0) | VAR-d30 · ImageNet 256x256 | sFID: 8.86 -> 7.92 | Improved |
| replace deep explicit stack with implicit equilibrium layer (cfg=2.0) | VAR-d30 · ImageNet 256x256 | FID: 2.05 -> 2.35 | Hurt |
| replace deep explicit stack with implicit layer | VAR-d30 · ImageNet 256x256 | parameters: 2010.0M -> 770.9M | Improved |
| use most aggressive iteration schedule s4 | VIAR · ImageNet 256x256 (RTX 4090) | throughput: 15.16 -> 32.08 images/s | Improved |
| use most aggressive iteration schedule s4 instead of s1 | VIAR · ImageNet 256x256 (RTX 4090) | FID: 2.16 -> 2.43 | Hurt |
| replace VAR-d30 by VIAR | VAR-d30 · training memory | parameter/gradient memory: 7.49 GB -> 2.87 GB | Improved |
| replace VAR-d30 by VIAR | VAR-d30 · training memory | optimizer-state memory: 14.98 GB -> 5.74 GB | Improved |
| use constant 20 iterations instead of constant 10 per scale (ablation) | VIAR · ImageNet 256x256 | FID: 2.16 -> 2.16 | No clear effect |
| use constant 5 iterations instead of 10 per scale (ablation) | VIAR · ImageNet 256x256 | FID: 2.16 -> 2.27 | Hurt |
| use decreasing schedule (20,5) instead of constant (10,10) (ablation) | VIAR · ImageNet 256x256 | FID: 2.16 -> 2.18 | Hurt |
| use decreasing schedule (10,5) instead of constant (10,10) (ablation) | VIAR · ImageNet 256x256 | FID: 2.16 -> 2.22 | Hurt |
| use decreasing schedule (20,10) instead of constant (10,10) (ablation) | VIAR · ImageNet 256x256 | FID: 2.16 -> 2.16 | No clear effect |
Closest papers by meaning
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.93)
- FasterVAR: Plug-and-Play Acceleration for Visual Autoregressive Models (ICML 2026 · similarity 0.91)
- Neighboring Autoregressive Modeling for Efficient Visual Generation (ICCV 2025 · similarity 0.90)
- FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.90)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.90)
- Fundamental Limits of Visual Autoregressive Transformers: Universal Approximation Abilities (ICML 2025 · similarity 0.89)
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.
