Research map / Image generation / Autoregressive and latent generation
EVAR: Edge Visual Autoregressive Models via Principled Pruning
Zefang Wang, Ying Li, Yanyu Li, Mingluo Su, Simin Xu, Guanzhong Tian, Huan Wang
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 pyramid-aware Hessian accumulation instead of direct concatenation of multi-scale tokens (ablation) | EVAR structured OBS pruning · VAR-d16, 40% sparsity, training-free | FID: 69.21 -> 58.21 | Improved |
| calibrate with tokens from all scales instead of a single scale (ablation) | EVAR structured OBS pruning · VAR-d16, 40% sparsity, training-free | FID: 86.51 -> 69.21 | Improved |
| add post-pruning fine-tuning (ablation) | EVAR pruned VAR-d16 (40% sparsity) · ImageNet 256 | FID: 58.21 -> 4.25 | Improved |
| add PSAD instead of vanilla KD (ablation) | EVAR pruned VAR-d16 (40% sparsity) · ImageNet 256 | FID: 4.26 -> 3.91 | Improved |
| weight distillation toward later scales with linear increase (ablation) | EVAR distillation · VAR-d16, 40% sparsity | FID: 4.26 -> 5.45 | Hurt |
| weight distillation toward later scales with exponential increase (ablation) | EVAR distillation · VAR-d16, 40% sparsity | FID: 4.26 -> 5.82 | Hurt |
| use naive exponential decay scale weighting instead of vanilla KD (ablation) | EVAR distillation · VAR-d16, 40% sparsity | FID: 4.26 -> 4.36 | Hurt |
| EVAR pyramid-aware OBS versus SlimGPT OBS with flat calibration | VAR-d16 pruning · 20% sparsity, training-free | FID: 56.87 -> 7.85 | Improved |
| EVAR versus best baseline (Taylor) after 1 epoch fine-tuning | VAR-d16 pruning · 40% sparsity, 1-epoch | FID: 8.35 -> 4.86 | Improved |
| prune VAR-d16 from 20% to 40% sparsity with EVAR | EVAR VAR-d16 · ImageNet 256 | FID: 3.67 -> 3.91 | Mixed |
| pin Core ML to CPU+GPU instead of All compute units (ablation) | EVAR on iPad Pro (M4) · single-image latency | latency (ms): 494 -> 277 | Improved |
| use CPU+NPU compute units instead of CPU+GPU (ablation) | EVAR on iPad Pro (M4) · single-image latency | latency (ms): 277 -> 1090 | Hurt |
Closest papers by meaning
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.90)
- Visual Implicit Autoregressive Modeling (ICML 2026 · similarity 0.89)
- Fundamental Limits of Visual Autoregressive Transformers: Universal Approximation Abilities (ICML 2025 · similarity 0.87)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.87)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.87)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · 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.
