Research map / Architectures and efficiency / Attention and transformers
AMS-KV: Adaptive KV Caching in Multi-Scale Visual Autoregressive Transformers
Boxun Xu, Yu Wang, Zihu Wang, Peng 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 |
|---|---|---|---|
| add AMS-KV caching to VAR-d30 | AMS-KV on VAR-d30 · ImageNet 256 | FID: 2.07 -> 2.09 | No clear effect |
| add AMS-KV caching to VAR-d30 | AMS-KV on VAR-d30 · ImageNet 256 | KV cache size: 22.41GB -> 4.77GB (78.72% down) | Improved |
| add AMS-KV caching to VAR-d36 | AMS-KV on VAR-d30 · ImageNet 512 | KV cache size: 53.15GB -> 15.26GB (71.29% down) | Improved |
| add AMS-KV caching to Infinity-2B | AMS-KV on Infinity-2B · GenEval | overall score: 0.711 -> 0.714 | No clear effect |
| add AMS-KV caching to VAR-d30 | AMS-KV on VAR-d30 · batch size 128 and 256 | feasibility: OOM -> 23.00 and 23.88 img/s | Improved |
| add AMS-KV caching to VAR-d30 | AMS-KV on VAR-d30 · batch size 16 | throughput: 18.73 -> 18.86 img/s | No clear effect |
| keep the KV cache of intermediate scales (ablation) | VAR-d30 · ImageNet 256 | FID: 2.12 -> 2.07 | No clear effect |
| keep the KV cache of local scales (ablation) | VAR-d30 · ImageNet 256 | FID: 3.04 -> 2.07 | Improved |
| keep the KV cache of condensed scales (ablation) | VAR-d30 · ImageNet 256 | FID: 2.48 -> 2.07 | Improved |
| use similarity-aware instead of uniform cache allocation under same budget (ablation) | VAR-d30 · ImageNet 256 | FID: 6.65 -> 3.37 | Improved |
| add the first scale as condensed scale (ablation) | VAR-d30 · ImageNet 256 | FID: 2.48 -> 2.16 | Improved |
| add the second scale as condensed scale (ablation) | VAR-d30 · ImageNet 256 | FID: 2.16 -> 2.09 | Improved |
Closest papers by meaning
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.94)
- HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling (NeurIPS 2026 · similarity 0.90)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.88)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
- PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation (NeurIPS 2025 · similarity 0.88)
- Fundamental Limits of Visual Autoregressive Transformers: Universal Approximation Abilities (ICML 2025 · 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.
