Research map / Image generation / Autoregressive and latent generation
Entropy-Aware Dynamic KV Cache Sparsification for Autoregressive Image Generation and Editing
Tong Tong, Ling Xing, Linjie Li, Rui Yan, Zhengyuan Yang, Lijuan Wang, Alex Jinpeng 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 |
|---|---|---|---|
| enable SparseAR entropy-aware KV sparsification at tau=1/6 | LlamaGen-XL · Text-to-image, FID-30K | FID-30K: 5.67 -> 5.55 | Improved |
| start sparsification earlier (tau 1/6 -> 1/8) (ablation) | SparseAR on Lumina-mGPT-768 · Text-to-image, FID-30K | FID-30K: 4.91 -> 10.88 | Hurt |
| start sparsification earlier (tau 1/6 -> 1/8) (ablation) | SparseAR on Lumina-mGPT-1024 · Text-to-image, FID-30K | FID-30K: 3.37 -> 4.08 | Hurt |
| start sparsification earlier (tau 1/6 -> 1/8) (ablation) | SparseAR on Lumina-mGPT-512 · Text-to-image, FID-30K | FID-30K: 6.54 -> 7.22 | Hurt |
| use entropy-aware SparseAR instead of Streaming KV compression | BLIP3o-Next image editing · ImgEdit score | ImgEdit: 1.98 -> 3.62 | Improved |
| use entropy-aware SparseAR instead of H2O | BLIP3o-Next image editing · FID-5k | FID-5k: 19.55 -> 13.80 | Improved |
| use SparseAR at tau=1/4 | BLIP3o-Next image editing · Latency | latency: 122.78s -> 111.74s | Improved |
| use fixed pattern compression keeping only condition, initial and recent tokens (ablation) | LlamaGen-XL · qualitative text-to-image | image detail | Hurt |
| compress KV in low-entropy regions (entropy 0 to 3) (ablation) | Lumina-mGPT-512 · text-to-image | generation quality | No clear effect |
| compress KV in higher-entropy regions (ablation) | Lumina-mGPT-512 · text-to-image | generation quality | Hurt |
| lower compression budget to 20% with tau 1/6 or 1/4 (ablation) | SparseAR on LlamaGen · FID | FID | Mixed |
Closest papers by meaning
- Autoregressive Image Generation Needs Only a Few Lines of Cached Tokens (ECCV 2026 · similarity 0.91)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.90)
- HeatKV: Head-tuned KV-cache Compression for Visual Autoregressive Modeling (NeurIPS 2026 · similarity 0.90)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.89)
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.89)
- ControlAR: Controllable Image Generation with Autoregressive Models (ICLR 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.
