Research map / Image generation / Autoregressive and latent generation
VVS: Accelerating Speculative Decoding for Visual Autoregressive Generation via Partial Verification Skipping
Haotian Dong, Ye Li, Rongwei Lu, Chen Tang, Shu-Tao Xia, Zhi Wang
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 partial verification skipping (VVS-U i=2) to relaxed-acceptance SD | LlamaGen-XL Stage II · MS-COCO 2017 | TPF: 1.83 -> 2.86 | Improved |
| add verification skipping vs LANTERN (Stage II) | LlamaGen-XL Stage II · MS-COCO 2017 | Recall: 0.5360 -> 0.4588 | Hurt |
| add verification skipping vs LANTERN (Stage I) | LlamaGen-XL Stage I · MS-COCO 2017 | wall-clock speedup: 1.45x -> 1.63x | Improved |
| sample the token path uniformly instead of picking the highest confidence (ablation) | VVS · MS-COCO (delta=0.1) | FID: 24.71 -> 23.88 | Improved |
| add dynamic truncation of selected token path (ablation) | VVS (confidence selection) · MS-COCO (delta=0.1) | FID: 26.24 -> 24.71 | Improved |
| add dynamic truncation of selected token path (ablation) | VVS (confidence selection) · MS-COCO (delta=0.2) | FID: 27.41 -> 26.20 | Improved |
| blend fresh features with most recently cached features instead of only stale features (ablation) | VVS · MS-COCO (delta=0.2) | FID: 32.63 -> 27.69 | Improved |
| blend fresh with most recent (s2=0) rather than staler (s2=3) features (ablation) | VVS · MS-COCO (delta=0.2) | TPF: 2.27 -> 2.31 | Improved |
| similarity-based dynamic skipping instead of uniform skipping (ablation) | VVS · MS-COCO FID-TPF trade-off | FID vs TPF | Improved |
| EAGLE-2 style SD without relaxation on visual AR | LlamaGen-XL Stage I · MS-COCO 2017 | wall-clock speedup: 1.00x -> 0.87x | Hurt |
Closest papers by meaning
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.93)
- Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient (CVPR 2025 · similarity 0.90)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (ICML 2025 · similarity 0.90)
- Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (NeurIPS 2025 · similarity 0.90)
- Parallelized Autoregressive Visual Generation (CVPR 2025 · similarity 0.89)
- RandAR: Decoder-only Autoregressive Visual Generation in Random Orders (CVPR 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.
