Research map / Image generation / Fast diffusion sampling
Continuous Speculative Decoding for Autoregressive Image Generation
Zili Wang, Zheng Zhang, Kun Ding, Qi Yang, Fei Li, SHIMING XIANG
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 denoising trajectory alignment (ablation) | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256, draft length 32 | acceptance rate: 0.07 -> 0.30 | Improved |
| add denoising trajectory alignment (ablation) | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256, draft length 32 | draft-target distance: 2.56 -> 1.13 | Improved |
| add denoising trajectory alignment (ablation) | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256, draft length 4 | acceptance rate: 0.14 -> 0.32 | Improved |
| add denoising trajectory alignment to the approximated ratio (ablation) | p(Yp)/q(Yq) criterion · probability ratio analysis | acceptance rate: 14% -> 32% | Improved |
| use approximated two-path ratio instead of exact path ratio p(Y)/q(Y) (ablation) | acceptance criterion · probability ratio analysis | acceptance rate: 0.0% -> 14% | Improved |
| add token pre-filling of 5% target tokens (ablation) | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256, draft length 32 | acceptance rate: 0.25 -> 0.30 | Improved |
| increase pre-filling from 5% to 15% (ablation) | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256, draft length 32 | speedup: 1.63x -> 1.61x | No clear effect |
| increase pre-fill ratio from 5% to 15% (ablation) | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256, draft length 32 | acceptance rate: 0.30 -> 0.33 | Improved |
| increase draft length from 4 to 32 | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256, bs=256 | speedup ratio: 1.62x -> 2.33x | Improved |
| increase draft length from 4 to 32 | continuous speculative decoding (MAR-H/MAR-B) · ImageNet 256 | acceptance rate: 0.30 -> 0.19 | Hurt |
| increase batch size from 1 to 256 | continuous speculative decoding (MAR-H/MAR-B, draft length 32) · ImageNet 256 | speedup ratio: 1.44x -> 2.33x | Improved |
| apply continuous speculative decoding with MAR-B draft | MAR-H · ImageNet 256 w/ CFG | FID: 1.55 -> 1.60 | No clear effect |
Closest papers by meaning
- LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding (ICLR 2025 · similarity 0.92)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.90)
- Hawk: Leveraging Spatial Context for Faster Autoregressive Text-to-Image Generation (NeurIPS 2025 · similarity 0.90)
- Speculative Jacobi-Denoising Decoding for Accelerating Autoregressive Text-to-image Generation (NeurIPS 2025 · similarity 0.90)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.89)
- ZipAR: Parallel Autoregressive Image Generation through Spatial Locality (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.
