Research map / Image generation / Autoregressive and latent generation
MoSs: Mixture of Scales for Efficient High-Resolution Autoregressive Image Generation
Yaoxiu Lian, Hao Liang, Zhihong Gou, Yijia Zhang, Jiaming Xu, Guohao Dai, Ningyi Xu
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 |
|---|---|---|---|
| apply MoSs5 parallel refinement | Infinity-2B · Infinity-2B 1024x1024 | latency (s): 1.29 -> 0.26 (4.96x speedup) | Improved |
| apply MoSs5 | Infinity-2B · Infinity-2B | KV cache (GB): 3.13 -> 0.45 | Improved |
| apply MoSs5 (more parallel scales) | Infinity-2B · GenEval | overall: 0.729 -> 0.705 | Hurt |
| apply MoSs2 | Infinity-2B · GenEval | overall: 0.729 -> 0.728 | No clear effect |
| apply MoSs2 | HART-0.7B · 1024x1024 | speedup: 1.56x speedup; KV cache 2.79 -> 0.90 GB | Improved |
| apply MoSs2 | HART-0.7B · GenEval | overall: 0.498 -> 0.488 | Hurt |
| add Cross-Scale Self-Attention instead of standard self-attention (ablation) | MoSs5 · GenEval | overall: 0.674 -> 0.705 | Improved |
| add Cross-Scale Self-Attention instead of standard self-attention (ablation) | MoSs2 · GenEval | overall: 0.720 -> 0.728 | Improved |
| add Triton RoPE2D kernel (ablation) | MoSs with CSSA · Infinity-2B | speedup: 1.87 -> 2.05 (MoSs2) | Improved |
| MoSs3 vs FastVAR | Infinity-2B · Infinity-2B | GenEval overall / speedup: 0.724 vs 0.721; both 2.87x | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Infinity — Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025) · better on 3 of 5 shared results
- HART — HART: Efficient Visual Generation with Hybrid Autoregressive Transformer (ICLR 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- HMAR: Efficient Hierarchical Masked Auto-Regressive Image Generation (CVPR 2025 · similarity 0.90)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.90)
- Memory-Efficient Visual Autoregressive Modeling with Scale-Aware KV Cache Compression (NeurIPS 2025 · similarity 0.89)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.89)
- Hierarchical Masked Autoregressive Models with Low-Resolution Token Pivots (ICML 2025 · similarity 0.89)
- NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 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.
