Research map / Image generation / Autoregressive and latent generation
AViTS:Adaptive Spatiotemporal Token Selection for Efficient Dynamic-Resolution Generation
Haoran Qin, zhengan yan, Shikang Zheng, Xiaobing Tu, Jiacheng Liu, Yuqi Lin, Chang Zou, Jinshan Liu, Peiliang Cai, Xiantao Zhang, Jinkui Ren, Linfeng Zhang
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 step-variance temporal importance to attention-only selection (ablation) | AViTS · DrawBench NFE 30 | ImageReward: 0.9875 -> 0.9959 | Improved |
| add attention spatial importance to step-variance-only selection (ablation) | AViTS · DrawBench NFE 30 | ImageReward: 0.9846 -> 0.9959 | Improved |
| use importance-based selection instead of random token selection (ablation) | AViTS · DrawBench NFE 30 | ImageReward: 0.9257 -> 0.9959 | Improved |
| use importance-based selection instead of even token selection (ablation) | AViTS · DrawBench NFE 30 | CLIP Score: 32.012 -> 32.361 | Improved |
| use attention+variance instead of edge detection (ablation) | AViTS · DrawBench NFE 30 | ImageReward: 0.9512 -> 0.9959 | Improved |
| AViTS replaces RALU at NFE 30 | dynamic-resolution sampling · GEdit-EN Qwen-Image-Edit | Overall Score: 7.52 -> 7.62 | Improved |
| AViTS replaces TaylorSeer at similar speedup | FLUX.1-Kontext-dev · GEdit-EN | Overall Score: 6.17 -> 6.70 | Improved |
| AViTS replaces FORA at similar speedup | Qwen-Image-Edit · GEdit-EN | Overall Score: 7.28 -> 7.55 | Improved |
| apply AViTS at 1440x1440 instead of direct sampling | FLUX.1-dev · DrawBench 1440x1440 | ImageReward: 0.7385 -> 0.9008 | Improved |
| apply AViTS at 2048x2048 instead of direct sampling | FLUX.1-dev · DrawBench 2048x2048 | ImageReward: 0.8371 -> 0.8905 | Improved |
| increase upsampling ratio rho (ablation) | AViTS · DrawBench | ImageReward | Mixed |
| use either attention-only (alpha=1) or variance-only (alpha=0) rather than intermediate alpha (ablation) | AViTS · FLUX.1-dev, Qwen-Image-Edit, Kontext | ImageReward / OS | Hurt |
Closest papers by meaning
- An Image is Worth 32 Tokens for Reconstruction and Generation (NeurIPS 2024 · similarity 0.86)
- KOALA: Empirical Lessons Toward Memory-Efficient and Fast Diffusion Models for Text-to-Image Synthesis (NeurIPS 2024 · similarity 0.85)
- Language-Guided Image Tokenization for Generation (CVPR 2025 · similarity 0.85)
- Which Tokens to Merge? Diffusion Dynamics for Efficient Image Generation (NeurIPS 2026 · similarity 0.85)
- Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models (NeurIPS 2026 · similarity 0.85)
- Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis (CVPR 2025 · similarity 0.85)
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.
