Research map / Image generation / Autoregressive and latent generation
AccelAes: Accelerating Diffusion Transformers for Training-Free Aesthetic-Enhanced Image Generation
Xuanhua Yin, CHUANZHI XU, Haoxian Zhou, Boyu Wei, Weidong Cai
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 spatial sparse computation to step-level cache (ablation) | AccelAes (Lumina-Next) · Pick-a-Pic | ImageReward: 0.743 -> 0.841 | Improved |
| add step-level cache alone (ablation) | dense Lumina-Next · Pick-a-Pic | ImageReward: 0.752 -> 0.743 | Hurt |
| add spatial allocation alone (ablation) | dense Lumina-Next · Pick-a-Pic | ImageReward: 0.752 -> 0.818 | Improved |
| add AesMask instead of uniform mask (ablation) | AccelAes (Lumina-Next) · Pick-a-Pic | ImageReward: 0.8222 -> 0.8410 | Improved |
| use aesthetic anchor tokens instead of non-aesthetic tokens (ablation) | AesMask · Pick-a-Pic | ImageReward: 0.8207 -> 0.8410 | Improved |
| add spatial FFN sparsity on top of spatial attention sparsity (ablation) | spatial-only variant · Pick-a-Pic | ImageReward: 0.818 -> 0.825 | No clear effect |
| raise skip_ratio from 0.50 to 0.70 (ablation) | AccelAes · Pick-a-Pic | ImageReward: 0.8410 -> 0.7643 | Hurt |
| lower skip_ratio from 0.50 to 0.40 (ablation) | AccelAes · Pick-a-Pic | ImageReward: 0.8410 -> 0.8334 | No clear effect |
| delay mask_step from 5 to 15 (ablation) | AccelAes · Pick-a-Pic | speedup: 2.11x -> 1.42x | Hurt |
| use static one-shot mask instead of per-step refresh (ablation) | AccelAes · hard non-centric prompts | ImageReward: 0.381 -> 0.407 | Improved |
| use static one-shot mask instead of per-step refresh (ablation) | AccelAes · hard non-centric prompts | speedup: 1.82x -> 2.08x | Improved |
| add Spatial-CFG (ablation) | AccelAes · blind human study | overall preference win rate: 66% | Improved |
Closest papers by meaning
- AttriCtrl: A Generalizable Framework for Controlling Semantic Attribute Intensity in Diffusion Models (ICLR 2026 · similarity 0.90)
- PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis (ICLR 2024 · similarity 0.88)
- Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.87)
- Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization (CVPR 2025 · similarity 0.87)
- SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers (NeurIPS 2026 · similarity 0.86)
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.
