Research map / Image generation / Autoregressive and latent generation
Massive Activations are the Key to Local Detail Synthesis in Diffusion Transformers
Chaofan Gan, Zicheng Zhao, Yuanpeng Tu, Xi Chen, Ziran Qin, Tieyuan Chen, Mehrtash Harandi, Weiyao Lin
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 Detail Guidance (no CFG) | SD3-Medium · Pick-a-Pic (Cond) | HPSv2.1: 21.84 -> 28.65 | Improved |
| add Detail Guidance (no CFG) | SD3.5-Large · Pick-a-Pic (Cond) | Blipscore: 70.09 to 83.66 | Improved |
| add Detail Guidance (no CFG) | SD3.5-Large · Pick-a-Pic (Cond) | Aesthetic: 5.94 to 6.16 | Improved |
| add Detail Guidance (no CFG) | Flux-dev · Pick-a-Pic (Cond) | HPSv2.1: 19.33 -> 27.88 | Improved |
| add Detail Guidance on top of CFG | SD3-Medium · Pick-a-Pic (CFG) | Aesthetic: 5.80 -> 6.03 | Improved |
| add Detail Guidance on top of CFG | SD3-Medium · Pick-a-Pic (CFG) | Clipscore: 26.64 -> 26.25 | No clear effect |
| add Detail Guidance on top of CFG | SD3.5-Large · Pick-a-Pic (CFG) | Blipscore: 92.62 -> 91.61 | Hurt |
| add Detail Guidance on top of CFG | Flux-dev · Pick-a-Pic (CFG) | HPSv2.1: 29.16 -> 29.25 | Improved |
| use DG instead of CFG | SD3-Medium · Pick-a-Pic | Aesthetic: 5.80 -> 6.01 | Improved |
| use DG instead of CFG | SD3-Medium · Pick-a-Pic | Clipscore: 26.64 -> 24.15 | Hurt |
| combine CFG with DG vs CFG-Zero | SD3-Medium · HPSv2.1 benchmark | HPSv2.1 avg: 30.57 -> 30.96 | Improved |
| disrupt massive activations in the model (ablation) | SD3 · Win probability vs original | HPSv2.1: 0.028 win probability | Hurt |
Closest papers by meaning
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (NeurIPS 2025 · similarity 0.89)
- Exploring Diffusion Transformer Designs via Grafting (NeurIPS 2025 · similarity 0.88)
- Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers (NeurIPS 2026 · similarity 0.87)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.86)
- Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.86)
- 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.
