Research map / Architectures and efficiency / Mixture of experts
Routing Matters in MoE: Scaling Diffusion Transformers with Explicit Routing Guidance
Yujie Wei, Shiwei Zhang, Hangjie Yuan, Yujin Han, Zhekai Chen, Jiayu Wang, Difan Zou, Xihui Liu, Yingya Zhang, Yu Liu, Hongming Shan
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 |
|---|---|---|---|
| use ProMoE two-step prototypical MoE instead of dense DiT at equal activated parameters | DiT-B/L/XL, Rectified Flow · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 9.02 -> 6.39 (B), 3.56 -> 2.79 (L), 3.23 -> 2.59 (XL) | Improved |
| use ProMoE under the DDPM objective | DiT-B DDPM · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 18.61 -> 17.90 | Improved |
| existing DiT-MoE token-choice routing instead of dense | DiT-MoE-B-Flow · ImageNet 256x256, 500K steps, CFG=1.5 | FID50K: 9.02 -> 8.94 | No clear effect |
| add semantic routing guidance via classification loss or k-means routing (ablation) | MoE DiT-B-Flow · ImageNet 256x256, 500K steps, CFG=1.5 | FID50K: 8.22 (DiffMoE) -> 5.91 (classification), 6.24 (k-means) | Improved |
| add prototypical routing (ablation) | Dense-DiT-B-Flow · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 9.02 -> 7.92 | Improved |
| add routing contrastive learning (ablation) | ProMoE-B-Flow · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 7.92 -> 6.75 | Improved |
| add conditional routing with unconditional experts (ablation) | ProMoE-B-Flow · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 6.75 -> 6.39 | Improved |
| add conditional routing to K-means routing (ablation) | K-means-based routing · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 8.75 -> 6.24 | Improved |
| use identity score activation instead of softmax or sigmoid (ablation) | prototypical routing · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 6.92 (softmax) -> 6.39 (identity) | Improved |
| add conventional load-balancing loss on top of RCL (ablation) | ProMoE-B-Flow · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 6.39 -> 6.53 | Hurt |
| raise RCL loss weight from 1 to 10 (ablation) | ProMoE-B-DDPM · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 18.34 -> 17.90 | No clear effect |
| increase activated routed experts from 1 to 3 (ablation) | ProMoE-L-Flow · ImageNet 256x256, 500K steps | FID50K (CFG=1.5): 2.79 -> 2.72 | Improved |
Closest papers by meaning
- Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts (ICML 2025 · similarity 0.93)
- EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing (ICLR 2025 · similarity 0.92)
- Diff-MoE: Diffusion Transformer with Time-Aware and Space-Adaptive Experts (ICML 2025 · similarity 0.89)
- Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE (NeurIPS 2026 · similarity 0.89)
- Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.88)
- ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation (NeurIPS 2025 · similarity 0.88)
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.
