[ICLR 2026] Routing Matters in MoE: Scaling Diffusion Transformers with Explicit Routing Guidance

Yujie Wei¹, Shiwei Zhang^2*, Hangjie Yuan³, Yujin Han⁴, Zhekai Chen^4,5, Jiayu Wang², Difan Zou⁴, Xihui Liu^4,5, Yingya Zhang², Yu Liu², Hongming Shan^1†
(*Project Leader, †Corresponding Author)

¹Fudan University ²Tongyi Lab, Alibaba Group ³Zhejiang University ⁴The University of Hong Kong ⁵MMLab

ProMoE is an MoE framework that employs a two-step router with explicit routing guidance to promote expert specialization for scaling Diffusion Transformers.

🤗 Overview

This codebase supports:

• Baselines: Dense-DiT, TC-DiT, EC-DiT, DiffMoE, and their variants.
• Proposed Models: ProMoE variants (S, B, L, XL) with Token-Choice (TC) and Expert-Choice (EC) routing.
• VAE Latent Preprocessing: Pre-encode raw images into latents and cache them for faster training; supports multi-GPU parallel processing.
• Sampling and Metric Evaluation: Image sampling, Inception feature extraction, and calculation of FID, IS, sFID, Precision, and Recall; supports multi-GPU parallel processing.

🔥 Updates

• [2026.03]: Release the model weights, generated 50K images, and evaluation results on both Hugging Face and ModelScope!
• [2026.02]: Release the training, sampling, and evaluation code of ProMoE.
• [2026.01]: 🎉 Our paper has been accepted by ICLR 2026!
• [2025.10]: Release the paper of ProMoE.

🗂️ Pretrained Models

We have released the model weights, generated 50K images, and corresponding evaluation results on both Hugging Face and ModelScope.

Model	Platform	Weights	50K Images	Eval Results (CFG=1.0)	Eval Results (CFG=1.5)
ProMoE-B-Flow (500K)	Hugging Face ModelScope	Link Link	cfg=1.0, cfg=1.5 cfg=1.0, cfg=1.5	FID 24.44, IS 60.38 FID 24.44, IS 60.38	FID 6.39, IS 154.21 FID 6.39, IS 154.21
ProMoE-L-Flow (500K)	Hugging Face ModelScope	Link Link	cfg=1.0, cfg=1.5 cfg=1.0, cfg=1.5	FID 11.61, IS 100.82 FID 11.61, IS 100.82	FID 2.79, IS 244.21 FID 2.79, IS 244.21
ProMoE-XL-Flow (500K)	Hugging Face ModelScope	Link Link	cfg=1.0, cfg=1.5 cfg=1.0, cfg=1.5	FID 9.44, IS 114.94 FID 9.44, IS 114.94	FID 2.59, IS 265.62 FID 2.59, IS 265.62

⚙️ Preparation

1. Requirements & Installation

conda create -n promoe python=3.10 -y
conda activate promoe
pip install -r requirements.txt

2. Dataset Preparation

Download ImageNet dataset, and modify cfg.data_path in config.py.

3. VAE Latent Preprocessing (Optional)

For faster training and more efficient GPU usage, you can precompute VAE latents and train with cfg.use_pre_latents=True.

Run latent preprocessing:

# bash
ImageNet_path=/path/to/ImageNet

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python preprocess/preprocess_vae.py --latent_save_root "$ImageNet_path/sd-vae-ft-mse_Latents_256img_npz"

🚀 Training

Training is launched via train.py with a YAML config:

python train.py --config configs/004_ProMoE_L.yaml

Notes:

• This repository currently supports Rectified Flow with Logit-Normal sampling (following SD3). For the DDPM implementation, please refer to this repository.
• By default, ProMoE utilizes Token-Choice routing. However, for DDPM-based training, we recommend using Expert-Choice in models/models_ProMoE_EC.py.
• Configuration files for all baseline models are provided in the configs directory.
• All results reported in the paper are obtained with qk_norm=False. For extended training steps (>2M steps), we suggest enabling qk_norm=True to ensure training stability.

💫 Sampling

Image generation is performed via the sample.py script, utilizing the same YAML configuration file used for training.

# use default setting
CUDA_VISIBLE_DEVICES=0 python sample.py --config configs/004_ProMoE_L.yaml

# use custom setting
CUDA_VISIBLE_DEVICES=0 python sample.py \
  --config configs/004_ProMoE_L.yaml \
  --step_list_for_sample 200000,300000 \
  --guide_scale_list 1.0,1.5,4.0 \
  --num_fid_samples 10000

Notes:

• By default, the script loads the checkpoint at 500k steps and generates 50,000 images using a single GPU, sweeping across guidance scales (CFG) of 1.0 and 1.5.
• To use multiple GPUs for sampling, specify the devices using CUDA_VISIBLE_DEVICES or by adding sample_gpu_ids in the configuration file. Please be aware that multi-GPU inference produces a globally different random sequence (e.g., class labels) compared to single-GPU inference.
• Generated images are saved as PNG files in the sample/ directory within the same parent directory as the checkpoint folder. Filenames include both the sample index and class label.
• If you only want to calculate FID, you can set cfg.save_inception_features=True to save Inception features and reduce cfg.save_img_num.

📝 Evaluation

We follow the standard evaluation protocol outlined in openai's guided-diffusion. All relevant code is located in the evaluation directory.

1. Environment Setup

Since the evaluation pipeline relies on TensorFlow, we strongly recommend creating a dedicated environment to avoid dependency conflicts.

conda create -n promoe_eval python=3.9 -y
conda activate promoe_eval
cd evaluation
pip install -r requirements.txt
conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0

2. Download Reference Batch

Download the reference statistics file VIRTUAL_imagenet256_labeled.npz (for 256x256 images) and place it in the evaluation directory.

3. Execution

To calculate the metrics, run the evaluation script by specifying the path to your folder of generated images.

CUDA_VISIBLE_DEVICES=0 python run_eval.py /path/to/generated/images

Acknowledgements

This code is built on top of DiffMoE, DiT, and guided-diffusion. We thank the authors for their great work.

🌟 Citation

If you find this code useful for your research, please cite our paper:

@inproceedings{wei2026promoe,
  title={Routing Matters in MoE: Scaling Diffusion Transformers with Explicit Routing Guidance},
  author={Wei, Yujie and Zhang, Shiwei and Yuan, Hangjie and Han, Yujin and Chen, Zhekai and Wang, Jiayu and Zou, Difan and Liu, Xihui and Zhang, Yingya and Liu, Yu and others},
  booktitle={International Conference on Learning Representations},
  year={2026}
}