RoSeDiT: Role-Separated Diffusion Transformers
Checkpoints of Role-Separated Diffusion Transformers: Global Rule Isolation and Deterministic Sampling Manifold Compression (Zhiguang Liu, Yi Shang; University of Missouri). Paper: https://arxiv.org/abs/2611.19009. Code: https://github.com/lz7fd/RoSeDiT
RoSeDiT is a diffusion transformer architecture that isolates the global rule from the image tokens. Twelve controller tokens maintain an evolving representation of the global condition, a listening mechanism lets adaLN compute its modulation from their states, and a structural pass restricts each image token to its 3x3 neighborhood and the controller tokens. RoSeDiT maintains the same level of parameter count as SiT by reusing the dense attention weights. Deterministic Sampling Manifold Compression (DSMC) is a two-stage training schedule for the continued training of diffusion models.
Files
Class-conditional ImageNet 256x256 models in the latent space of the Stable Diffusion VAE. FID is computed on 50,000 samples from the EMA weights with a 250-step Euler-Maruyama SDE sampler and no classifier-free guidance (ADM evaluation suite).
| File | Model | Steps | FID | Params | --model |
Flags | Size |
|---|---|---|---|---|---|---|---|
RoSeDiT-S-2_400K.pt |
RoSeDiT-S/2 | 400K | 51.42 | 33.0M | SiT-BR-B1W-DG-S/2 |
--encoder-depth=4 |
0.12 GB |
RoSeDiT-B-2_400K.pt |
RoSeDiT-B/2 | 400K | 27.18 | 130.6M | SiT-BR-B1W-DG-B/2 |
--encoder-depth=4 |
0.49 GB |
RoSeDiT-L-2_400K.pt |
RoSeDiT-L/2 | 400K | 15.34 | 458.2M | SiT-BR-B1W-DG-L/2 |
--encoder-depth=8 |
1.71 GB |
RoSeDiT-XL-2_400K.pt |
RoSeDiT-XL/2 | 400K | 12.63 | 675.3M | SiT-BR-B1W-DG-XL/2 |
--encoder-depth=8 --qk-norm |
2.52 GB |
RoSeDiT-XL-2_1400K.pt |
RoSeDiT-XL/2 | 1.4M | 7.98 | 675.3M | SiT-BR-B1W-DG-XL/2 |
--encoder-depth=8 --qk-norm |
2.52 GB |
RoSeDiT-XL-2_2400K.pt |
RoSeDiT-XL/2 | 2.4M | 7.61 | 675.3M | SiT-BR-B1W-DG-XL/2 |
--encoder-depth=8 --qk-norm |
2.52 GB |
RoSeDiT-XL-2_DSMC_2400K.pt |
RoSeDiT-XL/2 + DSMC | 2.4M | 5.84 | 675.3M | SiT-BR-B1W-DG-XL/2 |
--encoder-depth=8 --qk-norm |
2.52 GB |
RoSeDiT-B-2_REPA_400K.pt |
RoSeDiT-B/2 + REPA | 400K | 21.01 | 130.6M + 7.3M projector | SiT-BR-B1W-DG-B/2 |
--encoder-depth=4 --projector-embed-dims=768 |
0.51 GB |
Each file is a PyTorch dictionary with the EMA weights under the key ema and the settings under config. The files
hold no optimizer state: they are for sampling, not for resuming training. MD5SUMS lists the checksums.
Use
git clone https://github.com/lz7fd/RoSeDiT && cd RoSeDiT
pip install -r requirements.txt
hf download lz7fdmu/RoSeDiT --local-dir checkpoints
torchrun --nnodes=1 --nproc_per_node=1 generate.py \
--model "SiT-BR-B1W-DG-XL/2" --qk-norm \
--num-fid-samples 50000 \
--ckpt checkpoints/RoSeDiT-XL-2_DSMC_2400K.pt \
--path-type=linear --encoder-depth=8 --projector-embed-dims=None \
--per-proc-batch-size=64 --mode=sde --num-steps=250 --cfg-scale=1.0 --global-seed=0
scripts/sample_50k.sh <file> in the code repository picks the flags of each file. RoSeDiT-XL/2 is trained with RMS
normalization of the attention queries and keys, so --qk-norm is needed to load it. The 50,000-sample archives behind
RoSeDiT-XL/2 at 1.4M steps (FID 7.98) and RoSeDiT-XL/2 + DSMC at 2.4M steps (FID 5.84) are at
https://huggingface.co/datasets/lz7fdmu/RoSeDiT_samples
Citation
@article{liu2026rosedit,
title = {Role-Separated Diffusion Transformers: Global Rule Isolation and Deterministic Sampling Manifold Compression},
author = {Liu, Zhiguang and Shang, Yi},
journal = {arXiv preprint arXiv:2611.19009},
year = {2026}
}
License
MIT. The code is built on REPA.