WiT: Waypoint Diffusion Transformers via Trajectory Conflict Navigation

Official repository for the project page and future code release of:

WiT: Waypoint Diffusion Transformers via Trajectory Conflict Navigation

Project page

GitHub Pages:

• https://hainuo-wang.github.io/WiT/
• arXiv: https://arxiv.org/abs/2603.15132

Repository status

This repository is currently serving as the official public landing page for WiT. The training and inference code will be released in this same repository after the paper is formally published.

Main Result

Method	Params	Epochs	IS	FID-50K
JiT-B/16	131M	200	-	4.37
WiT-B/16	131M + 21M	200	270.7	3.34
JiT-B/16	131M	600	275.1	3.66
WiT-B/16	131M + 21M	600	280.2	3.03
JiT-L/16	459M	200	-	2.79
LF-DiT-L/16	465M	200	-	2.48
WiT-L/16	459M + 21M	200	289.1	2.38
JiT-L/16	459M	600	298.5	2.36
WiT-L/16	459M + 21M	265	293.7	2.36
WiT-L/16	459M + 21M	600	303.3	2.22
WiT-XL/16	676M + 21M	200	288.9	2.16
WiT-XL/16	676M + 21M	600	301.0	1.89

Summary

WiT mitigates trajectory conflict in pixel-space flow matching by:

• projecting pretrained vision features into low-dimensional semantic waypoints,
• predicting those waypoints from the current noisy pixel state,
• injecting them into the pixel diffusion transformer with Just-Pixel AdaLN.

On ImageNet 256x256, WiT reports:

• 2.27x faster convergence than JiT-L/16 at matched FID 2.36,
• 1.89 FID and 301.0 IS for WiT-XL/16 at 600 epochs,
• 1.62x more stable peak pairwise trajectory conflict than JiT.

Acknowledgments

We thank Qiming Hu for insightful discussions and feedback. This work was partially supported by computational resources from TPU Research Cloud (TRC).

Citation

@article{wang2026wit,
  title={WiT: Waypoint Diffusion Transformers via Trajectory Conflict Navigation},
  author={Wang, Hainuo and Li, Mingjia and Guo, Xiaojie},
  journal={arXiv preprint arXiv:2603.15132},
  year={2026}
}