Journal article · 2026 · Open access
Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers
IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, 1–17 · Published 11 November 2026 · DOI 10.1109/tpami.2026.3752263
In brief
“Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers” is a 2026 journal article by Yi Li and 7 co-authors, published in IEEE Transactions on Pattern Analysis and Machine Intelligence. OpenAlex counts 0 citations as of 18 November 2026. It is open access under a CC BY licence: the full text is free to read at IEEE Transactions on Pattern Analysis and Machine Intelligence.
- Citations
- 0
- Authors
- 8
Abstract
Training-free style transfer on Multimodal Diffusion Transformers (MM-DiT) remains challenging due to the persistent gap between trained adapters and injection-based training-free methods. Existing training-free methods typically perform forward feature injection, and thus struggle to reproduce the rich painterly textures achieved by trained adapters, often leading to only faint hue shifts. In this paper, we trace this limitation to the absence of backpropagation-induced spatial coupling and propose Hyperion, a fully training-free style-transfer pipeline for MM-DiT. At the optimization level, the pipeline introduces Adam-in-loop, which iteratively optimizes the diffusion latent at each denoising step with a frozen MM-DiT. This process recovers the gradient-driven coupling effect produced through the self-attention value path, where neighboring tokens receive highly similar updates and form coherent brushstroke textures. At the distribution level, the pipeline introduces Sinkhorn Latent Transport (SLT), a single closed-form optimal-transport step that pre-aligns latent color distributions before Adam optimization, improving the operational efficiency of Adam-in-loop. To further enhance stylization quality, we develop a frequency-domain Sinkhorn loss to sharpen painterly textures and a layout-residual-suppression step to reduce layout interference. The pipeline can also be extended to image-to-image stylization. Extensive experiments on a 39-style × 6-prompt benchmark (234 pairs) demonstrate that Hyperion achieves CSD $= 0.640$ and CLIP-T $= 0.296$, improving CSD by +5.2% over Adam-8 at $2.5\times$ lower forward-pass cost and by +16.6% over Adam-only under an equal compute budget. Notably, Hyperion outperforms four pre-trained style adapters on 19 of 21 diverse style families, achieving a mean CSD of 0.657 versus 0.506 for the strongest baseline.
Cite this paper
Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.
APA 7
Li, Y., Li, C., Fang, Z., Zhang, J., Wang, X., Huang, H., Luo, Y., & Guo, Y. (2026). Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, 1–17. https://doi.org/10.1109/tpami.2026.3752263
MLA 9
Li, Yi, et al. “Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers.” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PP, 2026, pp. 1–17. https://doi.org/10.1109/tpami.2026.3752263.
Chicago
Li, Yi, Chengyan Li, Zhirui Fang, Junchuan Zhang, Xuan Wang, Huaibo Huang, Yan Luo, and Yanqing Guo. 2026. “Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers.” IEEE Transactions on Pattern Analysis and Machine Intelligence PP: 1–17. https://doi.org/10.1109/tpami.2026.3752263.
Harvard
Li, Y. et al. (2026) ‘Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers’, IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, pp. 1–17. doi:10.1109/tpami.2026.3752263.
IEEE
Y. Li et al., “Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PP, pp. 1–17, 2026, doi: 10.1109/tpami.2026.3752263.
BibTeX
@article{li2026bridging,
title = {{Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers}},
author = {Li, Yi and Li, Chengyan and Fang, Zhirui and Zhang, Junchuan and Wang, Xuan and Huang, Huaibo and Luo, Yan and Guo, Yanqing},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
year = {2026},
volume = {PP},
pages = {1--17},
doi = {10.1109/tpami.2026.3752263},
url = {https://doi.org/10.1109/tpami.2026.3752263}
}
Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)
Topic and keywords
Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
- latent optimization
- optimal transport
- frequency domain loss
Work with this paper
Most-cited papers in Computer Science
- ImageNet classification with deep convolutional neural networksAlex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton · Communications of the ACM · 2017 · 114,726 citations
- Long Short-Term MemorySepp Hochreiter, Jürgen Schmidhuber · Neural Computation · 1997 · 109,097 citations
- A Mathematical Theory of CommunicationClaude E. Shannon · Bell System Technical Journal · 1948 · 91,736 citations
- Deep learningYann LeCun, Yoshua Bengio, Geoffrey E. Hinton · Nature · 2015 · 90,226 citations
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal NetworksShaoqing Ren, Kaiming He, Ross Girshick and 1 more · IEEE Transactions on Pattern Analysis and Machine Intelligence · 2016 · 74,143 citations
Questions about this paper
How do I cite this paper in APA style?
Li, Y., Li, C., Fang, Z., Zhang, J., Wang, X., Huang, H., Luo, Y., & Guo, Y. (2026). Bridging the Training Gap with Hyperion: A Training-Free Style Transfer Method on Diffusion Transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, PP, 1–17. https://doi.org/10.1109/tpami.2026.3752263
How many times has this paper been cited?
0 times, according to OpenAlex on 18 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.
Is this paper free to read?
Yes. It is open access at IEEE Transactions on Pattern Analysis and Machine Intelligence: sign in and open the full text from this page.
Who wrote this paper?
Yi Li, Chengyan Li, Zhirui Fang, Junchuan Zhang, Xuan Wang, Huaibo Huang, Yan Luo, Yanqing Guo.
What is the DOI of this paper?
10.1109/tpami.2026.3752263, permanently at https://doi.org/10.1109/tpami.2026.3752263. Sign in and open it from this page.
Bibliographic data from OpenAlex (CC0), last checked 18 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.