We propose LCUDiff, a stable one-step framework for human body restoration that upgrades latent diffusion from the 4-channel latent space to a 16-channel latent space. The method uses channel splitting distillation and prior-preserving adaptation to improve fidelity while keeping one-step efficiency, and further introduces a decoder router for per-sample decoder routing under diverse restoration conditions.
@article{gong2026lcudiff,title={{LCUDiff}: Latent Capacity Upgrade Diffusion for Faithful Human Body Restoration},status={preprint},author={Gong, Jue and Zhou, Zihan and Wang, Jingkai and Li, Shu and Liu, Libo and Lan, Jianliang and Zhang, Yulun},journal={arXiv preprint arXiv:2602.04406},year={2026},month=feb,}
We propose Light Up Your Face, an end-to-end universal and prompt-controllable face relighting framework built on one-step diffusion and identity-preserving guidance. The method supports diverse illumination edits while retaining identity consistency and fine facial details, and generalizes across in-the-wild portrait conditions.
@article{gong2026lightup,title={{Light Up Your Face}: Towards End-to-End Universal and Prompt Controllable Relighting},status={preprint},author={Gong, Jue and Chen, Zheng and Wang, Jingkai and Zhang, Yulun and Yang, Xiaokang},journal={arXiv preprint arXiv:2602.04300},year={2026},month=feb,}
SODiff improves JPEG artifact removal with semantic-oriented guidance for a pre-trained one-step diffusion model. It introduces a semantic-aligned image prompt extractor and a quality factor-aware time predictor to better handle varying compression levels.
@inproceedings{yang2026sodiff,title={{SODiff}: Semantic-Oriented Diffusion Model for {JPEG} Compression Artifacts Removal},status={accepted},author={Yang, Tingyu and Gong, Jue and Guo, Jinpei and Li, Wenbo and Guo, Yong and Zhang, Yulun},booktitle={The AAAI Conference on Artificial Intelligence},year={2026},month=feb,}
HAODiff targets practical human-centric restoration under mixed degradation and human motion blur. It introduces triple-branch dual-prompt guidance and a dedicated MPII-Test benchmark, enabling stronger robustness and better visual quality in a single diffusion step.
@article{gong2025haodiff,title={{HAODiff}: Human-Aware One-Step Diffusion via Dual-Prompt Guidance},status={accepted},author={Gong, Jue and Yang, Tingyu and Wang, Jingkai and Chen, Zheng and Liu, Xing and Gu, Hong and Zhang, Yulun and Yang, Xiaokang},journal={Advances in Neural Information Processing Systems},year={2025},month=dec,}
This work introduces the PERSONA benchmark for human body restoration and proposes OSDHuman, a one-step diffusion model guided by a high-fidelity image embedder. The method improves restoration quality while addressing the lack of large-scale, task-specific datasets for human-centric restoration.
@inproceedings{gong2025osdhuman,title={Human Body Restoration with One-Step Diffusion Model and a New Benchmark},status={accepted},author={Gong, Jue and Wang, Jingkai and Chen, Zheng and Liu, Xing and Gu, Hong and Zhang, Yulun and Yang, Xiaokang},booktitle={Proceedings of the International Conference on Machine Learning},year={2025},month=jul,}
OSDFace is a one-step diffusion model for face restoration that combines a visual representation embedder, identity-aware supervision, and GAN-based guidance. It improves realism and identity consistency while keeping inference efficient.
@article{wang2025osdface,title={{OSDFace}: One-Step Diffusion Model for Face Restoration},status={accepted},author={Wang, Jingkai and Gong, Jue and Zhang, Lin and Chen, Zheng and Liu, Xing and Gu, Hong and Liu, Yutong and Zhang, Yulun and Yang, Xiaokang},journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},year={2025},month=jun,}