ScaleResfusion: Residual Rectified Flow based on Residual Vector Field

Zhenning Shi</strong>, Chen Xu, Junhao Zhang, Kefei Zhang, Linjie Liu, Zhedong Zheng, Tao Li

arXiv / codes / video

• Proposes ScaleResfusion, a scalable Real-IR framework that starts sampling from noisy LQ images while remaining consistent with pre-trained text-to-image Rectified Flow models, overcoming the from-scratch training and scheduler-bound objectives of previous residual diffusion methods.
• Introduces the marginal-independent residual vector field $resv = \epsilon - x_0 + \gamma R = v_{\mathrm{RF}} + \gamma R$, which preserves the linear transport and marginal distributions of standard Rectified Flow, reducing adaptation to a compact residual offset that can be efficiently learned with LoRA.
• Builds a LoRA-based knowledge-distillation pipeline that scales residual restoration to billion-parameter backbones—from SD3 (2B) to FLUX2 (9B)—and achieves state-of-the-art performance on multiple Real-IR tasks with only four sampling steps.

LiveMoments: Reselected Key Photo Restoration in Live Photos via Reference-guided Diffusion

Clara Xue, Zizheng Yan, Zhenning Shi, Yuhang Yu, Jingyu Zhuang, Qi Zhang, Jinwei Chen, Qingnan Fan.

arXiv / codes / project page / video

• The first to address the problem of reselected key photo restoration in Live Photos.
• LiveMoments significantly improves perceptual quality and fidelity over existing solutions, including the recent flagships from vivo and iPhone.

Ultra-High-Definition Reference-Based Landmark Image Super-Resolution with Generative Diffusion Prior

Zhenning Shi, Zizheng Yan, Yuhang Yu, Clara Xue, Jingyu Zhuang, Qi Zhang, Jinwei Chen, Tao Li, Qingnan Fan

arXiv / codes

• The first to address ultra-high-definition (UHD) reference-based landmark image super-resolution under real-world degradations with a diffusion-based RefSR pipeline.
• Proposes TriFlowSR, a three-branch diffusion framework with Patch-Ref Attention to explicitly match LR and reference-HR features at the patch level, enabling better transfer of semantic and texture details and reducing artifacts.
• Introduces Landmark-4K, the first UHD RefSR dataset for landmarks (185 high-quality images, 49 categories; avg. resolution ~ $3295 \times 3295$.

Resfusion: Denoising diffusion probabilistic models for image restoration based on prior residual noise

Zhenning Shi, Haoshuai Zheng, Chen Xu, Changsheng Dong, Bin Pan, Xueshuo Xie, Along He, Tao Li, Huazhu Fu

arXiv / codes

• Start diffusion-based restoration process from noisy degraded images (not pure Gaussian noise) by injecting the residual term into the forward diffusion process, reducing unnecessary sampling.
• Proposes Resfusion, a general restoration framework that defines a weighted residual noise target (“resnoise”) and explicitly derives the quantitative relationship between the residual and noise terms, keeping the reverse process DDPM-consistent.
• Demonstrates competitive results with only five sampling steps, without requiring any physics-based priors.