CR-Diff: Cross-Resolution Diffusion Models via Network Pruning

Ren, Jiaxuan; Zhu, Junhan; Wang, Huan

CR-Diff: Cross-Resolution Diffusion Models via Network Pruning

Accepted by CVPR 2026 Findings

Jiaxuan Ren², Junhan Zhu¹, Huan Wang^1*

¹Westlake University, ²University of Electronic Science and Technology of China

^*Corresponding author: wanghuan [at] westlake [dot] edu [dot] cn

Paper Code arXiv

This paper proposes CR-Diff, a pruning-based method to improve cross-resolution consistency in UNet-based diffusion models. Pruning is used to mask certain parameters, helping the model generalize to unseen resolutions. Compared to the original SDXL trained at 1024×1024, which doesn't perform well on other resolutions, the pruned model produces more coherent results at sizes like 400×560 and 480×360. This suggests that some parameters may hinder generalization, and removing them can effectively improve performance across resolutions.

Overview of our CR-Diff method

Illustration of the proposed CR-Diff framework for improving cross-resolution generation in UNet-based diffusion models. The method follows a two-stage pruning and optimizing paradigm. First, target UNet blocks are assigned adaptive sparsity levels via a block-wise pruning strategy, where magnitude-based criteria are applied with differentiated ratios across downsampling, middle, and upsampling modules to extract resolution-stable parameter subsets. Subsequently, a pruned output amplification mechanism refines the denoising process by amplifying the pruned subnetwork outputs with a coefficient k > 1, effectively suppressing adverse contributions from the dense model. This process leads to more stable denoising trajectories and produces higher-quality images with improved structural coherence and semantic alignment across resolutions.

Quantitative Results

Performance comparison on unseen resolutions shows that CR-Diff consistently outperforms the dense baseline across most models and resolution settings. Metrics where CR-Diff achieves better results are highlighted in bold, demonstrating its overall advantage.

Performance comparison at default resolutions shows that CR-Diff either improves or preserves performance across the evaluated models. Bold values denote cases where it outperforms the dense baseline.

Qualitative Results

On a 5K MS-COCO subset, consistent gains of CR-Diff over dense SDXL in ImageReward and visual fidelity across resolutions.

On a 5K MS-COCO subset, consistent gains of CR-Diff over dense SD1.5 in ImageReward and visual fidelity across resolutions.

On a 5K MS-COCO subset, consistent gains of CR-Diff over dense SD2.1 in ImageReward and visual fidelity across resolutions.

BibTeX

@misc{ren2026crdiff,
      title={Cross-Resolution Diffusion Models via Network Pruning}, 
      author={Jiaxuan Ren and Junhan Zhu and Huan Wang},
      year={2026},
      eprint={2604.05524},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.05524},       
    }