TurboClear

One-Step Object-Effect Removal via Region-Calibrated Distribution Matching and Fusion

Jiawei Guo¹, Junxian Li¹, Yixin Tang¹, Bingya Zhang², Jiaxin Lu², Yulun Zhang^1,†, Shangchen Zhou^3,†
1 Shanghai Jiao Tong University    ² Honor Device Co., Ltd    ³ Imperial College London
^† Corresponding authors

TurboClear removes target objects and their associated visual effects with a single denoising step while preserving unaffected background regions.

Highlights

• One-step removal. TurboClear performs object-effect removal with one SDXL UNet evaluation.
• Region-Calibrated Distribution Matching. RDM calibrates the distillation signal across affected and preserved regions.
• Learnable Spatial Fusion. LSF combines removal-oriented generation with an identity-preserving reference stream.
• Efficient inference. TurboClear substantially reduces denoising computation while maintaining competitive visual quality.

Overview

TurboClear addresses the spatial asymmetry of object-effect removal: regions influenced by the target object should be regenerated, while the remaining image should stay unchanged. RDM provides region-aware supervision during one-step distillation, and LSF learns spatial gates for lightweight fusion at inference time.

Release Status

Resource	Status
Paper	arXiv:2608.01288
Model weights	TBD
Code	TBD
Usage guide	TBD

Citation

@article{guo2026turboclear,
  title={TurboClear: One-Step Object-Effect Removal via Region-Calibrated Distribution Matching and Fusion},
  author={Guo, Jiawei and Li, Junxian and Tang, Yixin and Zhang, Bingya and Lu, Jiaxin and Zhang, Yulun and Zhou, Shangchen},
  journal={arXiv preprint arXiv:2608.01288},
  year={2026}
}

Acknowledgements

TBD