TailSteer targets visual drift in multi-turn image editing
Ruikang Li and eight other authors submitted the TailSteer paper to arXiv. The proposed final-stage training framework keeps early denoising dynamics close to those of the base model and concentrates preservation corrections near the trajectory tail. These corrections are learned from the editor’s own on-policy denoising rollouts, using unchanged source content as a preservation reference. The authors report reduced cumulative drift over 15–20 editing turns across the FLUX, SenseNova, and Qwen model families, while largely preserving performance on text-to-image and single-image editing benchmarks.
The authors report that TailSteer improves long-horizon editing robustness by reducing cumulative visual drift while retaining generative and editing capabilities acquired through earlier training.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.06741 (cs)
[Submitted on 4 Nov 2026]
Title:TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing
Authors:Ruikang Li, Zixuan Chen, Jiarui Wu, Mingde Yao, Yiyuan Zhang, Yujin Wang, Hanming Deng, Lewei Lu, Tianfan Xue
View a PDF of the paper titled TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing, by Ruikang Li and 8 other authors