arXiv 2611.06741
TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing
AI35 upvotes
Authors: Ruikang Li, Zixuan Chen, Jiarui Wu, Mingde Yao, Yiyuan Zhang, Yujin Wang, Hanming Deng, Lewei Lu +1
Overview
The authors propose TailSteer, a final-stage training framework that makes image editors more robust across repeated editing turns. It keeps early denoising dynamics close to the base model and learns preservation corrections near the trajectory tail, using the editor's own on-policy rollouts with unchanged source content as a reference.
Introduction
Small unintended changes in an edited image accumulate when the output is reused across many editing turns, causing severe visual drift. Existing fixes either retrain on multi-turn data (which can alter pretrained priors) or rely on inference-time mechanisms that are hard to scale to foundation models.
Results
TailSteer substantially reduces cumulative drift over 15-20 editing turns across the FLUX, SenseNova, and Qwen model families. It does so while largely preserving base-model performance on text-to-image and single-image editing benchmarks.
Conclusion
This shows that long-horizon editing robustness can be added to existing foundation models without relearning their core capabilities, which matters for practical multi-turn editing.