# 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  
**Lab:** MMLab-CUHK · **Published:** 4 Nov 2026 · **Upvotes:** 39 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.06741) · [PDF](https://arxiv.org/pdf/2611.06741) · [Code](https://github.com/Lyricccco/tailsteer) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.06741.html)

## Curator's notes

**TL;DR:** TailSteer adds a final-stage training step that suppresses drift in multi-turn image editing while keeping base model abilities.

- **What's new:** It learns preservation corrections from on-policy rollouts without curated pairs, focusing on the denoising tail rather than altering early dynamics.
- **Key result:** TailSteer reduces cumulative drift over 15‑20 editing turns across FLUX, SenseNova, and Qwen models while preserving base performance.
- **Why it matters:** Builders can create longer interactive editing sessions without the images degrading over successive edits.

## Abstract

Strong single-turn editing performance does not guarantee reliable multi-turn interaction: as edited outputs are repeatedly reused, small unintended changes can accumulate into severe visual drift. This long-term drift is often mitigated through training on multi-turn data or cross-turn rollouts, potentially altering pretrained priors, or through specialized inference-time mechanisms that are difficult to scale to foundation models. We introduce TailSteer, a final-stage training framework that improves multi-turn robustness while retaining the generative and editing capabilities acquired through earlier training. TailSteer follows a simple principle: keep early denoising dynamics close to those of the base model and concentrate preservation corrections near the trajectory tail, where fine-grained deviations can be suppressed. These corrections are learned from the editor's own on-policy denoising rollouts, using unchanged source content as a preservation reference without requiring curated ground-truth editing pairs. Experiments across the FLUX, SenseNova, and Qwen families show that TailSteer substantially reduces cumulative drift over 15-20 editing turns while largely preserving base-model performance on extensive text-to-image and single-image editing benchmarks. These results demonstrate that long-horizon editing robustness can be unlocked in foundation models after TailSteer, rather than by relearning their existing capabilities.

## What you could build

1. Create a multi-step image editing web app that maintains visual consistency across many user prompts.
2. Develop a plugin for existing diffusion editors that adds TailSteer fine‑tuning for drift‑resistant workflows.
3. Build an automated storyboard generator that iteratively refines scenes without losing detail.

## Quiz yourself

What part of the denoising process does TailSteer modify to reduce drift?

**Answer:** It concentrates preservation corrections near the trajectory tail, keeping early denoising dynamics close to the base model.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.