TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing

TailSteer. Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing. By Ruikang Li and eight others, from the Chinese University of Hong Kong, SenseTime Research, and one other institution. Published on the fourth of November.

An image editor can be excellent at a single edit and still fall apart over a conversation. In multi-turn editing, each output becomes the input to the next request, and the same image is reused again and again. Small unintended changes that nobody would notice once begin to pile up, and after enough turns they accumulate into severe visual drift. Strong single-turn performance, in other words, does not guarantee reliable multi-turn interaction.

The usual remedies each carry a cost. One is to train on multi-turn data or on cross-turn rollouts, which can alter the pretrained priors the editor already has. The other is a specialized inference-time mechanism, and those are difficult to scale to foundation models.

TailSteer is offered as a final-stage training framework: something applied at the end, meant to improve multi-turn robustness while keeping the generative and editing capabilities earned in earlier training. Its principle is simple. Leave the early part of denoising alone, keeping its dynamics close to those of the base model, and concentrate preservation corrections near the tail of the trajectory. The tail is where fine-grained deviations can be suppressed, and fine-grained deviations are exactly what compounds into drift.

The corrections are learned from the editor's own on-policy denoising rollouts. Rather than relying on curated ground-truth editing pairs, which TailSteer does not require, it uses the unchanged source content as a preservation reference. Whatever part of the image an edit was not supposed to touch serves as the target to hold onto. Because the early dynamics stay near the base model, the broad behaviour of the editor is left in place, and the training effort is spent at the end of the trajectory.

Experiments span three model families: FLUX, SenseNova, and Qwen. Across them, TailSteer substantially reduces cumulative drift over fifteen to twenty editing turns. At the same time, the base models largely keep their performance on extensive text-to-image benchmarks and single-image editing benchmarks, which is the other half of the claim. A fix for drift that cost the editor its original skill would not count.

The authors read these results as showing that long-horizon editing robustness can be unlocked in a foundation model after the fact, with TailSteer, rather than by relearning capabilities the model already had. The design choice carrying that argument is where the correction lives: at the tail of the trajectory, not spread through all of it.