TailSteer targets the final stages of AI image editing to reduce cumulative drift

An AI image editor may successfully add an object while also slightly changing a face, background texture or scene color. When its output becomes the input for another edit, those unintended changes can accumulate. A successful one-off edit therefore does not guarantee a stable sequence of revisions.

The preprint TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing, submitted to arXiv on November 4, 2026, by Ruikang Li and colleagues, addresses this cumulative visual drift. The authors report improved preservation across four models from the FLUX, SenseNova and Qwen families. Their results also show an important distinction: keeping an image stable and correctly executing instructions are separate challenges.

Why focus on the end of an edit?

The models studied generate their outputs through a sequence of denoising steps. The paper describes this process as moving from coarse structure toward fine detail. Its trajectory analysis finds that broad image structure is already similar to the source before the final steps, while finer details recover mainly near the end. Unintended differences in regions that should remain unchanged nevertheless persist at the endpoint.

TailSteer concentrates preservation-oriented training in this late portion, or trajectory tail. Earlier in the process, the refined model is trained to match the original model’s predictions. Later, its training target gradually shifts toward an endpoint designed to preserve unchanged source content.

A preliminary experiment illustrates why both parts matter. Training only the final sampling step improved preservation when that modified model was used solely for the last update. Applying it throughout sampling instead produced desaturation and detail loss. The full method explicitly trains earlier behavior as well, allowing one refined model to handle the complete process.

Learning from the editor’s own outputs

Training examples come from two kinds of requests: instructions to leave an image unchanged, and instructions to make a localized edit. The original editor generates outputs, which are filtered for reliable preservation regions.

For a local edit, the researchers construct a target that combines unchanged source content outside the edited region with the editor’s generated modification inside it. This avoids requiring a manually prepared ideal output for every instruction. It also lets the editor retain its particular realization of the requested change.

During training, the refined model generates short denoising rollouts from cached states. Supervision is recalculated at the states it actually reaches, rather than relying entirely on the original model’s trajectory. Although the paper calls its approach “data-free” in one section, it still uses source images and instructions; the avoided requirement is curated ground-truth editing pairs or cross-turn training sequences.

The training machinery does not carry over into ordinary inference. The refined model runs the full sampling trajectory without the frozen reference model, preservation masks or model switching.

What the experiments found

The researchers evaluated FLUX.1 Kontext, FLUX.2 [klein] Base 9B, SenseNova-U1.5-8B-MoT and Qwen-Image-2.1. Each model received 2,000 training updates using eight NVIDIA H100 GPUs. Qualitative examples include editing sequences extending to 20 turns, plus multi-image compositing and pose-guided editing.

For a more controlled test, the authors introduce Add & Preserve: 100 natural images, each with a boxed region receiving ten successive object-addition or removal instructions. Content outside the box should remain unchanged. Separate measurements track outside-box drift and whether objects inside the box match the editing history.

For FLUX.2 [klein], one TailSteer configuration increased turn-ten structural similarity, or SSIM, from 0.289 to 0.621. Its perceptual difference score, LPIPS, fell from 0.524 to 0.290, where lower indicates less difference. Cumulative Object Accuracy changed much less, from 31.02% to 31.84%. This illustrates the main reported gain: better preservation, not a comparable improvement in completing edits.

Where the evidence stops

Results on broader editing benchmarks were mixed. Content consistency improved across the tested TailSteer settings on EdiVal, but instruction following did not improve consistently. Text-to-image and single-turn editing scores generally stayed close to the original models, although some declined.

The comparison with GPT-Image-2.5 Sunburst also needs qualification. TailSteer-Klein exceeded it on the three reported outside-box preservation metrics, but its object accuracy remained far lower: 31.84% versus 98.29% for the API model. That is not evidence of overall editing superiority.

Add & Preserve deliberately isolates localized changes; it cannot alone establish reliability for every kind of visual revision. Residual color shifts also remained, and an additional experiment suggested that bypassing repeated image-to-latent conversions could complement TailSteer. The preprint supports targeted late-stage training as a promising way to reduce drift, not a complete solution to long-running image editing.