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

AI summary

Overview

• Research area: Computer vision, generative image editing, and diffusion-model training.
• Technical level: Intermediate.
• Scope: The paper introduces TailSteer, a training method intended to make image editors preserve unchanged content more reliably during repeated, multi-turn editing without adding components at inference time.

What This Paper Is About

Image editors can perform well on a single request but gradually change details that users did not ask to edit when the edited image is reused for later requests. TailSteer aims to reduce this cumulative visual drift while retaining the model’s existing ability to generate and edit images. Its central idea is to leave most of the model’s denoising process close to its original behavior and focus preservation training on the later part of that process.

Key Contributions

1. A targeted training strategy: TailSteer steers the late portion of a model’s denoising trajectory toward better preservation, rather than broadly changing how the model edits images.
2. Training supervision from the model’s own outputs: The method uses the base editor’s rollouts and source images to create preservation targets. It does not require curated source–target editing pairs or multi-turn training examples.
3. On-policy training: TailSteer trains on states reached by the refined model itself, so its corrections are learned for the trajectories it will actually take, rather than only for trajectories produced by the original model.
4. A separate drift evaluation: The paper introduces Add & Preserve, which measures whether edits are completed while also tracking changes to image regions that should stay untouched.

Main Findings

• Less drift over repeated edits: The authors report reduced cumulative drift across the tested model families. Qualitative examples include editing sequences of up to 20 turns; the controlled Add & Preserve benchmark measures drift over 10 turns.
• The denoising tail is a useful intervention point: The paper’s analysis suggests that low-frequency structure is established earlier, while fine details are recovered closer to the end. Preservation errors remain at the endpoint, making late-stage correction a practical target.
• A final-step-only correction has a trade-off: On FLUX.2 [klein] Base 9B, using the fine-tuned model only for the last update reduced visible drift in a 20-turn example. Using that model throughout sampling instead led to desaturation and loss of detail, motivating TailSteer’s gradual, tail-focused training.
• Most single-turn capabilities are retained: Across text-to-image and single-turn editing evaluations, TailSteer’s scores are generally close to those of the corresponding base models. The paper also reports that preservation modules such as FreqEdit or Consistency LoRA V2 can cause larger declines on some single-turn editing measures.
• The results vary by model and steering onset: TailSteer improves EdiVal content consistency across the reported settings, but gains on MSE-Bench success rates are limited to selected steering onsets, and instruction-following gains are less consistent.
• Strong preservation results on Add & Preserve: TailSteer reduces drift in the benchmark’s outside-box PSNR, SSIM, and LPIPS measures, while Cumulative Object Accuracy is generally near the base model’s level. TailSteer-Klein surpasses GPT-Image-2.5 Sunburst on those three preservation metrics in the reported comparison; this does not mean it also surpasses it on edit completion.
• Multiple model families are tested: Experiments cover FLUX.1 Kontext [dev], FLUX.2 [klein] Base 9B, SenseNova-U1.5-8B-MoT, and Qwen-Image-2.1. The paper also reports qualitative tests involving multi-image compositing and pose-guided editing.

Methodology in Plain English

The researchers first collect source images and ask a frozen base editor to perform two kinds of edits: identity edits, which should leave the whole image unchanged, and local edits, which should change only a limited region. They filter the resulting examples for reliable preservation regions. For local edits, they use the difference between the source and generated image to identify areas that should remain unchanged; the generated edit itself is retained elsewhere.

These rollouts provide examples of where the editor’s own sampling process has introduced unwanted changes. TailSteer uses them to construct a target that restores the source content in preservation regions while keeping the model’s chosen edit in the changed region.

During training, the refined model is encouraged to match the original model’s behavior earlier in denoising. Later, its training target gradually shifts toward correcting preservation errors. The training states come from rollouts of the refined model, so the supervision reflects the states it actually reaches.

At inference, the refined model runs the full denoising process by itself. The frozen model, masks, and other training aids are not needed, and the method does not require an inference-time model switch or extra consistency module.

The experiments use 11,947 candidate source images from Open Images and MONET. Each backbone is trained for 2,000 updates, with an effective batch size of 16 on eight NVIDIA H100 GPUs.

Why This Matters

Research impact: TailSteer treats multi-turn robustness as a property that can be trained into a foundation model, rather than relying only on external memory, specialized inference procedures, or multi-turn demonstrations. Its results also suggest that training should focus on the part of denoising where fine details are finalized, rather than applying preservation pressure uniformly.

Potential applications:

• Iterative photo editing: Refining an image over repeated requests while keeping faces, backgrounds, and other unedited details stable.
• Product and marketing design: Making successive changes to a product image or scene without gradually altering previously approved colors or details.
• Multi-image compositing: Combining people or objects from different images while better preserving background and subject details.
• Pose-guided creative workflows: Applying successive pose or layout changes while retaining clothing and other visual details.

Industry relevance: TailSteer may help make image-generation systems more dependable for conversational and agentic creative workflows, where users refine an image rather than specifying the final result in one request. Because the trained model needs no additional components at inference time, the approach is designed to improve the model itself rather than add a separate runtime mechanism.

Future Directions

• Test how well TailSteer generalizes to more editing styles, instructions, and image domains than the identity and local edits used to construct its preservation supervision.
• Examine ways to address drift caused by repeated image encoding and decoding. The paper’s latent-feedback experiment suggests this could complement trajectory-tail steering, but reports only limited improvement from latent feedback alone.
• Study how steering onset and rollout-data scale affect the balance between preservation and edit completion across different models.
• Evaluate longer and more varied interactive workflows, including sequences where edits overlap or where users revise earlier instructions.

Target Audience

This paper is most relevant to researchers and practitioners working on diffusion models, image editing, generative-model fine-tuning, and conversational creative tools. It should also be accessible to readers who understand basic image-generation concepts and want to learn why repeated editing can cause drift and how training can reduce it.

Authors’ 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.