Ruikang Li, Zixuan Chen, Jiarui Wu, Mingde Yao, Yiyuan Zhang, Yujin Wang, Hanming Deng, Lewei Lu, Tianfan Xue
8 min
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.
Sam: So they don't dictate which valid replacement the model must invent. They correct the collateral changes around the replacement it already chose.
Alex: That's the distinction. They generate examples with a frozen copy of the original editor, then filter for reliable preservation regions. Some instructions request no change at all. Others request localized changes, with a preservation mask marking what should stay unchanged.
Sam: The paper calls this data-free training. There's still source imagery and filtering, though, so what is actually absent?
Alex: Manually curated ground-truth editing pairs are absent. It isn't training without images. The supervision comes from source images and the editor's own outputs, rather than human-prepared examples of the exact desired result.
Sam: Once those targets exist, why not train the model only to fix its final step?
Alex: They test that. Using the original model until the last update, then switching to a final-step-trained model, visibly reduces drift. In their illustrated sequence, results remain visually stable even after twenty editing turns.
Sam: But that still requires switching models. What happens if the trained version handles the whole generation process?
Alex: Images become desaturated and overly smooth. The same parameters operate throughout generation, so training only the last step can disturb earlier behavior. Full TailSteer therefore explicitly teaches the early steps to match the original editor, while gradually emphasizing preservation later.
Sam: That makes the early constraint part of the method, not merely a decision to leave early steps alone. Where does learning from the updated model come in?
Alex: That's their on-policy training: learning from states the updated model itself reaches. During training, it follows its own denoising path from cached starting points. At a sampled step, the authors recompute both the original model's prediction and the correction toward the preservation target.
Sam: So it learns to correct the path it's actually taking, rather than a path only the original model would take.
Alex: Yes. Earlier in the path, the target follows the original model. Later, it increasingly points toward the preservation target. At use time, the trained editor runs the whole process alone; the masks, reference model, and blended training targets disappear.
Sam: Let's separate the attractive mechanism from the evidence. How broadly did they test it?
Alex: They evaluated four foundation models across the FLUX, SenseNova, and Qwen families. Tests covered text-to-image generation, single-turn editing, repeated editing, and examples with multiple input images. Overall, the original generation and editing abilities were largely retained, though not every score held steady.
Sam: A model could preserve everything beautifully by refusing to make changes. How did they rule out that explanation?
Alex: They introduced Add & Preserve, using a hundred detail-rich natural images. Each image gets a white region enclosed by a red box. The model then receives ten successive instructions to add or remove objects inside it, while the surrounding image should stay unchanged.
Sam: That gives them separate places to measure success: requested changes inside, unintended damage outside.
Alex: They compare the outside region against the original image using several similarity measures. Inside, they check whether the objects match the editing history. For the Klein model, structural similarity outside the box rises from about point two nine to point six two after TailSteer.
Sam: That's a substantial preservation improvement. Did completion of the requested edits improve by a similar amount?
Alex: No. Its cumulative object accuracy stayed around a third, close to the original editor. The proprietary GPT comparison was near complete accuracy on that measure. TailSteer-Klein beat it on the outside-region preservation measures, not on overall editing ability.
Sam: So “better than the proprietary model” would lose the crucial qualification. It preserved the untouched region better in this test, while completing far fewer edits correctly.
Alex: That's the fair reading. Across the broader evaluations, content consistency improved, but instruction-following gains were less consistent. My assessment is that the evidence is stronger for slowing visual drift than for improving editing success generally.
Sam: The marked box also makes this unusually controlled. How much should that constrain our interpretation?
Alex: It isolates drift well, but doesn't represent every editing workflow. The paper also shows identity preservation, compositing, and pose-guided edits. Still, its cleanest quantitative drift evidence comes from localized changes inside a designated region, not unrestricted creative interaction.
Sam: And do the training comparisons show that the particular TailSteer recipe matters?
Alex: They use the same preservation data with alternative training strategies. Ordinary fine-tuning and tail-only training reduce drift but weaken editing capability. Removing on-policy training keeps preservation competitive, yet hurts spatial reasoning. Those comparisons support the need to preserve early behavior and train on the updated model's states.
Sam: Given that balance, who should read the full document, and where should they start?
Alex: Researchers training image editors or building iterative visual workflows should read it closely. Start with Tail Steering for the mechanism, then the training-strategy ablation for why each component matters. Read the Add & Preserve evaluation alongside the general editing results, so preservation doesn't stand in for instruction following.
Sam: And what should everyone else carry away?
Alex: TailSteer suggests that better repeated editing can come from refining how an existing model finishes each image. But preserving the picture and completing the edit are separate achievements.
Sam: Thanks for listening.