TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing
Ruikang Li, Zixuan Chen, Jiarui Wu, Mingde Yao, Yiyuan Zhang, Yujin Wang, Hanming Deng, Lewei Lu, Tianfan Xue
cs.CV, cs.AI
2026-11-05
TailSteer trains only the last denoising steps on-policy; FLUX.2 klein's background PSNR after 10 edit turns rises from 13.85 to 19.96 with single-turn scores nearly intact.
An image editor that scores well on a single edit can still fall apart in a conversation. Each output becomes the next input, so small unintended changes (softened textures, shifting facial detail, color casts in the background) get inherited and compound. After twenty turns the picture can look nothing like the original. Products like ChatGPT Images and Nano Banana 2 are built around this iterative workflow, so the failure is practical.
Existing fixes fall into three groups. Auxiliary modules (masks, consistency LoRAs, frequency corrections) constrain the pretrained prior and tend to suit only localized edits. Better representations (new VAEs, pixel diffusion) reduce per-turn error but still degrade over repeated edits. Memory and agentic planning leave the model untouched and add inference cost. TailSteer takes a fourth route: bake drift resistance into the base model, with nothing extra at inference.
The starting observation is that diffusion sampling fixes low-frequency structure early and recovers high-frequency detail near the end, while preservation error in unchanged regions is still nonzero at the final step. So correction belongs at the tail, and the early steps should stay as they were.
Where supervision comes from. No curated editing pairs. The frozen base editor generates its own rollouts of two kinds: identity edits ("keep the image exactly unchanged"), where the whole image is the preservation target, and local edits, filtered so the change is small and concentrated. A preservation mask M is derived from the source-versus-output difference. The target endpoint takes source content inside M and the model's own edit outside it, so the model is never forced toward one particular edit, only toward restoring what should not have changed.
Why not train just the last step. The paper first tries a terminal-only intervention. Swapping in the fine-tuned model for the final step keeps faces stable through 20 turns. Using it for all steps gives desaturated, over-smoothed images, because parameters are shared across timesteps.
Full method. A steering onset ks splits the trajectory:
At inference the single refined model runs the whole trajectory. No model switching, no adapters.
Training used 11,947 source images, 38,540 cached trajectories, 2,000 updates per backbone on 8 H100s. Four backbones were tested: FLUX.1 Kontext, FLUX.2 klein 9B, SenseNova-U1.5-8B, Qwen-Image-2.1.
On Add & Preserve, the authors' own benchmark (100 images, 10 successive add/remove edits inside a box, drift measured outside it):
| FLUX.2 klein | PSNR@10 | SSIM@10 | LPIPS@10 |
| Base | 13.85 | 0.289 | 0.524 |
| TailSteer (ks=46) | 19.96 | 0.621 | 0.290 |
SenseNova's PSNR@10 rises from 11.23 to about 18.2. The paper reports TailSteer-Klein beating GPT-Image-2.5 Sunburst (API) on all three preservation metrics. Object accuracy (COA) stays near base level, so the gain does not come from refusing edits.
Single-turn retention on Klein: GEdit-EN 7.47 for base versus 7.52 for TailSteer; RISE 14.17 versus 13.61. FreqEdit drops GEdit-EN to 7.20 and ConsLoRA to 6.58.
Ablations on Klein with identical data:
| Training scheme | GEdit-EN | PSNR@10 |
| Standard SFT | 5.24 | 19.77 |
| Last step only | 2.24 | 14.93 |
| No on-policy | 7.44 | 19.8 |
| Full TailSteer | 7.52 | 19.96 |
SFT resists drift but loses over two points of editing quality. Dropping on-policy training cuts RISE to 11.39 (13.61 for the full method) and breaks a spatial rearrangement task. More data helps, and mixing identity and local edits gives the best balance.
For teams running multi-turn editing on their own models, this is a cheap route: no architecture change, no inference module, no cross-turn data, and all supervision is self-generated. Because it only touches the denoising trajectory, it transferred across FLUX, SenseNova and Qwen.
It reads as a post-training patch, not a new editing capability. It also complements VAE bypassing: with latent feedback, Klein's turn-10 PSNR approaches its standard-feedback turn-5 level.