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.
Figures & tables
Figure 1: Single- and multi-input multi-turn editing. TailSteer substantially reduces cumulative drift over editing turns while retaining the base model’s (FLUX.2 [klein] Base 9B) editing capability.
Figure 2: Trajectory tail remains correctable. Low-frequency content is recovered early, whereas higher-frequency details are primarily recovered near the tail (top), while preservation error in unchanged regions continues to decrease but remains non-zero at the endpoint (bottom).
Figure 3: Data curation of preservation-aware trajectory endpoints. We construct supervision from rollouts of the frozen base editor using two complementary edit types: (a) identity editing provides full-image preservation trajectories, while (b) local editing provides partial-preservation trajectories with spatially localized changes. Only samples that pass the corresponding sanity checks are retained and cached as training trajectories for subsequent TailSteer optimization.
Figure 4: One-step terminal intervention: benefits and limitations. (a) We fine-tune FLUX.2-klein-9B with supervision only at the final sampling step, steering from the cached pre-terminal state xτbase toward the preservation-aware target x1tgt . (b) The base model accumulates visible identity drift over repeated edits (top row). Applying the fine-tuned model only at the final step markedly improves long-horizon preservation (middle row), whereas using it throughout sampling causes desaturation and detail loss (bottom row). Therefore, it is necessary to preserve the earlier dynamics.
Figure 5: On-policy training of TailSteer. At randomly sampled training states xk1 and xk2 , the refined model receives region-dependent supervision. In the prior region, its prediction is matched to the frozen base model’s velocity vbase ; in the transition region, the target v∗ progressively blends vbase with the preservation velocity vpres toward x1tgt . To obtain these states on-policy, the refined model rolls out from the nearest preceding cached anchor in prior region, and from the fixed steering-onset state in transition region. The velocity target is then recomputed at the state actually reached by the refined model, with gradients applied only to its prediction at the sampled step.
Figure 6: 20-turn qualitative comparison. While competing methods exhibit cumulative visual drift, TailSteer faithfully preserves the original tabletop color and previously established details, including the coffee cup, latte art, and cookie on the right, even after 20 successive editing turns.
Figure 7: Multi-input multi-turn editing. TailSteer retains base models’ multi-image editing capabilities while substantially reducing cumulative drift, preserving backgrounds during multi-person and multi-object compositing and the girl’s dress details across successive skeleton-guided edits.
Backbone
Baseline
ks
Text-to-image
Single-turn image editing
Qwen-IB EN ↑
Qwen-IB ZH ↑
BizGen ES (%) ↑
GEdit-EN ↑
GEdit-ZH ↑
RISE Acc. (%) ↑
ImgEdit ↑
FLUX.1 Kontext [dev]
Base model
–
41.19 ↑ 0.00
19.23 ↑ 0.00
0 1.70 ↑ 0.00
5.99 ↑ 0.00
0.89 ↑ 0.00
0 8.89 ↑ 0.00
3.77 ↑ 0.00
W/ FreqEdit
–
–
–
–
5.20 ↓ 0.79
0.64 ↓ 0.25
0 6.94 ↓ 1.94
3.62 ↓ 0.15
After TailSteer
22
41.63 ↑ 0.44
19.17 ↓ 0.06
0 2.17 ↑ 0.48
5.93 ↓ 0.06
0.81 ↓ 0.07
0 6.94 ↓ 1.94
3.80 ↑ 0.03
25
41.50 ↑ 0.32
19.19 ↓ 0.04
0 2.07 ↑ 0.38
5.95 ↓ 0.04
0.81 ↓ 0.07
0 7.22 ↓ 1.67
3.83 ↑ 0.06
27
41.30 ↑ 0.11
18.89 ↓ 0.34
0 1.70 = 0.00
5.97 ↓ 0.02
0.86 ↓ 0.02
0 7.50 ↓ 1.39
3.83 ↑ 0.06
Table 1: Single-turn capability retention. Text-to-image generation and single-turn editing performance across base models. Teal ↑ and orange ↓ indicate improvement/degradation from each base model. Bold number marks the best non-base score per backbone and metric, including ties.
Figure 8: Cumulative drift across turns. Outside-box metrics on Add & Preserve over 10 turns. TailSteer reduces drift across all backbones, approaching or surpassing GPT-Image-2.5-Sunburst in preservation. Dashed curves show further gains from bypassing inter-turn VAE for TailSteer-Klein.
Figure 9: Qualitative comparison after 10 editing turns. TailSteer substantially reduces background drift while retaining the requested object edits across different backbones. For Klein, bypassing inter-turn VAE conversions further preserves the original background colors and details.
Figure 10: Ablation of training strategies. With the same training data, full TailSteer better balances preserving the base model’s editing capabilities with enhancing multi-turn drift resistance.
Ablative training scheme
Single-turn image editing
Add & Preserve
(Use FLUX.2 [klein] Base 9B as backbone)
GEdit-EN ↑
GEdit-ZH ↑
RISE Acc. (%) ↑
ImgEdit ↑
COA (%) ↑
PSNR@10 ↑
SSIM@10 ↑
LPIPS@10 ↓
Training only last-step
2.24 ↓ 5.28
2.17 ↓ 5.22
0 3.33 ↓ 10.28
2.15 ↓ 2.04
13.10 ↓ 18.73
14.93 ↓ 5.03
0.399 ↓ 0.222
0.594 ↑ 0.305
Standard Supervised Fine-tuning (SFT)
5.24 ↓ 2.28
5.28 ↓ 2.11
0 8.89 ↓ 0 4.72
3.41 ↓ 0.77
31.07 ↓ 0 0.76
19.77 ↓ 0.19
0.608 ↓ 0.013
0.287 ↓ 0.003
Training only trajectory tail (from step-46)
6.09 ↓ 1.42
6.02 ↓ 1.37
10.28 ↓ 0 3.33
3.92 ↓ 0.26
29.62 ↓ 0 2.22
20.04 ↑ 0.08
0.606 ↓ 0.015
0.313 ↑ 0.023
TailSteer w/o on-policy training
7.44 ↓ 0.08
7.40 ↑ 0.01
11.39 ↓ 0 2.22
4.17 ↓ 0.01
31.94 ↑ 0 0.11
19.81 ↓ 0.15
0.607 ↓ 0.014
0.305 ↑ 0.015
TailSteer (full method, steering from step-46)
7.52 ↑ 0.00
7.39 ↑ 0.00
13.61 ↑ 00.00
4.19 ↑ 0.00
31.84 ↑ 00.00
19.96 ↑ 0.00
0.621 ↑ 0.000
0.290 ↑ 0.000
Table 3: Training-strategy ablation. All variants share same training data and full-steps sampling.
Ablative training data scale
Cached
Single-turn image editing
Add & Preserve
(Use FLUX.2 [klein] Base 9B as backbone)
Trajectories
GEdit-EN ↑
GEdit-ZH ↑
COA (%) ↑
PSNR@10 ↑
SSIM@10 ↑
LPIPS@10 ↓
Identity-edit data
1,000
7.05 ↓ 0.47
6.82 ↓ 0.56
30.27 ↓ 1.57
14.15 ↓ 5.81
0.482 ↓ 0.139
0.400 ↑ 0.110
Local-edit data
7.40 ↓ 0.12
7.35 ↓ 0.04
30.31 ↓ 1.52
17.61 ↓ 2.35
0.557 ↓ 0.064
0.330 ↑ 0.041
Identity-edit + Local-edit data
7.28 ↓ 0.24
7.26 ↓ 0.13
30.42 ↓ 1.41
16.88 ↓ 3.08
0.538 ↓ 0.083
0.349 ↑ 0.059
Identity-edit data
5,000
6.92 ↓ 0.59
6.97 ↓ 0.42
31.23 ↓ 0.61
16.61 ↓ 3.35
0.540 ↓ 0.081
0.353 ↑ 0.063
Local-edit data
7.26 ↓ 0.26
7.27 ↓ 0.12
31.11 ↓ 0.72
18.85 ↓ 1.11
0.614 ↓ 0.007
0.278 ↓ 0.012
Table 4: Data scale and trajectory-type ablation. All variants are trained for nearly one epoch.
Image editing models produce impressive results for individual edits, but their quality often deteriorates over multiple editing turns. Each editing turn introduces small distribution shifts that accumulate over successive edits, creating a gap between the input distributions encountered during the editor's original training and multi-turn inference. We introduce a training scheme that makes multi-turn image editing stable under this accumulating distribution shift. A key is to force the editor to operate on its own outputs from previous editing turns during training, bringing the principle of Self-Forcing to image editing. The main challenge is that self-generated editing trajectories deviate from ground-truth trajectories, making direct supervision unreliable. We address this challenge through a distillation objective that leverages a teacher image editor to provide a reliable training signal when the model operates on its own previous outputs. We introduce an asymmetric conditioning scheme: while the student performs edits sequentially on its own outputs, the teacher remains anchored to the original clean image and receives a compound editing instruction. Experiments show substantially reduced degradation over long multi-turn sequences while maintaining editing quality and instruction adherence.
Multi-turn image editing is essential for iterative design, yet current models often struggle with identity drift and error accumulation over successive steps. While existing research leverages video priors for consistency, their reliance on bidirectional attention is fundamentally misaligned with the causal, sequential nature of interactive editing. In this paper, we propose AnchorEdit, the first autoregressive (AR) diffusion-based framework designed specifically for high-resolution, long-term multi-turn editing. AnchorEdit bridges the gap between video priors and causal inference through a three-stage training curriculum: identity-preserving sing-turn pretraining, causal AR forcing fine-tuning with a novel self-rollout strategy to mitigate exposure bias, and consistency distillation for efficient 4-step generation. During inference, we introduce a memory mechanism to anchor the initial subject identity and ensure stable extrapolation across extended editing trajectories. To evaluate performance, we provide a new high-resolution multi-turn editing benchmark designed to stress-test long-horizon stability. Extensive experiments demonstrate that AnchorEdit achieves state-of-the-art results, maintaining exceptional subject fidelity and instruction following even over 10+ interaction rounds.
Hang Xu, Xiaoxiao Ma, Guohui Zhang +7
University of Science and Technology of China · JD Explore Academy
Instruction-based image editing has achieved strong performance in single-turn settings, yet practical editing is often iterative, with each instruction applied to the output of the previous turn. We find that existing editing models degrade rapidly under recursive editing and attribute this failure to a train-test mismatch in the conditioning distribution: models are trained on clean source images but must repeatedly condition on their own imperfect outputs at inference time. To address this, we propose MT-OPSD, an on-policy self-distillation framework that trains the model on self-generated conditioning states with editing supervision from a clean-conditioned teacher, without requiring multi-turn annotations. We further introduce LME-Bench, a benchmark of 100 ten-turn editing sessions for evaluating long-horizon robustness. Experiments across three editing backbones show that MT-OPSD substantially improves long-horizon editing success and reduces multi-turn collapse while largely preserving single-turn editing quality.