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.
Figures & tables
Figure 2: Method overview. Left: The student performs edits along a self-generated rollout, where each output becomes the input to the next turn. Given the previous output Ie−1 and the current instruction, the few-step student produces a clean estimate x^es . Right: Score distillation supervision with asymmetric conditioning . The student’s estimate x^es is noised to level τ and fed to both a frozen teacher and a trainable auxiliary editor. Unlike the student, both are conditioned identically on the original clean image I0 and a compound instruction aggregating all edits up to turn e ; however, the teacher represents the desired clean-conditioned distribution, while the auxiliary editor is trained to model the distribution actually produced by the student at edit e . The difference between their predictions yields the distillation gradient, which pushes the student’s multi-turn output distribution toward the teacher’s clean, non-degraded distribution.
Figure 3: Multi-turn image editing. We show selected turns from 20-step editing sequences. Each displayed instruction is followed by four no-op instructions (“keep the image as is”), and the images show the results after those no-ops. While the original FLUX.2 Klein accumulates color shifts and loses detail, fine-tuning it with our training paradigm (+Ours) yields a more stable editing process, preserving the scene and prior edits while following the requested changes.
Figure 4: Ten successive removals. Each turn removes one object from a synthetic garage scene: the drill, broom, car tire, ladder and bicycle (turns 1–5), then the lawn mower, vise, toolbox, pegboard and shelves (turns 6–10). By the last turn, every other method has visibly damaged the scene, while our model leaves the walls, the floor and the workbench as they were. Emu Edit leaves pale, blocky patches on the wall, VAE-LFA darkens the image under a coarse, mottled texture, FreqEdit casts the walls and floor yellow, and MT-OPSD washes the room out into a blurred, almost white space.
Quality
Preservation
Editing
Method
Quality (KRIS, 1–5) ↑
Δ HPSv3 (turn 1 → 10) ↑
Consistency (KRIS, 1–5) ↑
Anchor dist. (DINOv2) ↓
Background dist. (LPIPS) ↓
Instruction following (KRIS, 1–5) ↑
FLUX.2-klein-base-9B (undistilled)
Base model
1.42 ∗
−3.17∗
1.08 ∗
0.660 ∗
0.531 ∗
3.10 ∗
MT-OPSD [ 61 ]
1.95 ∗
−2.54∗
1.29 ∗
0.505 ∗
0.455 ∗
3.16 ∗
FLUX.2-klein-9B (4-step distilled)
Base model
2.73 ∗
−0.78∗
3.03 ∗
0.314 ∗
0.351 ∗
3.31
Table 1: Quantitative comparison on EdiVal-10. All metrics are computed after 10 cumulative edits and averaged over the 252 chains, except the anchor distance, which is averaged over the 148 chains whose anchor object lies mostly outside the edited objects (Appendix B ). Quality is measured by the KRIS-Bench image-quality score of the final image, assigned by a VLM judge (1–5), and by the change in HPSv3 from turn 1 to turn 10. Preservation is measured by the KRIS-Bench consistency score assigned by the same judge on the whole image (1–5; higher means fewer unrequested changes), the DINOv2 distance between the source and the final image on unedited anchor objects, and the LPIPS distance on the background outside all edited objects. Editing is measured by the KRIS-Bench instruction-following score at turn 10 (1–5); it checks that stability does not come from skipping edits. Methods are grouped by backbone; the best result in each column is shown in bold and the second best is underlined. ∗ : significantly different from our model (more than two standard errors, paired over chains).
Figure 5: Conditioning ablations. Turns 3, 6, 8 and 9 of a ten-turn editing sequence on a source scene generated with FLUX.2-klein-9B independently of the training data; the instruction below each column is the one applied at that turn. Our model changes only what each instruction names, and the image stays clean. The base model degrades the image: it turns yellow and washed out over the turns, and the knitted texture requested for the teddy bear at turn 9 also covers the rocking horse. Training with a teacher that also sees the previous image does not prevent this degradation: blotchy noise spreads over the walls and the rug from turn to turn. When the student also sees the source image, earlier edits are undone: the stool made yellow at turn 3 is blue again by turn 6, and the ball made green at turn 6 is red again by turn 8.
Quality
Preservation
Editing
Variant
Quality (KRIS, 1–5) ↑
Δ HPSv3 (turn 1 → 10) ↑
Consistency (KRIS, 1–5) ↑
Anchor dist. (DINOv2) ↓
Background dist. (LPIPS) ↓
Instruction following (KRIS, 1–5) ↑
Ours
3.19
−0.17
3.85
0.203
0.254
3.38
Teacher sees previous image
2.75 ∗
−0.86∗
2.81 ∗
0.307 ∗
0.366 ∗
3.32
Teacher sees only previous image and instruction
2.71 ∗
−1.12∗
2.28 ∗
0.366 ∗
0.381 ∗
3.23 ∗
Student sees source image
3.46 ∗
−0.08
4.08 ∗
0.188
0.230 ∗
2.80 ∗
5-edit training rollouts (vs. 10)
3.23
−0.28∗
3.29 ∗
0.265 ∗
0.334 ∗
3.25 ∗
Table 2: Ablations on EdiVal-10. Each row changes one aspect of our model; all variants are trained for 100 student updates and evaluated after 10 cumulative edits on all 252 chains (the anchor distance on the 148 chains whose anchor object lies mostly outside the edited objects; Appendix B ). ∗ : significantly different from our model (more than two standard errors, paired over chains).
Figure 6: Limitations. We show selected steps from a nine-edit sequence. Both the base model and ours follow the final instruction “Make the hiker smile,” but also remove the sunglasses added in a previous edit. While our method reduces accumulated drift, it inherits the editing capabilities and limitations of the base model.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Repeated content replacement over 40 edits. The unmodified distilled FLUX.2 Klein 9B model cycles from a dog in a living room to a cat in a living room, a cat on a beach, a dog on a beach, and a dog in a living room. We show the first, fifth, and tenth cycles of this sequential rollout. Visual quality remains coherent at late turns, while subject appearance and scene details change across cycles. Replacing both the subject and background every two turns limits long-term preservation demands, illustrating that sequence length alone does not necessarily lead to progressive visual degradation.
SR at turn k↑
CR at turn k↓
Method
3
5
8
10
3
5
8
10
FLUX.2-klein-base-9B (undistilled)
Base model
0.65 ∗
0.41 ∗
0.19 ∗
0.09 ∗
0.01
0.07
0.35 ∗
0.47 ∗
MT-OPSD [ 61 ]
0.64 ∗
0.50 ∗
0.27 ∗
0.10 ∗
0.03
0.03
0.06
0.16 ∗
FLUX.2-klein-9B (4-step distilled)
Base model
0.89
0.79 ∗
0.63 ∗
0.49
0.02
0.02
0.04
0.04
Appendix
Table 3: LME-Bench (100 sessions of 10 turns; Gemma-4-31B judge, 512-pixel protocol). SR: share of sessions whose turns 1 to k all succeed; CR: share of sessions collapsed by turn k . ∗ : significantly different from ours (exact sign test over sessions, p<0.05 ).
Method
Δ HPSv3 (source → edit) ↑
FLUX.2-klein-base-9B (undistilled)
Base model
−0.68∗
MT-OPSD [ 61 ]
−0.75∗
FLUX.2-klein-9B (4-step distilled)
Base model
−0.41∗
+ Emu Edit [ 42 ]
−0.56∗
Appendix
Table 4: Single edits on GEdit-Bench (606 English instructions, one edit each). Change in HPSv3 from the source to the edited image, averaged over instructions. ∗ : significantly different from our model (more than two standard errors, paired over instructions). † : VAE-LFA does not act on the first edit.
Figure 8: Successive background changes. Each training-free method carries part of an earlier background into the next turn. The teddy bear, alarm clock and toy train of a synthetic scene are placed in a lavender field at sunset (edit 1), in front of a red brick wall (edit 2), and on a smooth gradient backdrop from warm cream to light grey (edit 3). Our model keeps the three objects and gives a clean gradient. Emu Edit keeps the low, warm light of the sunset in front of the brick wall (edit 2), and at edit 3 leaves a clump of lavender stalks on the floor of the gradient backdrop and turns the blue train grey. VAE-LFA keeps the colour of the sunset: at edit 3 the whole image has a purple-magenta cast, the bear turns reddish and the clock face pink. FreqEdit prints the pattern of the brick wall into the smooth gradient. MT-OPSD leaves no trace of the earlier backgrounds but blurs the image and washes out its colours.
Figure 9: Additional multi-turn image editing results. We show selected turns from 20-step editing sequences. Each displayed instruction is followed by four no-op instructions (“keep the image as is”), and the images show the results after those no-ops. While the original FLUX.2 Klein accumulates color shifts and loses detail, fine-tuning it with our training paradigm (+Ours) yields a more stable editing process, preserving the scene and prior edits while following the requested changes.
Figure 10: Fifteen consecutive semantic edits. We compare the original FLUX.2 Klein with our fine-tuned model (+Ours) over a sequence of 15 edits. Each result becomes the input to the next edit. Our method reduces accumulated drift in color and scene details while retaining the original model’s ability to follow diverse editing instructions, including object additions, removals, and pose and expression changes.
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.
Multi-turn image editing aims to modify an image according to a sequence of instructions, making it better aligned with human-AI collaborative creation than single-turn editing. However, constructing high-quality multi-turn training data remains challenging, as conventional sequential rollout accumulates errors and introduces visual-quality drift across editing turns. To address this challenge, we introduce MTEdit, a multi-turn image editing dataset built through a pipeline centered on packed trajectory canvas (PTC) generation. By jointly generating multiple editing states within a unified visual context, this pipeline reduces visual-quality drift across turns. Covering both step-wise and context-aware editing scenarios, MTEdit comprises 500K multi-turn image editing trajectories, with up to 8 turns and an average of 5.38 turns per trajectory, making it one of the most extensive datasets for multi-turn image editing. We further introduce MTEdit-Bench, a benchmark designed to evaluate both scenarios by assessing models' ability to follow editing instructions, maintain content consistency, resist visual-quality drift, and effectively use historical context across turns. Using MTEdit, we fine-tune Qwen-Image-Edit-Causal to obtain Qwen-Image-MTEdit. On MTEdit-Bench, we evaluate a broad range of closed-source and open-source image editing models, including Qwen-Image-MTEdit. Experimental results show that Qwen-Image-MTEdit achieves the best performance among the evaluated open-source models, demonstrating the effectiveness of MTEdit in improving models' multi-turn image editing capabilities.
Houyuan Chen, Fanzhou Wang, Xiangyu Fan +7
SenseTime Research · HKUST MMLab · Stanford University +1
Text-guided image editing has advanced rapidly with diffusion models and unified multimodal foundation models. However, most existing methods remain confined to single-turn settings, overlooking the more realistic scenario of multi-turn in-context editing, where users iteratively refine an image through a sequence of instructions. In this setting, a model must follow each new instruction while preserving accumulated session-level constraints, challenged by two coupled failure modes: long-context dilution, where sparse textual constraints become difficult to recover from growing interleaved image-text histories, and state contamination, where earlier editing mistakes degrade subsequent generations. We introduce Edit-R2, a novel reinforcement learning post-training framework for unified multimodal models. Edit-R2 reconstructs the operative session intent, which effectively consolidates scattered historical constraints into an explicit reasoning trace before each editing turn. It further enables multi-turn RL over both reasoning and generation through a unified objective that jointly optimizes intent reconstruction generation in discrete text space and flow-matching image generation in continuous latent space, while a trajectory filtering mechanism suppresses corrupted rollouts to stabilize training under state contamination. To support systematic evaluation, we introduce MICE-Bench, a large-scale benchmark for multi-turn in-context editing with automated metrics for instruction following (IF), content consistency (CC), and global awareness (GA) over accumulated session constraints. Experiments show that Edit-R2 substantially improves multi-turn in-context editing and achieves competitive performance compared against strong baselines.
Yuxiao Ye, Haoran He, Fangyuan Kong +4
Hong Kong University of Science and Technology · Kuaishou Technology