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.
Figures & tables
Figure 1: Overview of our approach. (a,b) Packed trajectory canvas (PTC) generation reduces visual-quality drift compared with sequential rollout. (c) MTEdit offers more trajectories and turns than existing datasets. (d) Fine-tuning on MTEdit improves multi-turn editing performance in instruction following (IF), content consistency (CC), and visual-quality stability (DRS).
Dataset
#Traj.
#Edits
Max T.
Avg T.
Step-wise
Context-aware
Construction
MagicBrush [ 69 ]
3K
8K
3
2.58
✓
✗
Human
SEED-Data-Edit [ 17 ]
21K
81K
6
3.82
✓
✗
Human
ImgEdit [ 64 ]
115K
314K
3
2.73
✗
✓
Inpainting
Pico-Banana-400K [ 43 ]
77K
306K
6
3.97
✓
✗
Sequential rollout
MTEdit (Ours)
500K
2.69M
8
5.38
✓
✓
Packed trajectory canvas
Table 1: Comparison of representative multi-turn image editing datasets. #Traj. and #Edits denote the numbers of editing trajectories and edit pairs; Max T. and Avg T. denote the maximum and average numbers of turns; Step-wise and Context-aware indicate whether the dataset supports the two multi-turn editing settings; Construction describes how the dataset is built.
Figure 2: Overview of our data construction pipeline. (a) Seed image curation, which selects source images with high resolution and aesthetic quality; (b) PTC generation, in which VLMs derive a PTC prompt from each source image to guide closed-source image models in generating the corresponding PTC; and (c) PTC to multi-turn conversion, which processes the generated canvases into multi-turn editing trajectories with turn-level instructions and target images.
Figure 3: Statistics of MTEdit. (a) Edit-type taxonomy, (b) multi-turn data distribution, and (c) instruction word cloud.
Figure 4: Qualitative results on the step-wise sub-bench. Our method achieves competitive editing performance and maintains visual quality without noticeable degradation across editing turns.
Turn 1
Turn 2
Turn 3
Turn 4
Method
IF
CC
DRS
IF
CC
DRS
IF
CC
DRS
IF
CC
DRS
Nano Banana Pro [ 20 ]
9.75
9.79
4.17
9.73
9.53
4.22
9.63
9.35
4.22
9.53
9.06
4.19
Nano Banana 2 [ 21 ]
9.65
9.58
4.19
9.45
9.21
4.21
9.28
8.85
4.23
9.31
8.53
4.20
GPT Image 2 [ 42 ]
9.02
8.72
4.22
8.66
8.62
4.23
8.65
8.07
4.22
8.62
7.82
4.20
Seedream 5.0 Lite [ 4 ]
9.73
9.60
4.27
9.81
9.08
4.29
9.66
8.57
4.26
9.34
7.69
4.19
VINCIE [ 44 ]
1.13
4.31
2.47
0.01
0.02
1.66
0.00
0.00
1.40
0.00
0.00
1.33
Table 2: Quantitative results on the step-wise sub-bench. Closed-source scores are shown in gray ; among open-source methods and our model, different cell colors refer to best and 2nd-best results. For all metrics, higher values are better.
Figure 5: Visual quality comparison at Turn 8. From left to right: the source image and outputs from models trained on data from PTC generation and sequential rollout. Enlarged crops show fewer artifacts and cleaner details with our pipeline.
Turn 1
Turn 2
Turn 3
Turn 4
Method
IF
CC
DRS
IF
CC
DRS
IF
CC
DRS
IF
CC
DRS
PTC
9.64
9.61
4.17
9.63
9.27
4.26
9.30
8.59
4.21
8.67
7.67
4.22
Sequential rollout
9.66
9.50
4.19
9.32
9.04
4.27
8.97
8.20
4.18
7.95
7.58
4.19
Turn 5
Turn 6
Turn 7
Turn 8
Method
IF
CC
DRS
IF
CC
DRS
IF
CC
DRS
IF
CC
DRS
PTC
8.21
7.15
4.21
7.63
6.47
4.21
6.24
5.37
4.23
5.06
4.44
4.22
Table 3: Quantitative ablation results on the step-wise sub-bench. Method denotes the strategy used to construct the training data: PTC generation or sequential rollout. The best results are bolded . Higher values indicate better performance for all metrics.
Appendix figures & tables22 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Block causal attention and 3D interleaved RoPE for Qwen-Image-MTEdit. Image and text tokens are temporally interleaved as “ [Image] - [Text] - [Image] - ... ”.
Figure 7: Example of PTC generation used in the MTEdit construction pipeline.
Figure 8: Visual quality drift comparison. We compare our data with rollout-style results generated using Seedream 5.0 Lite/GPT Image 2/Nano Banana 2. Columns show results at Turns 1/4/8. The visual quality of rollout-style methods degrades noticeably starting from Turn 4 (e.g., spurious speckling appears on the ground near the car), whereas our results remain relatively stable.
Model
Method
Turn 1
Turn 2
Turn 3
Turn 4
Turn 5
Turn 6
Turn 7
Turn 8
Ours
PTC
4.28
4.30
4.30
4.29
4.29
4.29
4.28
4.27
Seedream 5.0 Lite [ 4 ]
Sequential rollout
4.30
4.10
4.16
3.95
3.88
3.70
3.65
3.51
Nano Banana 2 [ 21 ]
Sequential rollout
4.32
4.31
4.30
4.24
4.14
4.00
3.84
3.70
GPT Image 2 [ 42 ]
Sequential rollout
4.28
4.24
4.18
4.12
4.04
3.94
3.88
3.83
Appendix
Table 4: Quantitative comparison of visual-quality drift across data construction methods. PTC denotes packed trajectory canvas. We report the Drift Robustness Score ( DRS ) at each editing turn. Higher values indicate better visual quality; the best score at each turn is bolded .
Turn 1
Turn 2
Turn 3
Turn 4
Method
IF∗
CC∗
DRS
IF∗
CC∗
DRS
IF∗
CC∗
DRS
IF∗
CC∗
DRS
Nano Banana 2 [ 21 ]
8.70
9.04
4.28
8.70
8.64
4.27
8.32
8.38
4.25
7.64
7.46
4.19
Nano Banana Pro [ 20 ]
9.02
9.28
4.28
8.42
8.62
4.28
7.98
8.22
4.24
8.02
7.78
4.19
GPT Image 2 [ 42 ]
9.36
9.36
4.29
9.28
9.42
4.28
8.30
8.84
4.25
7.82
7.72
4.23
Seedream 5.0 Lite [ 4 ]
9.88
9.88
4.37
9.34
9.66
4.37
8.94
9.54
4.32
8.68
9.18
4.28
VINCIE [ 44 ]
1.58
5.30
2.79
0.20
0.20
1.76
0.06
0.06
1.71
0.00
0.00
1.51
Appendix
Table 5: Quantitative results on the persistent-constraint sub-bench. Closed-source scores are shown in gray ; among open-source methods and our model, different cell colors indicate the best and 2nd-best results. For all metrics, higher values indicate better performance. DRS is reported as an absolute score at every turn. After a case first has IF or CC below 6, subsequent IF and CC scores are set to zero; the triggering turn is retained. DRS uses actual scores at every turn, and all means include all cases.
Figure 9: Qualitative results on the persistent-constraint sub-bench. Yellow boxes highlight representative editing errors.
Turn 1
Turn 2
Turn 3
Turn 4
Method
IF∗
CC∗
DRS
IF∗
CC∗
DRS
IF∗
CC∗
DRS
IF∗
CC∗
DRS
Nano Banana 2 [ 21 ]
9.36
9.92
4.12
9.28
9.52
4.13
9.22
9.14
4.12
8.70
8.40
4.06
Nano Banana Pro [ 20 ]
9.56
9.76
4.09
9.30
9.56
4.09
8.24
8.82
4.08
7.84
8.12
4.00
GPT Image 2 [ 42 ]
9.72
9.78
4.14
9.34
9.84
4.18
8.78
9.14
4.17
8.42
8.68
4.18
Seedream 5.0 Lite [ 4 ]
9.80
9.82
4.20
8.46
9.76
4.19
7.54
8.12
4.17
7.30
7.44
4.15
VINCIE [ 44 ]
3.00
5.86
2.67
0.86
1.00
1.75
0.44
0.40
1.49
0.20
0.20
1.42
Appendix
Table 6: Quantitative results on the textual-reference sub-bench. Closed-source scores are shown in gray ; among open-source methods and our model, different cell colors indicate the best and 2nd-best results. For all metrics, higher values indicate better performance. DRS is reported as an absolute score at every turn. After a case first has IF or CC below 6, subsequent IF and CC scores are set to zero; the triggering turn is retained. DRS uses actual scores at every turn, and all means include all cases.
Figure 10: Qualitative results for textual-reference multi-turn image editing. Red boxes highlight representative editing errors.
Turn 1
Turn 2
Turn 3
Turn 4
Method
IF∗
CC∗
DRS
IF∗
CC∗
DRS
IF∗
CC∗
DRS
IF∗
CC∗
DRS
Nano Banana 2 [ 21 ]
9.68
9.78
4.25
8.24
9.58
4.20
7.70
7.82
4.18
6.38
6.50
4.13
Nano Banana Pro [ 20 ]
9.74
9.96
4.25
8.22
9.66
4.23
7.44
7.94
4.17
6.42
7.08
4.07
GPT Image 2 [ 42 ]
9.86
9.84
4.21
9.30
9.80
4.25
8.76
9.20
4.26
7.72
8.20
4.22
Seedream 5.0 Lite [ 4 ]
9.58
9.74
4.30
8.54
9.72
4.27
7.60
8.20
4.22
6.72
7.52
4.16
VINCIE [ 44 ]
2.92
6.30
2.80
1.16
1.30
1.78
0.30
0.32
1.59
0.08
0.10
1.49
Appendix
Table 7: Quantitative results on the short-instruction sub-bench. Closed-source scores are shown in gray ; among open-source methods and our model, different cell colors indicate the best and 2nd-best results. For all metrics, higher values indicate better performance. DRS is reported as an absolute score at every turn. After a case first has IF or CC below 6, subsequent IF and CC scores are set to zero; the triggering turn is retained. DRS uses actual scores at every turn, and all means include all cases.
Figure 11: Qualitative results on the short-instruction sub-bench. Later turns progressively lengthen the flower stems using concise, context-dependent instructions.
Figure 12: Representative examples from MTEdit. Rows show step-wise, persistent-constraint, textual-reference, and short-instruction editing trajectories.
Sub-bench
IF / IF ∗
CC / CC ∗
DRS
ρ↑
τb↑
ρ↑
τb↑
ρ↑
τb↑
Step-wise
0.83
0.71
0.85
0.75
0.92
0.83
Persistent-constraint
0.88
0.78
0.85
0.75
0.89
0.78
Textual-reference
0.92
0.82
0.87
0.71
0.83
0.75
Short-instruction
0.82
0.71
0.83
0.67
0.91
0.78
Appendix
Table 8: Alignment between MTEdit-Bench and human judgments. We report Spearman’s ρ and Kendall’s τb ; higher values indicate stronger agreement.
Figure 13: Failure case of PTC generation. Content from different panels may interfere with one another, requiring additional filtering during data construction.
Figure 14: Additional example of step-wise multi-turn image editing.
Figure 15: Additional example of step-wise multi-turn image editing.
Figure 16: Additional example of persistent-constraint multi-turn image editing.
Figure 17: Additional examples of textual-reference multi-turn image editing.
Figure 18: Additional examples of short-instruction multi-turn image editing.
Figure 19: Data examples for step-wise multi-turn image editing.
Figure 20: Data examples for persistent-constraint multi-turn image editing.
Figure 21: Data examples for textual-reference multi-turn image editing.
Figure 22: Data examples for short-instruction multi-turn image editing.
Multi-turn image editing aims to modify an image according to a sequence of instructions, enabling more effective interactive visual content creation for users. However, existing methods often exhibit noticeable drift after only a few editing turns, manifesting in two aspects: 1) unstable visual quality, with accumulating artifacts and degraded details; and 2) inconsistent image content, with unintended changes to preserved regions and prior edits. To address these challenges, we propose MT Forcing, a novel training paradigm for multi-turn image editing with unified multimodal models (UMMs) based on autoregressive diffusion. Specifically, we propose Single-Turn Distribution Matching Distillation (Single-Turn DMD), which supervises student self-rollouts with a teacher that applies cumulative edits to the source image in a single turn, providing visual-quality guidance without recursive error accumulation. Additionally, we introduce Teacher-Forcing Guidance, which supervises history-dependent transitions on reference trajectories during warmup and as a regularizer during distillation, encouraging cross-turn content consistency beyond what per-turn distribution matching alone can ensure. Extensive experiments demonstrate that our approach significantly improves the multi-turn image editing capabilities of UMMs, yielding more stable visual quality and stronger cross-turn consistency.
Houyuan Chen, Fanzhou Wang, Xiangyu Fan +9
SenseTime Research · HKUST MMLab · Stanford University +1
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.
Recent breakthroughs in instruction-based image editing have captured significant attention, as models are now capable of handling real-world editing demands with the practicality required by everyday users. However, editing models trained primarily for single-turn edits often break down in multi-turn editing--the natural interactive setting where a user iteratively refines an image based on the model's own previous outputs. This failure stems from the all-or-nothing requirement, where a single failed turn compromises the entire sequence, and error propagation, where exposure bias leads to compounding editing errors. To address these challenges, we introduce MT-EditFlow, a flow-matching reinforcement learning framework designed to optimize reward signals for sequential image editing. MT-EditFlow integrates a multi-turn perspective with a multi-reward formulation to provide a unified structure applicable to both GRPO and NFT-based reinforcement learning methods. We systematically analyze and optimize the reward signal by investigating effective scoring strategies for turn-level aggregation, VLM reasoning modes to trade off reward bias and variance, and advantage fusion levels to prevent reward hacking. Our findings reveal that broadcasting the aggregated advantage across the entire editing trajectory effectively bridges the gap between local planning and global multi-turn task success. Extensive experiments demonstrate that MT-EditFlow significantly improves performance across diverse base models. Notably, it boosts FLUX.1-Kontext-dev by 6.85 points in turn-3 overall performance, surpassing state-of-the-art open-source models such as Qwen-Image-Edit. By maintaining high marginal success rates and reducing exposure bias, MT-EditFlow provides a foundation for more reliable and natural human-AI collaboration in visual content creation.
Jiahui Huang, Yasi Zhang, Tianyu Chen +6
Apple · University of California, Los Angeles · University of Texas at Austin +1