- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper focuses on multi-turn image editing, where an image is edited repeatedly based on a sequence of instructions. The authors highlight a common problem with existing approaches: generating each edited image from the previous one can gradually introduce visual artifacts, change important details, and reduce image quality. To address this, the research propose a data generation strategy called Packed Trajectory Canvas (PTC). Instead of generating each editing step separately, PTC uses a vision-language model (VLM) to prepare a prompt describing four progressive edits. An image-generation model then takes the original image and this prompt to generate a single 2×2 canvas containing all four editing states. The canvas is divided into individual images, which are aligned using SIFT-based matching and warping to create consistent editing pairs. The authors also use automated filtering to improve the quality of the generated data. Using this approach, they construct MTEdit, a large-scale dataset containing 500K editing trajectories and 2.69M edit pairs, covering both step-wise and context-aware editing scenarios. They use this dataset to fine-tune Qwen-Image-Edit-Causal and report improvements in instruction following, content consistency, and visual quality across multiple editing turns. The paper also introduces MTEdit-Bench to evaluate these aspects of multi-turn editing, including how well models use information from previous instructions. Finally, the authors conduct a controlled ablation study comparing PTC with traditional sequential rollout generation to examine whether generating multiple editing states together provides better consistency and reduces visual degradation.
Critical strengths
-
The paper presents a practical way to construct multi-turn editing data and tests its main idea directly. PTC produces several editing states together, which may reduce the errors carried forward between turns. The comparison in Table 3 uses the same source images, prompts, base model and training budget for PTC and sequential rollout. At Turn 8, PTC achieves higher instruction-following scores (5.06 versus 4.26) and drift scores (4.22 versus 3.96). The data-level results in Table 4 also favor PTC. This controlled comparison is the strongest evidence for the proposed data-construction method.
-
The dataset addresses a useful gap in multi-turn editing. MTEdit contains 500K trajectories and includes both step-wise edits and requests that depend on earlier instructions, such as keeping a previous constraint or referring to an object from an earlier turn. Among the datasets compared in Table 1, it is the only one covering both settings. The project page is well organized, with full trajectory examples, a PTC-versus-rollout comparison and a data explorer, which makes it easier to judge data quality beyond the figures in the paper. This scope matters because a multi-turn editor must preserve earlier changes while following new requests.
-
The work has practical value for image-editing applications. In photo editing and content creation, users often make several changes to an image while expecting earlier edits to remain intact. This paper addresses this practical need by improving consistency across multiple editing turns.
Critical weaknesses
-
The headline model gains do not isolate the dataset's contribution. Table 3 gives useful evidence that PTC data are better than sequential-rollout data under the tested setup. The larger improvements over the base editor in Table 2 involve other changes: the proposed model uses editing history and architectural adaptations, including block-causal attention and interleaved RoPE (Section 3.5). Without training the same architecture on another multi-turn dataset at a comparable scale, it is hard to tell how much of the headline gain comes from MTEdit. The main evaluation also uses the authors' own benchmark, so performance on independent editing tasks remains unclear. This limits the dataset-level claim and leaves the controlled PTC ablation intact.
-
The evaluation leaves some uncertainty about close comparisons between models. Instruction following and content consistency are scored by one VLM judge, Qwen 3.5 Plus, from the same model family as the proposed editor. The agreement with human judgments in Table 8 helps, but an independent judge and confidence intervals would make small differences between models more convincing. The drift score measures visual quality and does not check whether the requested edit was completed, so it should be read alongside instruction following. These concerns matter mainly for the close rankings in Tables 2 and 5–7. The PTC-versus-rollout gains in Table 3 are more consistent.
Questions for authors
-
How many annotators and samples were used for the human-validation results in Appendix A.4?
-
Could the authors compare MTEdit with another multi-turn editing dataset using the same model, data size, and training settings?
-
Would the model rankings remain similar if a different VLM were used for evaluation? Could the authors also report confidence intervals to show how reliable the results are?
-
How well does PTC preserve parts of the image that should not change between editing turns? Could the authors provide a metric or a small human evaluation to support this?
9 comments
Step-wise — GPT-6.1 Sol
| Model | T1 | T2 | T3 | T4 | T5 | T6 | T7 | T8 |
|---|---|---|---|---|---|---|---|---|
| Nano Banana Pro | 9.17 / 8.57 | 9.04 / 8.12 | 9.09 / 7.73 | 9.09 / 7.54 | 9.05 / 7.35 | 8.84 / 7.18 | 8.81 / 7.05 | 8.96 / 6.83 |
| Nano Banana 2 | 9.18 / 8.71 | 8.41 / 7.72 | 8.24 / 7.20 | 8.23 / 6.86 | 8.07 / 6.51 | 7.98 / 6.37 | 7.48 / 6.25 | 7.51 / 5.98 |
| GPT Image 2 | 8.43 / 7.78 | 8.32 / 7.07 | 8.27 / 6.57 | 8.38 / 6.30 | 8.41 / 6.10 | 8.07 / 5.93 | 7.59 / 5.51 | 7.31 / 5.09 |
| Seedream 5.0 Lite | 9.43 / 8.92 | 9.45 / 8.01 | 9.30 / 7.38 | 9.22 / 6.78 | 9.01 / 6.30 | 8.92 / 6.13 | 8.67 / 5.94 | 8.84 / 5.89 |
| VINCIE | 1.69 / 6.27 | 0.14 / 0.33 | 0.02 / 0.05 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 |
| OmniGen2 | 5.94 / 6.96 | 3.72 / 3.81 | 2.53 / 2.35 | 1.81 / 1.59 | 0.85 / 0.86 | 0.54 / 0.49 | 0.36 / 0.28 | 0.11 / 0.11 |
| FLUX.1-kontext-dev | 7.39 / 7.95 | 6.06 / 6.31 | 5.35 / 5.09 | 4.33 / 4.27 | 3.81 / 3.55 | 3.12 / 2.87 | 2.56 / 2.34 | 2.34 / 1.94 |
| FLUX.2-klein-9B | 8.45 / 8.62 | 7.52 / 7.49 | 7.12 / 6.82 | 7.41 / 6.62 | 6.73 / 6.03 | 6.44 / 5.70 | 5.74 / 5.13 | 5.59 / 4.89 |
| Qwen-Image-Edit-2511 | 8.02 / 7.98 | 8.28 / 7.32 | 8.02 / 6.74 | 7.89 / 6.39 | 7.71 / 6.14 | 7.17 / 5.89 | 6.99 / 5.46 | 6.96 / 5.31 |
| BAGEL | 8.05 / 8.75 | 7.36 / 7.82 | 6.47 / 7.06 | 5.09 / 5.90 | 4.89 / 4.80 | 3.87 / 4.37 | 3.08 / 3.14 | 2.06 / 2.18 |
| Emu3.5 | 7.08 / 7.29 | 6.71 / 6.62 | 5.98 / 5.87 | 5.46 / 5.25 | 4.70 / 4.40 | 4.28 / 3.93 | 3.51 / 3.11 | 3.17 / 2.74 |
| SenseNova-U1 | 8.63 / 8.38 | 7.59 / 7.56 | 7.00 / 6.59 | 6.59 / 5.71 | 5.90 / 4.79 | 5.78 / 4.50 | 4.94 / 4.13 | 4.41 / 3.71 |
| Lance | 6.64 / 6.88 | 3.32 / 3.60 | 0.79 / 0.84 | 0.51 / 0.52 | 0.13 / 0.10 | 0.05 / 0.06 | 0.00 / 0.00 | 0.00 / 0.00 |
| Qwen-Image-Edit-Causal | 7.47 / 7.56 | 7.53 / 6.63 | 7.12 / 5.96 | 6.91 / 5.28 | 6.52 / 4.88 | 6.09 / 4.64 | 5.52 / 4.27 | 4.79 / 3.60 |
| Ours | 8.03 / 7.75 | 8.94 / 7.61 | 7.98 / 7.48 | 8.28 / 7.15 | 8.14 / 6.58 | 7.87 / 6.54 | 6.67 / 6.53 | 7.21 / 6.03 |
Step-wise — Gemini 3.1 Pro
| Model | T1 | T2 | T3 | T4 | T5 | T6 | T7 | T8 |
|---|---|---|---|---|---|---|---|---|
| Nano Banana Pro | 8.12 / 8.15 | 7.58 / 7.70 | 7.19 / 7.27 | 7.12 / 6.63 | 6.32 / 5.84 | 5.93 / 5.06 | 4.45 / 3.82 | 3.27 / 2.76 |
| Nano Banana 2 | 7.72 / 7.82 | 7.19 / 7.38 | 6.64 / 6.38 | 6.01 / 5.86 | 5.07 / 4.76 | 4.22 / 3.84 | 3.95 / 3.44 | 3.17 / 2.78 |
| GPT Image 2 | 7.51 / 7.51 | 7.05 / 6.83 | 6.88 / 6.38 | 6.81 / 5.81 | 5.74 / 4.76 | 4.75 / 3.70 | 3.07 / 2.58 | 2.06 / 1.69 |
| Seedream 5.0 Lite | 7.76 / 7.73 | 7.57 / 7.58 | 7.36 / 6.40 | 6.22 / 4.98 | 4.54 / 3.49 | 2.61 / 1.72 | 1.06 / 0.69 | 0.24 / 0.16 |
| VINCIE | 1.07 / 3.98 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 |
| OmniGen2 | 5.06 / 6.32 | 2.02 / 2.35 | 0.63 / 0.57 | 0.12 / 0.09 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 |
| FLUX.1-kontext-dev | 6.61 / 7.28 | 4.96 / 5.40 | 3.75 / 4.12 | 2.64 / 2.84 | 1.65 / 1.53 | 0.96 / 0.94 | 0.45 / 0.41 | 0.28 / 0.18 |
| FLUX.2-klein-9B | 7.16 / 7.69 | 6.20 / 6.78 | 5.43 / 5.62 | 4.86 / 4.79 | 4.16 / 3.93 | 3.42 / 3.13 | 2.26 / 2.01 | 1.68 / 1.37 |
| Qwen-Image-Edit-2511 | 7.30 / 7.65 | 6.66 / 6.66 | 6.11 / 5.88 | 5.06 / 4.48 | 4.31 / 3.77 | 3.91 / 3.14 | 2.94 / 2.30 | 2.23 / 1.69 |
| BAGEL | 6.35 / 7.30 | 4.79 / 5.62 | 3.72 / 4.15 | 2.54 / 3.06 | 1.64 / 1.73 | 0.81 / 1.03 | 0.44 / 0.45 | 0.16 / 0.15 |
| Emu3.5 | 5.34 / 6.17 | 2.67 / 2.63 | 2.10 / 2.18 | 1.44 / 1.32 | 1.05 / 1.03 | 0.80 / 0.72 | 0.62 / 0.61 | 0.42 / 0.38 |
| SenseNova-U1 | 7.16 / 7.51 | 6.32 / 6.19 | 5.01 / 4.40 | 3.41 / 2.62 | 1.22 / 0.95 | 0.13 / 0.10 | 0.00 / 0.00 | 0.00 / 0.00 |
| Lance | 5.74 / 6.47 | 2.39 / 2.19 | 0.31 / 0.27 | 0.06 / 0.05 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 |
| Qwen-Image-Edit-Causal | 6.77 / 7.33 | 5.44 / 5.31 | 4.80 / 4.34 | 3.48 / 3.00 | 2.24 / 1.75 | 0.97 / 0.78 | 0.84 / 0.65 | 0.35 / 0.29 |
| Ours | 7.30 / 7.43 | 7.04 / 7.12 | 6.55 / 6.42 | 5.92 / 5.57 | 5.23 / 4.53 | 4.59 / 3.94 | 3.21 / 3.04 | 2.73 / 2.34 |
Persistent-constraint — GPT-6.1 Sol
| Model | T1 | T2 | T3 | T4 |
|---|---|---|---|---|
| Nano Banana 2 | 7.46 / 7.80 | 7.36 / 6.80 | 6.76 / 6.10 | 6.12 / 5.40 |
| Nano Banana Pro | 7.70 / 7.74 | 7.16 / 6.50 | 6.84 / 6.06 | 6.76 / 5.70 |
| GPT Image 2 | 7.54 / 7.44 | 7.98 / 6.96 | 6.90 / 6.26 | 6.12 / 5.42 |
| Seedream 5.0 Lite | 8.88 / 7.90 | 8.06 / 7.60 | 7.16 / 6.70 | 6.54 / 6.04 |
| VINCIE | 0.94 / 6.06 | 0.02 / 0.04 | 0.00 / 0.00 | 0.00 / 0.00 |
| BAGEL | 7.04 / 8.24 | 4.72 / 5.90 | 2.98 / 4.08 | 2.20 / 2.60 |
| Emu3.5 | 6.36 / 6.58 | 4.70 / 5.10 | 4.16 / 4.38 | 2.92 / 2.82 |
| SenseNova-U1 | 8.04 / 7.56 | 6.48 / 6.66 | 4.86 / 4.82 | 3.76 / 3.80 |
| Lance | 5.62 / 6.00 | 2.00 / 2.18 | 0.42 / 0.40 | 0.10 / 0.10 |
| Ours | 8.00 / 7.52 | 8.56 / 7.22 | 7.08 / 6.58 | 6.90 / 5.88 |
Persistent-constraint — Gemini 3.1 Pro
| Model | T1 | T2 | T3 | T4 |
|---|---|---|---|---|
| Nano Banana 2 | 7.08 / 7.32 | 6.18 / 5.94 | 5.46 / 5.32 | 4.56 / 4.50 |
| Nano Banana Pro | 6.86 / 7.08 | 5.88 / 5.84 | 5.38 / 5.26 | 4.64 / 4.60 |
| GPT Image 2 | 7.18 / 7.54 | 6.70 / 7.06 | 5.80 / 6.16 | 5.04 / 5.20 |
| Seedream 5.0 Lite | 7.92 / 8.02 | 7.42 / 7.74 | 6.36 / 6.82 | 5.24 / 5.52 |
| VINCIE | 2.48 / 4.94 | 0.00 / 0.00 | 0.00 / 0.00 | 0.00 / 0.00 |
| BAGEL | 6.78 / 7.30 | 4.04 / 4.96 | 2.10 / 2.68 | 1.18 / 1.56 |
| Emu3.5 | 6.54 / 6.58 | 4.18 / 4.94 | 2.20 / 2.50 | 1.28 / 1.40 |
| SenseNova-U1 | 7.26 / 7.26 | 6.22 / 6.18 | 4.28 / 4.40 | 2.20 / 2.34 |
| Lance | 6.12 / 6.30 | 2.54 / 3.18 | 0.66 / 0.66 | 0.08 / 0.20 |
| Ours | 7.78 / 7.94 | 7.80 / 7.72 | 6.78 / 7.06 | 5.92 / 5.88 |
Textual-reference — GPT-6.1 Sol
| Model | T1 | T2 | T3 | T4 |
|---|---|---|---|---|
| Nano Banana 2 | 8.80 / 8.32 | 8.40 / 7.58 | 7.96 / 7.16 | 7.62 / 6.50 |
| Nano Banana Pro | 8.50 / 7.94 | 8.50 / 7.88 | 7.96 / 7.08 | 7.68 / 6.64 |
| GPT Image 2 | 8.28 / 7.70 | 8.38 / 7.18 | 7.44 / 6.80 | 6.86 / 6.00 |
| Seedream 5.0 Lite | 8.66 / 7.80 | 7.28 / 7.70 | 6.62 / 6.42 | 6.28 / 5.80 |
| VINCIE | 2.86 / 6.00 | 0.76 / 0.82 | 0.26 / 0.22 | 0.20 / 0.16 |
| BAGEL | 7.80 / 8.40 | 4.12 / 7.56 | 2.36 / 3.40 | 1.46 / 2.04 |
| Emu3.5 | 6.16 / 6.40 | 4.96 / 4.82 | 3.54 / 3.50 | 2.56 / 2.36 |
| SenseNova-U1 | 8.67 / 7.98 | 7.39 / 7.31 | 6.90 / 6.43 | 6.27 / 5.61 |
| Lance | 5.52 / 6.00 | 1.96 / 2.00 | 0.40 / 0.40 | 0.00 / 0.00 |
| Ours | 8.20 / 7.74 | 8.66 / 7.50 | 7.76 / 7.36 | 7.90 / 6.94 |
Textual-reference — Gemini 3.1 Pro
| Model | T1 | T2 | T3 | T4 |
|---|---|---|---|---|
| Nano Banana 2 | 7.62 / 7.52 | 7.40 / 7.18 | 6.88 / 6.84 | 6.04 / 6.04 |
| Nano Banana Pro | 7.48 / 7.52 | 7.34 / 7.40 | 6.92 / 7.04 | 5.98 / 6.10 |
| GPT Image 2 | 7.48 / 7.76 | 7.30 / 7.14 | 6.62 / 6.76 | 6.12 / 5.92 |
| Seedream 5.0 Lite | 7.40 / 7.46 | 6.74 / 7.16 | 6.08 / 6.36 | 5.30 / 5.18 |
| VINCIE | 3.08 / 4.86 | 0.34 / 0.40 | 0.14 / 0.14 | 0.14 / 0.14 |
| BAGEL | 6.60 / 7.14 | 4.08 / 5.16 | 2.48 / 2.60 | 1.52 / 1.80 |
| Emu3.5 | 6.48 / 6.58 | 4.16 / 4.58 | 2.60 / 3.04 | 1.74 / 1.80 |
| SenseNova-U1 | 7.20 / 7.36 | 6.42 / 6.34 | 5.38 / 5.30 | 3.06 / 3.26 |
| Lance | 6.28 / 6.38 | 3.60 / 3.86 | 1.38 / 1.46 | 0.42 / 0.38 |
| Ours | 7.54 / 7.72 | 7.14 / 7.04 | 6.38 / 6.44 | 5.54 / 5.82 |
Short-instruction — GPT-6.1 Sol
| Model | T1 | T2 | T3 | T4 |
|---|---|---|---|---|
| Nano Banana 2 | 8.64 / 8.24 | 8.62 / 7.92 | 7.88 / 6.82 | 6.60 / 5.82 |
| Nano Banana Pro | 8.78 / 8.22 | 8.38 / 7.80 | 7.68 / 7.04 | 5.52 / 5.74 |
| GPT Image 2 | 8.34 / 7.56 | 8.64 / 7.28 | 7.56 / 6.60 | 6.82 / 5.94 |
| Seedream 5.0 Lite | 8.50 / 7.86 | 7.20 / 7.52 | 6.78 / 5.84 | 5.46 / 5.20 |
| VINCIE | 2.96 / 6.58 | 0.98 / 1.38 | 0.00 / 0.00 | 0.00 / 0.00 |
| BAGEL | 7.00 / 8.50 | 3.34 / 6.72 | 2.04 / 2.98 | 1.34 / 1.68 |
| Emu3.5 | 6.50 / 6.60 | 4.10 / 4.40 | 2.50 / 2.46 | 1.56 / 1.66 |
| SenseNova-U1 | 8.52 / 7.90 | 7.52 / 7.14 | 7.00 / 6.30 | 5.34 / 5.26 |
| Lance | 5.80 / 6.20 | 2.20 / 2.54 | 0.86 / 0.70 | 0.08 / 0.06 |
| Ours | 8.30 / 7.90 | 9.20 / 7.86 | 8.18 / 7.38 | 7.00 / 6.78 |
Reviewer Lstb
We appreciate your recognition of our work and your valuable suggestions for strengthening the paper. Below, we respond to each of your concerns in detail.
Q1:
How many annotators and samples were used for the human-validation results in Appendix A.4 of the revised paper?
A total of 20 annotators participated in the human evaluation. We sampled 30% of the cases from each sub-benchmark and included outputs from all models evaluated on that sub-benchmark. These details have been added to Appendix A.4 of the revised paper.
Q2:
Could the authors compare MTEdit with another multi-turn editing dataset using the same model, data size, and training settings?
We have added a comparison with Pico-Banana-400K, an existing multi-turn editing dataset. We use 10,000 trajectories from each of three training-data sources: PTC-generated MTEdit data, sequential rollout data, and Pico-Banana-400K. We separately fine-tune the same Qwen-Image-Edit-Causal base model for 5,000 steps with identical training hyperparameters. This comparison matches the number of training trajectories and the training schedule.
We evaluate the resulting checkpoints on both the step-wise sub-benchmark of MTEdit-Bench and the independently developed MSE-Bench.
Step-wise
Each cell reports IF / CC / DRS. The best scores are bolded.
PTC-generated data yields higher IF from Turn 2 onward, higher CC at every turn, and higher DRS from Turn 3 onward than both alternatives. At Turn 8, PTC achieves IF / CC / DRS of 5.06 / 4.44 / 4.22, compared with 1.03 / 2.27 / 3.79 for Pico-Banana-400K.
MSE-Bench
We also evaluate the same checkpoints on MSE-Bench using GPT-4o and its cumulative success-rate protocol.
PTC achieves the highest cumulative success rates at Turns 1 and 3–5. Its SR@5 is 0.65, compared with 0.54 for sequential rollout and 0.50 for Pico-Banana-400K. These results support the training utility of PTC-generated data on both our benchmark and an independent benchmark with a different evaluator.
In the revised paper, the experimental settings and step-wise results are provided in Section 4.3, and the MSE-Bench ablation is reported in Appendix A.3.6.
Q3
Would the model rankings remain similar if a different VLM were used for evaluation?
Yes. We additionally evaluated the models using GPT-6.1 Sol and Gemini 3.1 Pro as VLM judges, following the same evaluation protocol. Although absolute scores and some individual-turn rankings vary across judges, the main conclusion that our model leads the evaluated open-source methods remains consistent. These results have been added to Appendix A.9 of the revised paper.
Each cell reports IF / CC for step-wise editing and IF* / CC* for context-aware editing. Closed-source results are italicized; the best scores among open-source models are bolded.