acceptodds
Under review as a conference paper at ICLR 2027

MTEdit: Packed Trajectories Beat Rollouts for Multi-Turn Image Editing

Abstract

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. Code and data will be publicly available upon acceptance. For additional results, see https://sadlksjada.github.io/mtedit.github.io/.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multi turn image editing, image editing dataset and benchmark

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_LstbOpenReview ↗
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
  1. 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.

  2. 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.

  3. 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
  1. 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.

  2. 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
  1. How many annotators and samples were used for the human-validation results in Appendix A.4?

  2. Could the authors compare MTEdit with another multi-turn editing dataset using the same model, data size, and training settings?

  3. 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?

  4. 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

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.

Training-data sourceT1T2T3T4T5T6T7T8
PTC9.64 / 9.61 / 4.179.63 / 9.27 / 4.269.30 / 8.59 / 4.218.67 / 7.67 / 4.228.21 / 7.15 / 4.217.63 / 6.47 / 4.216.24 / 5.37 / 4.235.06 / 4.44 / 4.22
Sequential rollout9.66 / 9.50 / 4.199.32 / 9.04 / 4.278.97 / 8.20 / 4.187.95 / 7.58 / 4.197.66 / 6.66 / 4.164.21 / 6.16 / 4.034.92 / 4.97 / 4.034.26 / 3.84 / 3.96
Pico-Banana-400K9.61 / 9.57 / 4.179.55 / 9.01 / 4.238.86 / 8.35 / 4.138.03 / 7.48 / 4.017.15 / 6.46 / 4.115.00 / 5.45 / 3.993.54 / 4.21 / 3.881.03 / 2.27 / 3.79

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.

Training-data sourceSR@1SR@2SR@3SR@4SR@5
PTC0.950.820.780.680.65
Sequential rollout0.920.830.740.620.54
Pico-Banana-400K0.940.810.740.650.50

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.

Step-wise — GPT-6.1 Sol

ModelT1T2T3T4T5T6T7T8
Nano Banana Pro9.17 / 8.579.04 / 8.129.09 / 7.739.09 / 7.549.05 / 7.358.84 / 7.188.81 / 7.058.96 / 6.83
Nano Banana 29.18 / 8.718.41 / 7.728.24 / 7.208.23 / 6.868.07 / 6.517.98 / 6.377.48 / 6.257.51 / 5.98
GPT Image 28.43 / 7.788.32 / 7.078.27 / 6.578.38 / 6.308.41 / 6.108.07 / 5.937.59 / 5.517.31 / 5.09
Seedream 5.0 Lite9.43 / 8.929.45 / 8.019.30 / 7.389.22 / 6.789.01 / 6.308.92 / 6.138.67 / 5.948.84 / 5.89
VINCIE1.69 / 6.270.14 / 0.330.02 / 0.050.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.00
OmniGen25.94 / 6.963.72 / 3.812.53 / 2.351.81 / 1.590.85 / 0.860.54 / 0.490.36 / 0.280.11 / 0.11
FLUX.1-kontext-dev7.39 / 7.956.06 / 6.315.35 / 5.094.33 / 4.273.81 / 3.553.12 / 2.872.56 / 2.342.34 / 1.94
FLUX.2-klein-9B8.45 / 8.627.52 / 7.497.12 / 6.827.41 / 6.626.73 / 6.036.44 / 5.705.74 / 5.135.59 / 4.89
Qwen-Image-Edit-25118.02 / 7.988.28 / 7.328.02 / 6.747.89 / 6.397.71 / 6.147.17 / 5.896.99 / 5.466.96 / 5.31
BAGEL8.05 / 8.757.36 / 7.826.47 / 7.065.09 / 5.904.89 / 4.803.87 / 4.373.08 / 3.142.06 / 2.18
Emu3.57.08 / 7.296.71 / 6.625.98 / 5.875.46 / 5.254.70 / 4.404.28 / 3.933.51 / 3.113.17 / 2.74
SenseNova-U18.63 / 8.387.59 / 7.567.00 / 6.596.59 / 5.715.90 / 4.795.78 / 4.504.94 / 4.134.41 / 3.71
Lance6.64 / 6.883.32 / 3.600.79 / 0.840.51 / 0.520.13 / 0.100.05 / 0.060.00 / 0.000.00 / 0.00
Qwen-Image-Edit-Causal7.47 / 7.567.53 / 6.637.12 / 5.966.91 / 5.286.52 / 4.886.09 / 4.645.52 / 4.274.79 / 3.60
Ours8.03 / 7.758.94 / 7.617.98 / 7.488.28 / 7.158.14 / 6.587.87 / 6.546.67 / 6.537.21 / 6.03

Step-wise — Gemini 3.1 Pro

ModelT1T2T3T4T5T6T7T8
Nano Banana Pro8.12 / 8.157.58 / 7.707.19 / 7.277.12 / 6.636.32 / 5.845.93 / 5.064.45 / 3.823.27 / 2.76
Nano Banana 27.72 / 7.827.19 / 7.386.64 / 6.386.01 / 5.865.07 / 4.764.22 / 3.843.95 / 3.443.17 / 2.78
GPT Image 27.51 / 7.517.05 / 6.836.88 / 6.386.81 / 5.815.74 / 4.764.75 / 3.703.07 / 2.582.06 / 1.69
Seedream 5.0 Lite7.76 / 7.737.57 / 7.587.36 / 6.406.22 / 4.984.54 / 3.492.61 / 1.721.06 / 0.690.24 / 0.16
VINCIE1.07 / 3.980.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.00
OmniGen25.06 / 6.322.02 / 2.350.63 / 0.570.12 / 0.090.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.00
FLUX.1-kontext-dev6.61 / 7.284.96 / 5.403.75 / 4.122.64 / 2.841.65 / 1.530.96 / 0.940.45 / 0.410.28 / 0.18
FLUX.2-klein-9B7.16 / 7.696.20 / 6.785.43 / 5.624.86 / 4.794.16 / 3.933.42 / 3.132.26 / 2.011.68 / 1.37
Qwen-Image-Edit-25117.30 / 7.656.66 / 6.666.11 / 5.885.06 / 4.484.31 / 3.773.91 / 3.142.94 / 2.302.23 / 1.69
BAGEL6.35 / 7.304.79 / 5.623.72 / 4.152.54 / 3.061.64 / 1.730.81 / 1.030.44 / 0.450.16 / 0.15
Emu3.55.34 / 6.172.67 / 2.632.10 / 2.181.44 / 1.321.05 / 1.030.80 / 0.720.62 / 0.610.42 / 0.38
SenseNova-U17.16 / 7.516.32 / 6.195.01 / 4.403.41 / 2.621.22 / 0.950.13 / 0.100.00 / 0.000.00 / 0.00
Lance5.74 / 6.472.39 / 2.190.31 / 0.270.06 / 0.050.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.00
Qwen-Image-Edit-Causal6.77 / 7.335.44 / 5.314.80 / 4.343.48 / 3.002.24 / 1.750.97 / 0.780.84 / 0.650.35 / 0.29
Ours7.30 / 7.437.04 / 7.126.55 / 6.425.92 / 5.575.23 / 4.534.59 / 3.943.21 / 3.042.73 / 2.34

Persistent-constraint — GPT-6.1 Sol

ModelT1T2T3T4
Nano Banana 27.46 / 7.807.36 / 6.806.76 / 6.106.12 / 5.40
Nano Banana Pro7.70 / 7.747.16 / 6.506.84 / 6.066.76 / 5.70
GPT Image 27.54 / 7.447.98 / 6.966.90 / 6.266.12 / 5.42
Seedream 5.0 Lite8.88 / 7.908.06 / 7.607.16 / 6.706.54 / 6.04
VINCIE0.94 / 6.060.02 / 0.040.00 / 0.000.00 / 0.00
BAGEL7.04 / 8.244.72 / 5.902.98 / 4.082.20 / 2.60
Emu3.56.36 / 6.584.70 / 5.104.16 / 4.382.92 / 2.82
SenseNova-U18.04 / 7.566.48 / 6.664.86 / 4.823.76 / 3.80
Lance5.62 / 6.002.00 / 2.180.42 / 0.400.10 / 0.10
Ours8.00 / 7.528.56 / 7.227.08 / 6.586.90 / 5.88

Persistent-constraint — Gemini 3.1 Pro

ModelT1T2T3T4
Nano Banana 27.08 / 7.326.18 / 5.945.46 / 5.324.56 / 4.50
Nano Banana Pro6.86 / 7.085.88 / 5.845.38 / 5.264.64 / 4.60
GPT Image 27.18 / 7.546.70 / 7.065.80 / 6.165.04 / 5.20
Seedream 5.0 Lite7.92 / 8.027.42 / 7.746.36 / 6.825.24 / 5.52
VINCIE2.48 / 4.940.00 / 0.000.00 / 0.000.00 / 0.00
BAGEL6.78 / 7.304.04 / 4.962.10 / 2.681.18 / 1.56
Emu3.56.54 / 6.584.18 / 4.942.20 / 2.501.28 / 1.40
SenseNova-U17.26 / 7.266.22 / 6.184.28 / 4.402.20 / 2.34
Lance6.12 / 6.302.54 / 3.180.66 / 0.660.08 / 0.20
Ours7.78 / 7.947.80 / 7.726.78 / 7.065.92 / 5.88

Textual-reference — GPT-6.1 Sol

ModelT1T2T3T4
Nano Banana 28.80 / 8.328.40 / 7.587.96 / 7.167.62 / 6.50
Nano Banana Pro8.50 / 7.948.50 / 7.887.96 / 7.087.68 / 6.64
GPT Image 28.28 / 7.708.38 / 7.187.44 / 6.806.86 / 6.00
Seedream 5.0 Lite8.66 / 7.807.28 / 7.706.62 / 6.426.28 / 5.80
VINCIE2.86 / 6.000.76 / 0.820.26 / 0.220.20 / 0.16
BAGEL7.80 / 8.404.12 / 7.562.36 / 3.401.46 / 2.04
Emu3.56.16 / 6.404.96 / 4.823.54 / 3.502.56 / 2.36
SenseNova-U18.67 / 7.987.39 / 7.316.90 / 6.436.27 / 5.61
Lance5.52 / 6.001.96 / 2.000.40 / 0.400.00 / 0.00
Ours8.20 / 7.748.66 / 7.507.76 / 7.367.90 / 6.94

Textual-reference — Gemini 3.1 Pro

ModelT1T2T3T4
Nano Banana 27.62 / 7.527.40 / 7.186.88 / 6.846.04 / 6.04
Nano Banana Pro7.48 / 7.527.34 / 7.406.92 / 7.045.98 / 6.10
GPT Image 27.48 / 7.767.30 / 7.146.62 / 6.766.12 / 5.92
Seedream 5.0 Lite7.40 / 7.466.74 / 7.166.08 / 6.365.30 / 5.18
VINCIE3.08 / 4.860.34 / 0.400.14 / 0.140.14 / 0.14
BAGEL6.60 / 7.144.08 / 5.162.48 / 2.601.52 / 1.80
Emu3.56.48 / 6.584.16 / 4.582.60 / 3.041.74 / 1.80
SenseNova-U17.20 / 7.366.42 / 6.345.38 / 5.303.06 / 3.26
Lance6.28 / 6.383.60 / 3.861.38 / 1.460.42 / 0.38
Ours7.54 / 7.727.14 / 7.046.38 / 6.445.54 / 5.82

Short-instruction — GPT-6.1 Sol

ModelT1T2T3T4
Nano Banana 28.64 / 8.248.62 / 7.927.88 / 6.826.60 / 5.82
Nano Banana Pro8.78 / 8.228.38 / 7.807.68 / 7.045.52 / 5.74
GPT Image 28.34 / 7.568.64 / 7.287.56 / 6.606.82 / 5.94
Seedream 5.0 Lite8.50 / 7.867.20 / 7.526.78 / 5.845.46 / 5.20
VINCIE2.96 / 6.580.98 / 1.380.00 / 0.000.00 / 0.00
BAGEL7.00 / 8.503.34 / 6.722.04 / 2.981.34 / 1.68
Emu3.56.50 / 6.604.10 / 4.402.50 / 2.461.56 / 1.66
SenseNova-U18.52 / 7.907.52 / 7.147.00 / 6.305.34 / 5.26
Lance5.80 / 6.202.20 / 2.540.86 / 0.700.08 / 0.06
Ours8.30 / 7.909.20 / 7.868.18 / 7.387.00 / 6.78
Reviewer_SXWYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses error accumulation and visual-quality drift in multi-turn image-editing data built by sequential rollout. Its core idea is packed trajectory canvas (PTC) generation. A VLM plans four cumulative edits, and a closed-source image model (GPT Image 2 or Nano Banana 2) renders them jointly as a 2×2 canvas. Gemini 3.1 Pro then filters failed canvases and reconstructs turn-level instructions. The panels are split and aligned with SIFT-based warping, and longer trajectories are produced by restarting from the fourth panel. Fine-tuning Qwen-Image-Edit-Causal yields the strongest open-source results on all sub-benchmarks (Tables 2, 5–7).

Critical strengths

This paper introduce the MTEdit dataset (500K trajectories, 2.69M edit pairs, up to 8 turns) covers step-wise editing and three context-aware settings and MTEdit-Bench, which uses VLM-judged instruction following (IF) and content consistency (CC) with failure propagation, plus a no-reference quality score (DRS). Fine-tuning Qwen-Image-Edit-Causal yields the strongest open-source results on all sub-benchmarks.

Critical weaknesses
  1. The evaluation protocol is not fully substantiated.

    • CC* protocol. Section 3.4 (p. 8) and A.8.9 describe CC* as a joint judgment over the full trajectory. In the supplementary records, however, the textual-reference and short-instruction CC scores come from the "original" report in the per-turn format of A.8.7; only the persistent-constraint CC scores carry the context-aware protocol tag.
    • DRS. DRS is an absolute no-reference score that is never zeroed after failure, so it does not measure drift relative to the source. In Table 2, Lance keeps DRS 4.11–4.30 at Turns 4–8 while its IF falls to ≤2.59. DRS stability alone therefore cannot support the drift-resistance claim.
    • Human validation. The human study (A.4/Table 8) omits the number of cases, annotators, and ranked models.

    These points are fixable by documenting the final protocol and the human-study design.

  2. Use of historical context is not isolated. The main comparison changes both the architecture (A.1) and access to history: the base model sees only the previous image (§4.1) and is absent from the contextual sub-benchmarks. The data ablation covers step-wise editing only. Some contextual cases also do not require history. Case 06_persistent_rules_000005 repeats "yellow" in every turn, and case 06_persistent_rules_000029 states each color explicitly without a persistent rule. Evaluating the same checkpoint with and without earlier turns would resolve this.

  3. Benchmark construction, independence, and data quality are unclear. Section 3.4 describes manually crafted prompts. Yet the supplementary textual-reference and short-instruction records list per-turn gt_images (e.g., step…-rank…-seed…-turnNN-gt.jpg). Some prompts also read like descriptions of generated outputs, e.g., "(side effect: the left lamp head of the streetlight disappears accidentally)" (02_textual_reference_000067). If the benchmark was derived with a pipeline similar to MTEdit's, it would favor the MTEdit-trained model. In addition, no training–benchmark source separation is described, and neither filter retention rates nor an audit of retained trajectories is reported (§3.2; Fig. 13). I am not asserting that leakage occurred, but generalization beyond MTEdit-Bench remains uncertain.

Questions for authors
  1. Which evaluation run produced each column of Tables 2 and 5–7? Were the CC* scores for the textual-reference and short-instruction sub-benchmarks computed with A.8.9 or A.8.7, and was the same protocol applied to all models? For Table 8, how many cases, annotators, and ranked models were used?

  2. In §4.3, how were "identical per-turn prompts" obtained, given that PTC instructions are reconstructed after generation? Was the same filtering applied to the rollout arm, and what is the variance across seeds? Please also verify the rollout Turn-6 IF value (4.21, rising to 4.92 at Turn 7). Separately, how do the contextual results of the same checkpoint change when earlier turns are withheld?

6 comments

Response to Reviewer SXWY

Thank you for your positive assessment of our work and for highlighting several key issues. Your constructive feedback has helped us improve the paper. We address your concerns point by point below.


Q1 and W1:

In the revised paper, which evaluation run produced each column of Tables 2 and 7–9? The textual-reference and short-instruction CC scores come from the "original" report in the per-turn format; were the CC* scores computed with the system prompt in Appendix A.8.9 or A.8.7? Was the same protocol applied to all models?

In the revised paper, IF and CC in Table 2 were computed using the system prompts in Appendices A.8.6 and A.8.7, respectively. IF* and CC* in Tables 7–9 were computed using the context-aware system prompts in Appendices A.8.8 and A.8.9, respectively. All models were evaluated using the same protocol.

The discrepancy arose during the preparation of the supplementary materials: although the tables in the main paper and appendix reported the latest evaluation results, we inadvertently included outdated result records for the textual-reference, persistent-constraint, and short-instruction tasks in the supplementary package. We apologize for this error and have corrected the supplementary materials to match the reported results. This also clarifies some of the discrepancies raised in W1 and W3.


For Table 14 of the revised paper, how many cases, annotators, and ranked models were used?

We sampled 30% of the cases from each sub-benchmark and included outputs from all models evaluated on that sub-benchmark. A total of 20 annotators participated in the study. We apologize for omitting these details and have added them to Appendix A.4 of the revised paper.


Q2 and W2:

how were "identical per-turn prompts" obtained?

Specifically, PTC generation and sequential rollout used the same planned per-turn editing instructions during data generation. The final training prompts were subsequently reconstructed from the actual visual changes. We have clarified this distinction in Section 4.3 of the revised paper. Thank you for pointing out this ambiguity.


Was the same filtering applied to the rollout arm?

Yes. We applied the same quality-filtering procedure to the sequential rollout data and likewise reconstructed the turn-level training prompts from the actual visual changes.


What is the variance across seeds?

For the PTC and sequential rollout settings in Table 4 of the revised paper, we report the turn-wise standard deviations of IF and CC across three different random seeds.

TurnPTC: IF SDPTC: CC SDSequential rollout: IF SDSequential rollout: CC SD
10.0410.0420.0150.057
20.0650.0500.0760.015
30.0710.0710.0750.040
40.0590.0490.0350.036
50.0620.1070.0490.066
60.0990.0310.0850.055
70.0880.0870.0670.097
80.0210.0700.0910.084

Please also verify the rollout Turn-6 IF value (4.21, rising to 4.92 at Turn 7).

We have verified that the reported values—4.21 at Turn 6 and 4.92 at Turn 7—are correct. The failure-propagation rule does not require the average IF score to decrease monotonically: failed trajectories receive zero scores in subsequent turns, while surviving trajectories may achieve higher scores on later instructions.

For illustration, consider two hypothetical trajectories:

TrajectoryTurn-6 IFTurn-7 IF
A: fails at Turn 620
B: remains valid, with CC ≥ 6610
Average over both trajectories4.05.0

Thus, the average IF can increase even though the zeroing rule is strictly applied.

How do the contextual results of the same checkpoint change when earlier turns are withheld?Evaluating the same checkpoint with and without earlier turns.

We evaluated the same Qwen-Image-MTEdit checkpoint with and without historical context. With historical context, the model receives the source image, preceding instructions and generated images, and the current instruction. Without historical context, it receives only the latest image and the current instruction.

Each cell reports IF / CC / DRS for step-wise editing and IF* / CC* / DRS for context-aware editing. Better scores are bolded.

Step-wise

ModelT1T2T3T4T5T6T7T8
Ours9.74 / 9.54 / 4.239.59 / 9.28 / 4.269.32 / 9.19 / 4.309.35 / 8.82 / 4.338.99 / 8.12 / 4.328.70 / 7.60 / 4.338.00 / 7.51 / 4.357.45 / 6.42 / 4.33
Ours without context9.72 / 9.54 / 4.239.37 / 8.39 / 4.279.17 / 7.56 / 4.208.71 / 6.67 / 4.198.00 / 5.64 / 4.136.62 / 4.56 / 4.095.24 / 3.56 / 4.073.80 / 2.57 / 4.04

Persistent-constraint

ModelT1T2T3T4
Ours9.70 / 9.80 / 4.289.52 / 9.54 / 4.289.26 / 9.12 / 4.308.20 / 8.32 / 4.27
Ours without context9.68 / 9.82 / 4.287.82 / 8.82 / 4.256.64 / 6.80 / 4.175.90 / 5.82 / 4.07

Textual-reference

ModelT1T2T3T4
Ours9.42 / 9.54 / 4.109.10 / 9.30 / 4.208.40 / 9.04 / 4.188.24 / 8.12 / 4.17
Ours without context9.42 / 9.50 / 4.108.56 / 8.08 / 4.087.22 / 7.94 / 4.045.44 / 5.30 / 3.95

Short-instruction

ModelT1T2T3T4
Ours9.62 / 9.88 / 4.219.46 / 9.60 / 4.288.82 / 9.26 / 4.267.48 / 8.12 / 4.25
Ours without context9.64 / 9.84 / 4.217.84 / 7.78 / 4.116.70 / 6.24 / 3.904.62 / 4.90 / 3.83

Withholding historical context generally lowers IF, CC, and DRS, with the performance gap becoming more pronounced in later turns. For context-aware tasks, the absence of historical context can lead to edits that entirely miss the intended request, as illustrated in Figure 14 of the revised paper. A more detailed discussion is provided in Appendix A.3.5 of the revised paper.


W1:

Concern About DRS

DRS primarily measures visual quality, while CC is more appropriate for assessing unintended deviations from the source image. DRS should therefore be interpreted alongside IF and CC.

For Lance, DRS remains between 4.11 and 4.30 at Turns 4–8 even though IF falls to 2.59 or below. This indicates that visually plausible outputs can still fail to follow the requested edits. One possible explanation is that Lance’s autoregressive generation mechanism helps maintain visual quality when continuing from its own outputs, even as editing accuracy deteriorates. For the diffusion-based models evaluated in our experiments (excluding Lance), the observed DRS trends remain informative about visual-quality degradation across turns.

W2:

The base model sees only the previous image (Section 4.1 of the revised paper) and is absent from the contextual sub-benchmarks.

The results for Qwen-Image-Edit-Causal with and without historical context on all four sub-benchmarks are as follows. With historical context, the model receives the source image, preceding instructions and generated images, and the current instruction; without historical context, it receives only the latest image and current instruction.

All results below are for Qwen-Image-Edit-Causal. Each cell reports IF / CC / DRS for step-wise editing and IF* / CC* / DRS for context-aware editing.

A more detailed analysis of these results is provided in Appendix A.3.5 of the revised paper.

Step-wise

Historical contextT1T2T3T4T5T6T7T8
With context9.63 / 9.07 / 4.189.19 / 8.24 / 4.278.07 / 7.55 / 4.225.45 / 6.17 / 4.222.62 / 4.19 / 4.170.59 / 1.90 / 4.200.20 / 0.41 / 4.130.00 / 0.15 / 4.13
Without context9.66 / 9.11 / 4.188.76 / 8.21 / 4.228.54 / 7.26 / 4.207.75 / 6.38 / 4.176.87 / 5.20 / 4.155.74 / 4.29 / 4.134.11 / 3.21 / 4.123.17 / 2.41 / 4.06

Persistent-constraint

Historical contextT1T2T3T4
With context9.56 / 8.68 / 4.237.84 / 7.72 / 4.265.06 / 5.78 / 4.212.52 / 4.02 / 4.18
Without context9.52 / 8.66 / 4.237.82 / 7.88 / 4.236.50 / 6.60 / 4.156.00 / 5.52 / 4.10

Textual-reference

Historical contextT1T2T3T4
With context9.40 / 9.30 / 4.098.18 / 7.94 / 4.087.02 / 6.70 / 4.023.56 / 5.52 / 3.99
Without context9.44 / 9.30 / 4.098.52 / 8.18 / 3.967.26 / 7.14 / 3.935.76 / 4.80 / 3.80

Short-instruction

Historical contextT1T2T3T4
With context9.16 / 9.18 / 4.227.94 / 7.30 / 4.186.06 / 5.84 / 4.083.16 / 4.36 / 4.01
Without context9.26 / 9.40 / 4.217.74 / 7.86 / 4.086.50 / 5.70 / 3.964.30 / 4.20 / 3.84

The data ablation covers step-wise editing only.

We have extended the data ablation to all three context-aware tasks: persistent-constraint, textual-reference, and short-instruction editing. For each task, PTC generation and sequential rollout use the same planned per-turn editing instructions to construct 10,000 four-turn trajectories per method. Both datasets undergo the same filtering procedure, with training prompts reconstructed from the actual visual changes. We then fine-tune the same Qwen-Image-Edit-Causal base model on each dataset for 5,000 steps using identical training hyperparameters.

Each cell below reports IF* / CC* / DRS.

Persistent-constraint

Training-data sourceT1T2T3T4
PTC9.50 / 9.54 / 4.289.54 / 9.32 / 4.259.08 / 8.72 / 4.208.22 / 8.18 / 4.16
Sequential rollout9.42 / 9.24 / 4.229.62 / 9.02 / 4.238.50 / 8.22 / 4.167.90 / 8.18 / 4.02

Textual-reference

Training-data sourceT1T2T3T4
PTC9.34 / 9.28 / 4.129.02 / 9.20 / 4.178.46 / 8.40 / 4.148.00 / 8.12 / 4.15
Sequential rollout9.50 / 9.24 / 4.138.90 / 8.82 / 4.118.64 / 8.32 / 4.057.84 / 7.64 / 3.99

Short-instruction

Training-data sourceT1T2T3T4
PTC9.42 / 9.70 / 4.209.36 / 9.46 / 4.208.26 / 8.84 / 4.217.32 / 7.34 / 4.23
Sequential rollout9.46 / 9.50 / 4.238.90 / 9.12 / 4.187.89 / 8.32 / 3.997.10 / 6.94 / 4.08

PTC-generated data generally yields better performance in later turns. At Turn 4, it achieves higher IF* and DRS across all three tasks, while CC* is higher for textual-reference and short-instruction editing and tied for persistent-constraint editing. These controlled comparisons extend the evidence beyond step-wise editing and support the benefit of PTC-generated training data for sustaining context-aware editing performance. The results and detailed analysis have been added to Appendix A.3.6 of the revised paper.


Some contextual cases also do not require history.

This issue also arose because we inadvertently included an earlier version when preparing the supplementary materials. Thank you for bringing this to our attention. We have replaced the outdated materials with the latest results.

Reviewer_xSfEOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses a concrete bottleneck in multi-turn image editing: training trajectories built by sequential rollout accumulate errors and suffer visual-quality drift as turns increase. The proposed fix is a data-construction idea, i.e., packed trajectory canvas (PTC) generation: a VLM plans a 4-turn edit sequence, a closed-source editor renders all four cumulative states as panels of a single 2×2 canvas in one pass, and the panels are then split, instruction-annotated, and SIFT-aligned into training pairs. The authors release MTEdit, MTEdit-Bench, and Qwen-Image-MTEdit, a fine-tuned model that is best among evaluated open-source systems. Evidence includes a controlled data-construction comparison, a matched-budget training ablation, full benchmark results, and a human-alignment study of the benchmark.

Critical strengths

The root cause of rollout drift, each turn reuses the previous, already-degraded output as input, is correctly identified, and PTC removes it at the source by generating all turn states in one shared visual context. The idea is simple in hindsight, practically important, and the title's claim is directly tested rather than asserted.

MTEdit dominates the dataset comparison on trajectories, turns, and scenario coverage. It is the only compared resource covering both step-wise and context-aware editing at this scale, and the three context-aware subtypes (persistent-constraint, textual-reference, short-instruction) are well-chosen, realistic interaction patterns.

The 8-turn step-wise sub-bench targets the regime where drift appears; the failure-propagation rule correctly reflects the causal structure of multi-turn editing; DRS is grounded in an external finding that its two components detect iterative-replication degradation.

Critical weaknesses

The same closed-model ecosystem (Nano Banana 2 / GPT Image 2) generates the training data and appears as benchmark competitors; the judge (Qwen 3.5 Plus) shares the Qwen ecosystem with the base model. The human-alignment study mitigates the judge concern at rank level, but absolute-score calibration against humans is not reported, and the "best among open-source" claim should be read knowing the closed-source teachers sit above. This is inherent to frontier-distilled dataset work and is handled transparently, but it deserves explicit discussion.

The 8-turn construction quietly reintroduces a rollout boundary. Longer trajectories are built by taking the last panel of one PTC as the source image of the next pack, i.e., a coarse-grained rollout every 4 turns. Table 4 shows no visible drift discontinuity at the Turn 4 to 5 boundary, but this is not analyzed directly; the claim "packed beats rollout" strictly holds within packs. A per-turn DRS zoom on pack boundaries, or a failure analysis there, would close this.

Turn-level instructions are retro-fitted to generated panels. Instructions are derived from the PTC after generation, so they describe what visibly changed rather than what a user would plausibly ask next. This is a subtle distribution shift versus real user intent; a small human or VLM audit of instruction naturalness would strengthen the data-quality story.

Questions for authors

What is the PTC rejection rate at the Gemini filtering stage, and what is the end-to-end cost per accepted trajectory vs. sequential rollout?

Is there any measurable quality discontinuity at pack boundaries (Turn 4 to 5) in 8-turn samples? Would 3×3 packs with the stronger upcoming editors change the layout conclusion?

How often do the retro-fitted turn instructions diverge from the planned PTC prompt, and did you audit instruction naturalness?

Can you report IF/CC without the zeroing rule (or trajectory survival curves) as a robustness check?

What license will govern MTEdit given that targets are closed-model outputs, and will MTEdit-Bench source images be redistributable?

7 comments

Response to Reviewer xSfE

We sincerely thank you for your positive assessment of our work and your constructive feedback, which has helped us improve the paper. We address your concerns point by point below.


Q1:

What is the PTC rejection rate at the Gemini filtering stage?

The Gemini filtering rejection rates are 5.3% for Nano Banana 2 and 1.1% for GPT Image 2.


What is the end-to-end cost per accepted trajectory vs. sequential rollout?

PTC costs $0.442/$0.496 per 8-turn trajectory with Nano Banana 2/GPT Image 2, versus $1.206/$1.256 for sequential rollout, including image generation and Gemini annotation.

Here, we compare 4K PTC canvases with 2K images generated through sequential rollout. A detailed cost breakdown is provided below.

Cost per 8-turn trajectory (USD)Nano Banana 2: PTCNano Banana 2: rolloutGPT Image 2: PTCGPT Image 2: rollout
Image generation$0.302$0.806$0.356$0.856
Gemini annotation$0.140$0.400$0.140$0.400
Total$0.442$1.206$0.496$1.256

This analysis has been added to Appendix A.2.2 of the revised paper.


Q2 and W2:

Is there any measurable quality discontinuity at pack boundaries (Turn 4 to 5) in 8-turn samples?

Yes. However:

  1. It should be assessed alongside the results before truncation.
  2. It is more evident in CC than in DRS.

Before applying the truncation rule, our MTEdit achieves the following results on the step-wise benchmark:

TurnIFCCDRS
19.749.544.23
29.799.434.27
39.499.354.30
49.759.184.33
59.498.554.32
69.408.254.33
78.758.304.35
89.358.074.33

The untruncated results show a clear drop in CC at the pack boundary: CC decreases from 9.18 at Turn 4 to 8.55 at Turn 5, the largest adjacent-turn decline.

DRS does not show an obvious discontinuity because: This is because PTC 1 and PTC 2 correspond to the first and second image-generation passes, respectively, and the change in visual quality between these two passes is relatively small.

This analysis has been added to Appendix A.3.1 of the revised paper.


Would 3×3 packs with the stronger upcoming editors change the layout conclusion?

We believe so. For example, GPT Image 2.6, a leading image editing model, already produces 3×3 PTCs with promising visual quality. However, its current maximum output resolution of 2880×2880 limits each panel to 960×960 pixels.

This makes 3×3 layouts viable for applications with lower resolution requirements. We will continue to assess advances in image generation models, particularly their ability to produce higher-resolution PTCs, and revisit our layout choice as these capabilities improve.

See Appendix A.2.3 and Figure 12 of the revised paper for this discussion and examples of 3×3 PTCs.


Q3 and W3:

How often do the retro-fitted turn instructions diverge from the planned PTC prompt?

Approximately 14% of the rewritten prompts deviate from the planned PTC prompt.


Discussion of instruction naturalness.

We sampled 400 cases from the training data and evaluated instruction naturalness on a scale of 1–5. The evaluation system prompt is provided in Appendix A.8.10 of the revised paper.

We also conducted a human evaluation, with each instruction independently rated by 6–10 annotators.

EvaluatorAverage Naturalness (out of 5)
Human4.63
Qwen 3.5 Plus4.70
Gemini 3.1 Pro4.96
GPT-6.1 Sol4.84

These results indicate that the instructions are generally natural.

This discussion has been added to Appendix A.2.5 of the revised paper.


Q4

Can you report IF/CC without the zeroing rule (or trajectory survival curves) as a robustness check?

Yes. These results have been added to Appendix A.10 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.

Step-wise:

ModelT1T2T3T4T5T6T7T8
Nano Banana Pro9.75 / 9.799.73 / 9.539.73 / 9.439.73 / 9.249.84 / 9.019.74 / 8.749.22 / 8.569.80 / 8.31
Nano Banana 29.65 / 9.589.65 / 9.389.78 / 9.239.75 / 8.919.74 / 8.779.72 / 8.529.27 / 8.109.74 / 8.20
GPT Image 29.02 / 8.728.84 / 8.679.08 / 8.309.03 / 8.049.05 / 7.658.86 / 7.398.21 / 6.797.92 / 6.41
Seedream 5.0 Lite9.73 / 9.609.91 / 9.189.76 / 8.669.74 / 7.969.77 / 7.509.88 / 6.879.56 / 6.629.78 / 6.42
VINCIE1.13 / 4.310.81 / 2.740.97 / 1.810.79 / 1.241.12 / 0.800.76 / 0.690.78 / 0.430.61 / 0.40
OmniGen26.75 / 7.466.35 / 6.576.52 / 5.156.62 / 4.095.92 / 3.395.16 / 2.895.89 / 2.555.23 / 2.40
FLUX.1-kontext-dev8.29 / 8.778.27 / 8.228.04 / 7.817.91 / 7.337.97 / 6.867.88 / 6.357.07 / 5.988.28 / 5.98
FLUX.2-klein-9B9.09 / 9.198.98 / 9.068.93 / 8.599.18 / 8.448.91 / 8.289.15 / 7.918.39 / 7.749.21 / 7.70
Qwen-Image-Edit-25119.30 / 9.439.30 / 9.049.27 / 8.609.58 / 8.469.43 / 8.059.35 / 7.838.45 / 7.419.24 / 7.22
BAGEL8.69 / 9.148.38 / 8.848.09 / 8.557.28 / 8.237.82 / 7.866.49 / 7.256.67 / 6.836.12 / 6.54
Emu3.57.40 / 7.907.63 / 7.267.66 / 7.067.96 / 6.887.71 / 6.577.59 / 6.327.08 / 5.847.77 / 5.77
SenseNova-U19.29 / 9.068.95 / 8.518.90 / 7.938.50 / 6.828.74 / 6.078.89 / 5.558.57 / 5.098.63 / 4.87
Lance8.20 / 8.397.28 / 6.636.77 / 5.496.78 / 4.605.56 / 4.055.05 / 3.573.82 / 3.333.87 / 3.03
Qwen-Image-Edit-Causal8.89 / 9.218.91 / 8.409.10 / 7.769.21 / 7.309.20 / 6.519.11 / 6.088.39 / 5.549.02 / 5.35
Ours9.74 / 9.549.79 / 9.439.49 / 9.359.75 / 9.189.49 / 8.559.40 / 8.258.75 / 8.309.35 / 8.07

Persistent-constraint:

ModelT1T2T3T4
Nano Banana 28.70 / 9.049.22 / 9.029.12 / 8.868.88 / 8.36
Nano Banana Pro9.02 / 9.289.34 / 9.288.68 / 9.069.28 / 8.78
GPT Image 29.36 / 9.369.40 / 9.488.70 / 9.169.24 / 9.06
Seedream 5.0 Lite9.88 / 9.889.34 / 9.668.94 / 9.548.68 / 9.38
VINCIE1.58 / 5.301.34 / 2.960.64 / 1.680.34 / 0.96
BAGEL7.90 / 8.986.88 / 8.626.02 / 8.165.72 / 7.92
Emu3.57.88 / 8.346.98 / 7.925.96 / 7.046.10 / 6.24
SenseNova-U19.64 / 9.468.52 / 8.648.10 / 7.887.08 / 6.74
Lance8.44 / 8.305.80 / 6.605.06 / 5.444.80 / 4.52
Ours9.70 / 9.809.52 / 9.549.26 / 9.128.60 / 8.62

Textual-reference:

ModelT1T2T3T4
Nano Banana 29.36 / 9.929.28 / 9.709.28 / 9.208.78 / 8.56
Nano Banana Pro9.56 / 9.769.30 / 9.568.64 / 9.328.46 / 9.12
GPT Image 29.72 / 9.789.34 / 9.849.06 / 9.468.86 / 9.12
Seedream 5.0 Lite9.80 / 9.828.46 / 9.768.66 / 9.528.90 / 9.56
VINCIE3.00 / 5.862.30 / 4.081.58 / 2.921.32 / 1.90
BAGEL8.30 / 9.365.16 / 9.245.56 / 8.584.60 / 7.98
Emu3.57.36 / 7.807.10 / 6.826.68 / 6.246.10 / 5.12
SenseNova-U19.42 / 9.568.76 / 8.788.64 / 8.628.26 / 8.08
Lance8.10 / 8.506.52 / 7.225.64 / 5.925.26 / 4.66
Ours9.42 / 9.549.28 / 9.488.70 / 9.348.92 / 8.86

Short-instruction:

ModelT1T2T3T4
Nano Banana 29.68 / 9.788.24 / 9.588.42 / 9.027.98 / 8.26
Nano Banana Pro9.74 / 9.968.42 / 9.868.62 / 9.707.54 / 9.22
GPT Image 29.86 / 9.849.30 / 9.808.94 / 9.588.46 / 9.14
Seedream 5.0 Lite9.58 / 9.748.54 / 9.728.18 / 9.508.04 / 9.64
VINCIE2.92 / 6.302.46 / 4.362.10 / 3.341.94 / 2.40
BAGEL8.00 / 9.204.00 / 8.723.36 / 8.502.22 / 8.02
Emu3.57.54 / 8.387.16 / 7.427.02 / 6.465.48 / 5.02
SenseNova-U19.32 / 9.468.54 / 9.128.46 / 8.747.78 / 7.90
Lance8.62 / 8.606.94 / 7.706.28 / 6.564.76 / 5.22
Ours9.62 / 9.889.64 / 9.749.00 / 9.448.10 / 8.66

W1:

Concern About the Judge (Qwen 3.5 Plus) and the Base Model Sharing the Qwen Ecosystem.

We additionally evaluated the models using GPT-6.1 Sol and Gemini 3.1 Pro as independent VLM judges. The results are broadly consistent with those obtained using Qwen 3.5 Plus. 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, including ties.

Step-wise — GPT-6.1 Sol

ModelT1T2T3T4T5T6T7T8
Nano Banana Pro9.17 / 8.579.04 / 8.129.09 / 7.739.09 / 7.549.05 / 7.358.84 / 7.188.81 / 7.058.96 / 6.83
Nano Banana 29.18 / 8.718.41 / 7.728.24 / 7.208.23 / 6.868.07 / 6.517.98 / 6.377.48 / 6.257.51 / 5.98
GPT Image 28.43 / 7.788.32 / 7.078.27 / 6.578.38 / 6.308.41 / 6.108.07 / 5.937.59 / 5.517.31 / 5.09
Seedream 5.0 Lite9.43 / 8.929.45 / 8.019.30 / 7.389.22 / 6.789.01 / 6.308.92 / 6.138.67 / 5.948.84 / 5.89
VINCIE1.69 / 6.270.14 / 0.330.02 / 0.050.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.000.00 / 0.00
OmniGen25.94 / 6.963.72 / 3.812.53 / 2.351.81 / 1.590.85 / 0.860.54 / 0.490.36 / 0.280.11 / 0.11
FLUX.1-kontext-dev7.39 / 7.956.06 / 6.315.35 / 5.094.33 / 4.273.81 / 3.553.12 / 2.872.56 / 2.342.34 / 1.94
FLUX.2-klein-9B8.45 / 8.627.52 / 7.497.12 / 6.827.41 / 6.626.73 / 6.036.44 / 5.705.74 / 5.135.59 / 4.89
Qwen-Image-Edit-25118.02 / 7.988.28 / 7.328.02 / 6.747.89 / 6.397.71 / 6.147.17 / 5.896.99 / 5.466.96 / 5.31
BAGEL8.05 / 8.757.36 / 7.826.47 / 7.065.09 / 5.904.89 / 4.803.87 / 4.373.08 / 3.142.06 / 2.18
Emu3.57.08 / 7.296.71 / 6.625.98 / 5.875.46 / 5.254.70 / 4.404.28 / 3.933.51 / 3.113.17 / 2.74
SenseNova-U18.63 / 8.387.59 / 7.567.00 / 6.596.59 / 5.715.90 / 4.795.78 / 4.504.94 / 4.134.41 / 3.71
Lance6.64 / 6.883.32 / 3.600.79 / 0.840.51 / 0.520.13 / 0.100.05 / 0.060.00 / 0.000.00 / 0.00
Qwen-Image-Edit-Causal7.47 / 7.567.53 / 6.637.12 / 5.966.91 / 5.286.52 / 4.886.09 / 4.645.52 / 4.274.79 / 3.60
Ours8.03 / 7.758.94 / 7.617.98 / 7.488.28 / 7.158.14 / 6.587.87 / 6.546.67 / 6.537.21 / 6.03

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.