acceptodds
Under review as a conference paper at ICLR 2027

Leveraging Diffusion MLLMs for Image Editing via Dynamic Masked Generation

Abstract

Diffusion Multimodal Large Language Models (dMLLMs) have recently emerged as a powerful paradigm for multimodal generation. Their localized and progressive masked generation property naturally provides a promising foundation for training-free text-based image editing, \ie, explicitly preserving editing-irrelevant regions from the source image, while treating regions to be edited as masked tokens for regeneration guided by the target prompt. Existing dMLLM-based editing method, however, relies on fixed edit masks, static source guidance, and pre-defined generation trajectories, resulting in inflexible edit localization and degraded source consistency. In this paper, we propose DynEdit, a general training-free framework that better exploits the intrinsic generation properties of dMLLMs for flexible and faithful image editing. Specifically, we introduce a parallel dual-branch framework that enables dynamic masked generation with three key designs: 1) Dynamic edit localization to automatically identify editable tokens throughout generation. 2) Adaptive source guidance to adjust source preservation according to specific editing areas and generation steps. 3) Edit-aware token decoding to dynamically refine the decoding schedule for more reliable generation. Extensive experiments on diverse image benchmarks have demonstrated that DynEdit consistently improves editing quality, editing efficiency, and source preservation under both mask-free and external mask-grounded settings, exhibiting strong generalization in various editing scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Multimodal Large Language Models, Image Editing, Masked Generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_XJRxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DynEdit, a training-free image editing framework for diffusion multimodal language models. A source branch extracts reconstruction residuals using the masking pattern of the evolving target branch. Prediction disagreement identifies editable tokens, spatial and temporal schedules adjust source guidance, and source-consistent tokens are decoded before editable tokens. Experiments evaluate mask-free and mask-grounded editing on PIE-Bench, PIE-Bench++, and GIDE-Bench, with additional backbone and component ablations.

Critical strengths
  1. A coherent editing mechanism with practical value. Synchronizing source residual extraction with target generation allows localization, preservation, and decoding order to evolve together. In the mask-free setting, this removes the need for an external grounding model and provides a useful capability beyond editing within a fixed mask.

  2. Meaningful evidence of improved editing–preservation trade-offs. The evaluation includes same-backbone discrete baselines, identical-mask comparisons, generation budgets, and detailed appendix ablations. On PIE-Bench, reported global structure distance improves from 3.84 to 2.81 and latency decreases from 31.1 to 21.3 seconds relative to GIDE. These benefits support the method’s practical contribution, although background preservation is weaker in this setting.

Critical weaknesses
  1. The component gains are not fully separated from changes in editing strength. Removing several components reduces alignment while improving preservation, and the schedule ablations show that adjusting guidance can recover similar operating points. Appendix D provides useful evidence for spatial adaptation, but a comparison at matched edit-region alignment would better establish whether the complete mechanism improves the trade-off over simpler alternatives. Eq. 11 also couples residual strength to Gumbel noise amplitude, complicating attribution specifically to source guidance. This affects how much independent algorithmic value can be assigned to the proposed components.

  2. The robustness of automatic localization is not sufficiently characterized. Argmax disagreement is discontinuous, and the source and target branches also differ in their visible image context. Previously decoded tokens appear to remain fixed, so an early incorrect decision may prevent a required edit. The process visualizations demonstrate successful examples but do not establish how often this failure occurs. Since reliable automatic localization is a central benefit over fixed-mask editing, this uncertainty matters to the method’s applicability.

Questions for authors
  1. At comparable edit-region alignment, does DynEdit retain better preservation than GIDE or the principal ablated variants? A limited comparison using existing PIE-Bench results or a small guidance sweep would address this question.
  2. Do the adaptive-guidance gains persist if Gumbel noise amplitude is held fixed across the compared variants?
  3. Can previously decoded tokens be reopened? If not, what evidence characterizes failures caused by early localization or fixation errors?
7 comments

We thank the reviewer for recognizing the practical contributions of DynEdit. We address these concerns with additional experiments and failure-case analyses below.

[W1 & Q1]. Component contributions and matched-alignment comparisons

We thank the reviewer for this important suggestion. To better distinguish improvements in source preservation from changes in editing strength, we conduct additional controlled comparisons at matched edit-region alignment (CLIP-E), covering both individual module ablations and the comparison with GIDE.

1. Component contributions at matched editing strength.

We adjust the initial guidance strength of each ablated variant to achieve comparable CLIP-E scores, allowing us to directly compare source preservation at similar editing strength.

Table 1. CLIP-E-matched module ablation on PIE-Bench (revised Sec. 4.5, Table 4). Guidance is adjusted for each ablated variant.

MethodGuidanceCLIP-G ↑CLIP-E ↑BG PSNR ↑BG LPIPS ↓BG MSE ↓BG SSIM ↑Struct. Dist. ↓
Ours0.700025.2722.2122.529.679.4577.342.81
w/o M10.500025.0922.2419.0715.0321.8372.015.70
w/o M20.327025.1322.2321.0612.2618.2275.884.38
w/o M30.544225.0722.1920.9011.9513.4974.224.08
w/o M2 time adaptation0.423525.2222.2322.1710.8111.5077.033.36
w/o M2 spatial adaptation0.697725.2022.2022.0510.2511.0677.103.16

As shown in Table 1 (also revised Sec. 4.5, Table 4), all variants achieve nearly identical CLIP-E scores. DynEdit also obtains the highest CLIP-G and the best source preservation across all five metrics.

In particular, removing M1, M2, or M3 increases the global structure distance despite maintaining comparable edit-region alignment. Removing temporal or spatial adaptation similarly degrades source preservation.

These matched-alignment results are included in revised Sec. 4.5, Table 4 (page 10). These results show that the observed improvements are not explained solely by differences in CLIP-E-measured editing strength. Even when the ablated variants recover comparable target alignment through guidance adjustment, they consistently sacrifice more source information than the complete method, which also retains the highest global target alignment. This provides direct evidence for the complementary contributions of the proposed modules within our coordinated framework.

2. Comparison with GIDE at matched editing strength.

We further adjust DynEdit's initial guidance strength to match GIDE's edit-region alignment, enabling a more controlled comparison of their editing–preservation trade-offs.

Table 2. Comparison with GIDE at matched edit-region alignment on PIE-Bench.

MethodCLIP-G ↑CLIP-E ↑BG PSNR ↑BG LPIPS ↓BG MSE ↓BG SSIM ↑Dist. ↓Global PSNR ↑Global LPIPS ↓Global MSE ↓Global SSIM ↑
GIDE24.7022.1724.377.868.3679.243.8417.8919.8530.4861.45
Ours (0.7404)25.1822.1822.909.288.7477.782.6318.5117.3617.8362.45

As shown in Table 2, at nearly identical CLIP-E scores, DynEdit achieves higher global target alignment and consistently better global preservation across all five metrics. In particular, the structure distance decreases from 3.84 to 2.63, representing a 31.5% reduction.

GIDE retains an advantage in background preservation, which is expected because it directly copies source tokens outside the externally estimated editing mask. However, DynEdit achieves better overall structural consistency, while the background preservation gap is relatively modest at this matched operating point.

Importantly, GIDE relies on external SAM-generated masks, whereas DynEdit performs mask-free editing without external localization. DynEdit also reduces inference time from 31.1s to 21.3s, corresponding to approximately 1.46× faster inference.

Furthermore, our experiments under identical external masks (Sec. 4.3) demonstrate that DynEdit consistently outperforms GIDE across the reported metrics from 16 steps onward, further validating its effectiveness when external grounding is available.

Overall, these controlled comparisons indicate that DynEdit's improvements are not simply attributable to choosing a different CLIP-E operating point. The proposed framework provides stronger global structural preservation at comparable edit-region alignment, while offering greater flexibility and efficiency across different editing settings.

[W1 & Q2]. Adaptive guidance with fixed Gumbel amplitude

We thank the reviewer for this suggestion. As clarified in revised Appendix B (lines 847–855), the complementary weighting between source-residual guidance and Gumbel noise in Eq. (11) follows the stochastic fusion formulation adopted in GIDE [GIDE Sec. 3.1 and Algorithm 1]. DICE similarly adopts complementary coefficients to simplify hyperparameter selection [DICE Appendix E.2].

This is an empirical parameterization that jointly controls source-residual correction and sampling stochasticity through a single coefficient, rather than a theoretically required coupling. DynEdit follows this formulation and focuses on making the coefficient spatially and temporally adaptive.

To examine whether the effectiveness of our adaptive guidance depends on this complementary weighting, we conduct additional experiments with a fixed Gumbel noise amplitude of 1, while retaining the adaptive source-residual guidance.

We report both the influence of initial guidance strength and the ablation of adaptive guidance components under this alternative formulation.

1. Influence of initial guidance strength.

Table 3. Effect of initial guidance strength with complementary and fixed Gumbel noise on PIE-Bench. All experiments use 64 steps and cosine guidance decay.

Initial GuidanceOriginalFixed Gumbel
CLIP-G ↑CLIP-E ↑BG PSNR ↑Dist. ↓CLIP-G ↑CLIP-E ↑BG PSNR ↑Dist. ↓
0.625.7722.7521.663.4425.9122.8621.003.84
0.725.2722.2122.522.8125.5122.5021.893.13
0.824.9221.8523.282.4525.0522.0322.952.58

As shown in Table 3, fixing the Gumbel amplitude generally increases target alignment while reducing source preservation, consistent with the increased stochastic perturbation during generation. Nevertheless, both formulations exhibit the same overall trend: stronger initial source guidance improves preservation at the expense of editing strength.

This indicates that the complementary noise weighting influences the editing–preservation operating point, but does not fundamentally change the role of source guidance.

2. Ablation of adaptive guidance components.

Table 4. Adaptive guidance ablation with complementary and fixed Gumbel noise on PIE-Bench. All experiments use an initial guidance strength of 0.7 and cosine decay.

MethodOriginal CLIP-G ↑Original CLIP-E ↑Original BG PSNR ↑Original Dist. ↓Fixed Gumbel CLIP-G ↑Fixed Gumbel CLIP-E ↑Fixed Gumbel BG PSNR ↑Fixed Gumbel Dist. ↓
Full25.2722.2122.522.8125.5122.5021.893.13
w/o M223.0220.1326.931.0023.1820.2726.761.20
w/o spatial25.2022.1222.293.1025.4322.3921.433.44
w/o time23.9520.9625.911.3924.0421.1025.821.51

As shown in Table 4, the effects of adaptive guidance remain consistent when the Gumbel amplitude is fixed.

Removing M2 or its temporal adaptation substantially reduces target alignment while favoring source reconstruction. Meanwhile, removing spatial adaptation degrades both target alignment and source preservation, confirming that spatially differentiated guidance contributes to the overall editing–preservation balance.

Importantly, these trends closely match those observed under the original complementary weighting. Therefore, the qualitative contributions of adaptive guidance persist even without dynamically scaling the Gumbel noise amplitude.

Overall, the fixed-Gumbel experiments confirm that the complementary weighting primarily serves as a practical parameterization of the residual–noise balance, while the contributions of spatially and temporally adaptive source guidance remain valid under an alternative noise configuration. We retain the original formulation for consistency with prior work, its rationale is now explained in revised Appendix B (lines 847–855).

[W2 & Q3]. Robustness of localization, irreversible token fixation, and failure cases

We thank the reviewer for raising these important concerns. We first clarify the motivation behind our disagreement-based localization. Argmax disagreement is a deliberate design choice motivated by the discrete generation mechanism of dMLLMs, where token predictions are ultimately converted into discrete token decisions. Rather than comparing the entire predicted distributions, we directly identify positions where the two branches favor different tokens, providing a simple token-level signal of source–target discrepancy.

Regarding the different visible contexts, the two branches share the same masking pattern but retain their respective source and evolving target contexts. Since localization is performed only on the remaining masked positions, comparing their predictions under these distinct contexts allows us to capture discrepancies arising from both the editing prompts and the evolving visual content. We consider this a natural way to exploit the progressive masked-generation mechanism of dMLLMs, rather than requiring identical visible tokens across branches. Nevertheless, argmax disagreement is not guaranteed to accurately capture semantic differences, particularly under uncertain predictions, and alternative distribution-aware signals may offer further improvements. We therefore discuss how our framework mitigates localization errors and acknowledge its remaining limitations below.

1. Irreversible token fixation and robustness of progressive localization.

As explicitly stated in revised Appendix J (lines 1877–1889), previously decoded tokens cannot be reopened or remasked in our current implementation, following the standard progressive masked-generation strategy of dMLLMs. Therefore, incorrect early decoding decisions cannot be directly corrected and may affect subsequent predictions through the fixed context.

Our framework is designed to mitigate, rather than eliminate, this limitation by exploiting the progressive and localized generation process of dMLLMs.

Specifically, during the initial steps, strong source guidance encourages source-consistent predictions, while the decoding schedule fixes only a small number of tokens. M3 further prioritizes tokens classified as non-editable, reducing the likelihood of prematurely fixing editing-relevant content.

As generation progresses, source-consistent background and structural tokens are preferentially decoded, providing increasingly informative visual context for subsequent predictions. Meanwhile, M2 adapts source guidance according to the evolving localization results, balancing source preservation and target-oriented generation.

Importantly, the three modules form a mutually supportive progressive process. M1 provides the dynamic localization information required by M2 and M3, while M2 influences subsequent source–target predictions through adaptive guidance, and M3 establishes visual context through edit-aware decoding. Thus, localization guides generation, while the evolving generation context in turn supports subsequent localization.

We acknowledge that individual localization decisions are not guaranteed to be correct. For example, as source guidance weakens at later stages, some background tokens may be incorrectly classified as editable. However, the previously established source-consistent context, together with the remaining source guidance, can help constrain their subsequent predictions and reduce unintended modifications.

This coordinated process helps mitigate uncertain localization, but does not guarantee recovery from incorrect early decisions, particularly when such decisions result in prematurely fixed tokens.

Reviewer_6EAiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents DynEdit, a training-free text-guided image editing method for diffusion multimodal large language models (dMLLMs). Its core idea is to replace fixed masks and predefined generation trajectories with a dynamic editing mechanism that is continuously updated during generation, enabling the desired modification while preserving irrelevant regions. The main contributions are: (1) parallel source and target branches that synchronously extract source-image information and generate target content at each generation step; (2) dynamic edit localization based on prediction discrepancies between the two branches, allowing mask-free editing; (3) adaptive source guidance conditioned on the editing region and generation stage; (4) an edit-aware decoding strategy that prioritizes background and non-editable regions before progressively generating the target content. Experiments show that DynEdit generally outperforms other discrete image editing methods. Ablation studies and visualizations of the generation process further demonstrate the effectiveness of DynEdit.

Critical strengths

1.DynEdit transforms fixed masks and predefined generation trajectories into a progressively updated, closed-loop process, which is both valuable and meaningful. At each generation step, it compares the predictions of the source and target branches and dynamically adjusts the editing region, guidance strength, and decoding order. By unifying edit localization, source information preservation, and token decoding within a dual-branch framework, the dynamic editing mask controls both the source guidance strength across different regions and the token decoding priority. This allows non-editable regions to stabilize first while editable regions are progressively refined, effectively leveraging the progressive generation characteristics of dMLLMs.

2.DynEdit adopts a fully training-free design, resulting in relatively low deployment and transfer costs. It requires neither fine-tuning of the dMLLMs nor additional auxiliary inputs, but instead performs editing directly with the pretrained model through inference-time control, giving the method practical value.

Critical weaknesses

1.The effectiveness of dynamic edit localization is not directly validated. DynEdit identifies editable tokens solely based on whether the argmax predictions of the source and target branches disagree. However, such disagreement may also arise from model uncertainty or other factors, and the paper does not experimentally verify whether the resulting dynamic editing masks accurately correspond to the target editing regions.

2.The three proposed modules are highly coupled, which weakens the current ablation studies. The mask produced by dynamic edit localization controls both adaptive source guidance and edit-aware decoding. Consequently, removing the localization module also disrupts the other two modules, making it difficult to isolate their individual contributions and reducing the persuasiveness of the ablation results.

Questions for authors

1.There appears to be a formatting error in the third row of Figure 3(b). Please check and revise it carefully.

2.How do M2 and M3 operate when the dynamic edit localization module (M1) is removed? Is the dynamic editing mask mt set to all zeros, replaced with a fixed mask, or determined using another rule? To what extent are the spatial adaptivity of M2 and the edit-aware decoding order of M3 degraded in this setting?

3.In Equation (11), αt\alpha_{t} jointly controls the source residual guidance and Gumbel noise. As the source residual guidance becomes stronger, the Gumbel noise is simultaneously reduced. What is the rationale behind this coupled design?

4.Prediction disagreement may arise from model uncertainty or other factors. Therefore, using hard argmax disagreement as the localization signal may be overly brittle. It ignores distributional distances and prediction confidence, potentially misidentifying editable tokens because of minor probability fluctuations. Please provide a detailed explanation of how DynEdit addresses this issue.

7 comments

We thank the reviewer for recognizing the value of coordinating localization, source preservation, and token decoding within a training-free framework. We address the concerns below with additional experiments and clarifications on the robustness of dynamic localization, module dependencies, and the complementary residual–noise weighting.

[W1 & Q4]. Effectiveness and robustness of dynamic localization

We thank the reviewer for this important concern. We agree that hard argmax disagreement may be affected by prediction uncertainty and does not guarantee accurate localization at every token. To address this concern, we provide additional quantitative evidence and clarify how our framework exploits progressive masked generation to mitigate the impact of potential localization errors.

1. Effectiveness of dynamic localization.

To directly evaluate the proposed localization, we measure its IoU with PIE-Bench ground-truth (GT) editing masks across all 64 generation steps. We further introduce a random baseline with the same number of editable tokens to distinguish meaningful spatial localization from the natural progression of the decoding schedule.

Specifically, we evaluate 556 PIE-Bench images with nonempty background regions. At each step, we restrict the GT mask to the remaining masked positions. Given that DynEdit identifies kk editable tokens, the random baseline uniformly selects the same number kk from these positions, ensuring identical candidate regions and editable-token counts.

Revised Appendix I, Fig. 17 reports the dynamic-mask IoU curves against the same-count random baseline.

As shown in revised Appendix I, Fig. 17, the IoU of DynEdit gradually increases throughout generation, consistent with the localized and progressive generation mechanism of dMLLMs. More importantly, DynEdit consistently outperforms random selection during intermediate steps. For example, at step 44, DynEdit achieves 25.91% IoU versus 17.30% for random selection (+8.61 percentage points), with similar improvements at steps 32 (+5.12) and 48 (+7.83). These results demonstrate that prediction disagreement provides meaningful spatial localization beyond the step-dependent decoding schedule and editable-token count.

Meanwhile, we emphasize that GT-mask IoU is a complementary diagnostic rather than a definitive measure of editing effectiveness. GT masks often cover entire objects or coarse semantic regions, whereas our method aims to identify fine-grained tokens requiring modification. For example, in the cake-shape editing case (revised Appendix F, Fig. 13), the GT mask covers the entire cake, while DynEdit identifies more localized regions corresponding to the evolving target shape. Such localization may yield a lower IoU despite being more suitable for editing.

More importantly, the final editing results provide complementary evidence. DynEdit achieves competitive editing performance against GIDE without requiring its external SAM masks. Under the identical-mask setting (Sec. 4.3), DynEdit further distinguishes editable tokens within the provided candidate regions (GT masks on PIE-Bench and SAM-grounded bounding boxes on PIE-Bench++) and outperforms GIDE across the four reported metrics from 16 steps onward. These results demonstrate the practical effectiveness of dynamic token-level localization across different editing settings.

2. Robustness to prediction uncertainty and localization errors.

Our localization is designed around the progressive masked-generation process of dMLLMs, rather than relying on a single prediction to determine a fixed editing region. As illustrated in revised Appendix F, Fig. 12, the localization, guidance, and decoding mechanisms jointly support the progressive identification and generation of editable regions.

During the initial steps, strong source guidance encourages source-consistent predictions, causing most remaining masked tokens to be classified as non-editable. Meanwhile, the decoding schedule initially fixes only a small number of tokens. Edit-aware decoding explicitly prioritizes tokens classified as non-editable, reducing the likelihood of prematurely fixing editing-relevant content under strong source guidance.

As generation progresses, M3 prioritizes source-consistent background and structural tokens, gradually establishing visual context for subsequent predictions. Meanwhile, M2 adapts source guidance according to the evolving localization results, balancing source preservation and target-oriented generation. The updated generation context and source–target predictions are then used by M1 to refine the localization at subsequent steps.

Importantly, the three modules form a mutually supportive progressive process. While M1 provides the localization information required by M2 and M3, these modules also influence subsequent localization: M2 adjusts the source–target predictions through adaptive guidance, while M3 establishes source-consistent context through editing-aware decoding. Thus, dynamic localization guides generation, and the evolving generation context in turn supports subsequent localization, helping mitigate the impact of uncertain token-level decisions.

The robustness mechanism and its limitations are now discussed in revised Appendix J (lines 1864–1889), with Fig. 18. We acknowledge that hard argmax disagreement can still misclassify individual tokens, particularly under uncertain predictions or weaker source guidance at later stages. However, by this stage, a substantial portion of source-consistent content has typically been decoded, providing contextual constraints for the remaining predictions. Together with the remaining source guidance, this helps reduce unintended modifications even when some background tokens are classified as editable.

We also acknowledge a limitation inherited from the standard progressive masked-generation strategy: previously decoded tokens cannot be reopened for correction. Consequently, an inaccurate localization decision may contribute to premature fixation; those fixed tokens can then affect later predictions and localization through their context. Our framework mitigates this risk through adaptive guidance and edit-aware decoding but does not eliminate it entirely. The limitation and corresponding failure-case visualizations are discussed in revised Appendix J, Fig. 18, with additional backbone-level limitations in revised Appendix L (page 38).

Overall, DynEdit exploits the localized and progressive nature of dMLLM generation to coordinate dynamic localization, source preservation, and token decoding. Rather than requiring perfectly accurate localization at every step, the framework progressively adapts its editing decisions as more visual context becomes available, enabling effective mask-free editing while retaining the limitations of the underlying decoding paradigm.

[W2 & Q2]. Module coupling and the operation of M2/M3 without M1

We thank the reviewer for raising this important question. We acknowledge that the three modules are inherently coupled by design, as M1 provides the dynamic localization information required by M2 and M3. However, this coupling is an intentional aspect of our framework: M1 guides subsequent source adaptation and decoding, while M2 and M3 influence the evolving generation process and thereby support subsequent localization.

1. How M2 and M3 operate without M1.

When M1 is removed, no dynamic editable/non-editable labels are available for the remaining masked tokens. Consequently, M2 retains only global time-adaptive source guidance, applying the same temporally decaying guidance strength to all masked tokens, while M3 reduces to standard confidence-based decoding, without editable/non-editable priority.

Under the original configuration with relatively strong initial source guidance (0.7) and a slowly decaying cosine schedule, this can cause editing-relevant tokens to be decoded prematurely, leading to weaker target alignment and a tendency toward source reconstruction.

We agree that removing M1 inevitably disables the localization-dependent operations of M2 and M3, making it difficult to fully isolate its contribution through conventional ablation. However, this dependency reflects the intended architecture of our framework, where localization serves as the foundation for subsequent guidance and decoding decisions.

2. Controlled ablation at matched editing strength.

To better assess the contributions of individual modules despite this dependency, we conduct additional controlled ablations at matched edit-region alignment. Importantly, independently removing M2 or M3 while retaining M1 allows us to evaluate their respective contributions within the localization-guided framework.

Specifically, we adjust the initial guidance strength of each ablated variant to match the CLIP-E score of the complete DynEdit, enabling a more direct comparison of source preservation under comparable editing strength.

Table 1. CLIP-E-matched module ablation on PIE-Bench (revised Sec. 4.5, Table 4). Guidance is adjusted for each ablated variant; BG denotes background preservation. LPIPS, SSIM, and structure distance are scaled by 10210^2, and MSE by 10310^3.

MethodGuidanceCLIP-G ↑CLIP-E ↑BG PSNR ↑BG LPIPS ↓BG MSE ↓BG SSIM ↑Struct. Dist. ↓
Ours0.700025.2722.2122.529.679.4577.342.81
w/o M10.500025.0922.2419.0715.0321.8372.015.70
w/o M20.327025.1322.2321.0612.2618.2275.884.38
w/o M30.544225.0722.1920.9011.9513.4974.224.08
w/o M2 time adaptation0.423525.2222.2322.1710.8111.5077.033.36
w/o M2 spatial adaptation0.697725.2022.2022.0510.2511.0677.103.16

As shown in Table 1 (also revised Sec. 4.5, Table 4), all variants achieve nearly identical CLIP-E scores (22.19–22.24). DynEdit also obtains the highest CLIP-G (25.27 versus 25.07–25.22 for ablated variants) and the best source preservation across all five metrics.

In particular, removing M1, M2, or M3 increases the global structure distance from 2.81 to 5.70, 4.38, and 4.08, respectively, despite maintaining comparable edit-region alignment. Removing temporal or spatial adaptation similarly degrades source preservation, further confirming the complementary contributions of these two designs.

These matched-alignment results are included in revised Sec. 4.5, Table 4 (page 10). These results show that the observed benefits are not explained solely by differences in CLIP-E-measured editing strength. Even when the ablated variants recover comparable target alignment through guidance adjustment, they consistently sacrifice more source information than the complete framework, which also retains the highest global target alignment.

We emphasize that the modules are functionally interdependent rather than independently operating components. M1 provides the localization information needed by M2 and M3, while their guidance and decoding decisions influence the evolving context used for subsequent localization. The controlled ablations demonstrate the effectiveness of these complementary components within this coordinated framework, without assuming that their contributions can be completely disentangled.

The matched-alignment results are now included in revised Sec. 4.5, Table 4; the functional dependencies are also clarified in revised Sec. 4.5 (lines 465–479).

Reviewer_KSHDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DynEdit, a training-free method for text-based image editing with diffusion multimodal LLMs (dMLLMs) such as Lumina-DiMOO. At each step, it runs a source branch and a target branch in parallel and marks a token as editable when the two branches predict different tokens. It then weakens the source guidance over time and on editable tokens, and decodes non-editable tokens first. Unlike the prior method GIDE, it can edit without an external mask and needs no post-refinement step. On PIE-Bench, PIE-Bench++ and GIDE-Bench, it beats three discrete editing methods re-implemented on the same backbone, outperforms GIDE on all four reported metrics from 16 steps onward when both use the same masks, and also improves over GIDE on GIDE-Bench.

Critical strengths
  1. The comparison with the same masks is convincing. When DynEdit and GIDE start from the same external masks, DynEdit is better on all four metrics on both benchmarks from 16 steps onward, and with 16 steps it already beats GIDE at 64 steps. Because the mask source is held fixed, this is the strongest evidence in the paper.
  2. The method is simple and practical. It needs no training, no segmentation model and no post-refinement, works on two backbones, and is faster than GIDE (21.3s vs. 31.1s). Running the source branch with the same masking pattern as the target is a reasonable fix for GIDE's precomputed inversion.
Critical weaknesses
  1. The core mechanism is not well supported. The paper says the disagreement mask finds the regions to edit, but never checks it against the ground-truth masks in PIE-Bench. Appendix F says almost all tokens are marked non-editable early and editable late, so the mask may mostly follow the step count. The ablation does not separate the modules either: removing any one of them gives nearly the same result, close to plain reconstruction (CLIP-G 22.8–23.0, Table 3c). Spatially adaptive guidance barely matters (25.20 vs. 25.27). What clearly helps is decaying the guidance over time, which is common in diffusion editing. The disagreement-mask idea is also close to DiffEdit, which is not cited.
  2. Key comparisons are missing. Lumina-DiMOO can already edit images by itself, but the paper does not compare against it, so it is unclear why one would use a training-free method on this backbone. In the mask-free setting, the results against GIDE are a trade-off rather than a clear win: GIDE preserves the background better on PIE-Bench and follows the text better on PIE-Bench++. Changing λ0\lambda_0 from 0.6 to 0.8 moves CLIP-G by 0.85, more than the 0.57 gap to GIDE, so comparing single operating points is not enough.
Questions for authors
  1. Sec. 4.5 says that without edit-aware decoding, editable tokens are reconstructed too early. But Appendix F says almost all tokens are non-editable in early steps. Why does this module matter so much?
  2. How well does the dynamic edit mask match the PIE-Bench ground-truth masks across steps (e.g., IoU)?
  3. How does DynEdit compare with Lumina-DiMOO's own editing mode on these benchmarks?
  4. What is the "minimum guidance" mentioned in Appendix F? And why is the Gumbel noise scaled by (1−αt)(1-\alpha_t) in Eq. 11?
  5. How were the re-implemented baselines tuned? DICE gets 9.01 background PSNR, far below the 27.53 reconstruction level.
12 comments

We thank the reviewer for the detailed and constructive feedback, and for recognizing the practical value of DynEdit, particularly the convincing comparisons under identical external masks. We address the concerns and questions below, with additional experiments and clarifications on dynamic localization, component contributions, and the editing–preservation trade-off.

[W1 & Q2]. Effectiveness of dynamic localization and its relationship to ground-truth masks

We thank the reviewer for this suggestion. To directly evaluate the proposed dynamic localization, we measure its IoU with PIE-Bench ground-truth (GT) editing masks across all 64 denoising steps. We further introduce a random baseline with the same number of editable tokens to investigate whether the observed localization is merely determined by the denoising step and the number of selected tokens. Specifically, we evaluate 556 PIE-Bench images with nonempty background regions. At each step, we restrict the GT mask to the remaining masked positions. Given that DynEdit identifies kk editable tokens, the random baseline uniformly selects the same number kk from these remaining positions. Both methods are thus evaluated under identical candidate regions and editable-token counts.

Revised Appendix I, Fig. 17 reports the dynamic-mask IoU curves against the same-count random baseline.

As shown in revised Appendix I, Fig. 17, the IoU of DynEdit gradually increases throughout generation. We acknowledge that this behavior is related to the progressive generation process of dMLLMs, but it does not imply that localization is determined solely by the step count. Our framework intentionally prioritizes decoding source-consistent regions to establish reliable visual context, allowing editable regions to be progressively identified and refined. Crucially, the random baseline follows exactly the same step-dependent candidate set and editable-token count, yet consistently achieves lower IoU during intermediate steps. For example, at step 44, DynEdit achieves 25.91% IoU, compared with 17.30% for random selection, an improvement of 8.61 percentage points. Similar advantages are observed at steps 32 (+5.12 points) and 48 (+7.83 points). The curves converge only near the final steps, when very few masked tokens remain. These results demonstrate that prediction disagreement provides meaningful spatial localization beyond the progression of the decoding schedule or the increasing proportion of editable tokens.

Meanwhile, we emphasize that IoU with GT masks is not the sole criterion for evaluating dynamic localization, for two reasons.

  1. GT masks do not necessarily represent the optimal fine-grained editing regions. They often cover entire objects or coarse semantic regions, whereas our dynamic localization aims to identify individual tokens requiring modification. For example, in the cake-shape editing case (revised Appendix F, Fig. 13), the GT mask covers the entire cake, while DynEdit can identify finer-grained regions corresponding to the evolving target shape. Such localization may yield a lower IoU despite being more suitable for editing.
  2. Localization quality should ultimately be assessed together with editing effectiveness. A higher mask IoU does not necessarily guarantee better target alignment or source preservation. In our experiments, GIDE relies on external SAM masks, while DynEdit achieves competitive or better editing performance without such masks. More importantly, under the identical-mask setting (Q2, Sec. 4.3), DynEdit further identifies editable tokens within the externally provided candidate region (GT masks on PIE-Bench and SAM-grounded bounding-box masks on PIE-Bench++) and outperforms GIDE across the four reported metrics from 16 steps onward.

The quantitative evaluation and its interpretation are now included in revised Appendix I, Fig. 17 (pages 34–35). Overall, our dynamic localization is designed to exploit the localized and progressive generation mechanism of dMLLMs. The additional IoU analysis confirms that its localization is not merely a consequence of the denoising step, while the editing results across different settings further demonstrate its practical effectiveness.

[W1]. Effectiveness of the proposed modules

We thank the reviewer for this important observation. We agree that the original ablation results, where removing different modules leads to similarly reduced target alignment, may not sufficiently distinguish their contributions. To address this concern, we conduct two additional analyses: (1) controlled comparisons at matched editing strength, and (2) extended ablations across different guidance schedules and editing settings.

1. Controlled ablation at matched editing strength.

To investigate whether the improvements merely reflect different editing–preservation operating points, we adjust the initial guidance strength of each ablated variant to match the edit-region alignment (CLIP-E) of the complete DynEdit, enabling a direct comparison of source preservation at comparable editing strength.

Table 1. CLIP-E-matched module ablation on PIE-Bench (revised Sec. 4.5, Table 4). Guidance is adjusted for each ablated variant; BG denotes background preservation. LPIPS, SSIM, and structure distance are scaled by 10210^2, and MSE by 10310^3.

MethodGuidanceCLIP-G ↑CLIP-E ↑BG PSNR ↑BG LPIPS ↓BG MSE ↓BG SSIM ↑Struct. Dist. ↓
Ours0.700025.2722.2122.529.679.4577.342.81
w/o M10.500025.0922.2419.0715.0321.8372.015.70
w/o M20.327025.1322.2321.0612.2618.2275.884.38
w/o M30.544225.0722.1920.9011.9513.4974.224.08
w/o M2 time adaptation0.423525.2222.2322.1710.8111.5077.033.36
w/o M2 spatial adaptation0.697725.2022.2022.0510.2511.0677.103.16

As shown in Table 1 (also revised Sec. 4.5, Table 4), all variants achieve nearly identical CLIP-E scores (22.19–22.24). DynEdit also obtains the highest CLIP-G (25.27 versus 25.07–25.22 for ablated variants) and the best source preservation across all five metrics, including background PSNR, LPIPS, MSE, SSIM, and global structure distance.

In particular, removing M1, M2, and M3 increases the structure distance from 2.81 to 5.70, 4.38, and 4.08, respectively, despite maintaining comparable edit-region alignment. Similarly, removing temporal or spatial adaptation increases the structure distance to 3.36 and 3.16, while also degrading the background preservation metrics.

These matched-alignment results are included in revised Sec. 4.5, Table 4 (page 10). These results show that the observed benefits are not explained solely by changes in CLIP-E-measured editing strength. Even when the ablated variants recover comparable target alignment through guidance adjustment, they consistently sacrifice more source information than the complete method, which also retains the highest global target alignment.

2. Extended ablations across different schedules and settings.

To further explain the similar performance degradation observed in the original ablation study, we extend the experiments to three configurations: mask-free editing with linear or cosine guidance decay, and GT-mask-grounded editing with exponential decay.

Table 2. Extended ablation study across different guidance schedules and editing settings on PIE-Bench. We evaluate the contributions of each module under mask-free and GT-mask-grounded configurations.

MethodMask-free · 0.7 + linearMask-free · 0.7 + cosineGT mask · 0.7 + exp
C-G ↑C-E ↑P-bg ↑Dist. ↓C-G ↑C-E ↑P-bg ↑Dist. ↓C-G ↑C-E ↑P-bg ↑Dist. ↓
Ours26.1623.0720.044.6525.2722.2122.522.8125.9723.6926.215.30
w/o M124.9522.0420.254.1422.8320.0727.120.9024.8822.5626.163.30
w/o M223.0220.1226.990.9823.0220.1326.931.0022.8420.1027.480.90
w/o M325.2122.2120.663.7622.8320.1427.110.9324.9022.6026.223.06
w/o time24.2221.3125.321.6923.9520.9625.911.3924.7922.3626.812.50
w/o spatial26.1422.9019.555.1325.2022.1222.293.1025.8423.5226.145.62

Specifically, removing M1 or M3 consistently weakens target alignment, but the degradation is less pronounced under the faster linear decay than under the original cosine schedule. Removing M2 substantially reduces editing capability across all settings, as the guidance strength remains fixed without temporal and spatial adaptation. Furthermore, removing temporal adaptation consistently reduces editing capability, while removing spatial adaptation negatively affects both target alignment and source preservation across all settings, demonstrating that its contribution extends beyond marginal improvements in individual metrics to the overall editing–preservation balance.

Importantly, the original 0.7 + cosine configuration retains relatively strong source guidance during early and intermediate steps, which can cause ablated variants to favor source reconstruction, resulting in disproportionately large drops in target alignment. The additional settings demonstrate that these numerical changes are influenced by the guidance schedule, rather than indicating that the modules perform identical functions.

Overall, these two analyses provide complementary evidence: the matched-alignment comparisons demonstrate that each module contributes to a more favorable editing–preservation trade-off, while the extended ablations confirm consistent module behavior across different schedules and settings. The matched ablation is now included in revised Sec. 4.5, Table 4, and the functional coupling is clarified in revised Sec. 4.5 (lines 465–479).

[W1]. Relationship to DiffEdit

We thank the reviewer for pointing out DiffEdit. We have added its citation and an explicit comparison in revised Sec. 2 (lines 145–154).

Both DiffEdit and DynEdit exploit prediction differences under different text conditions to infer editing regions. However, their localization mechanisms serve different roles within fundamentally different generation processes. DiffEdit estimates an editing mask before the editing process and subsequently uses this fixed mask for source-trajectory blending during continuous DDIM denoising. In contrast, DynEdit operates on discrete masked-token generation, where the editing mask is dynamically updated at each step based on the evolving source–target predictions. Rather than merely identifying a region for editing, the resulting token-level labels jointly control spatially adaptive source guidance and the decoding priority of remaining masked tokens. This allows localization, preservation, and generation order to evolve together throughout the progressive decoding process.

Therefore, while both methods share the general intuition of prediction disagreement, DynEdit integrates this signal into a different generation paradigm and uses it to enable a dynamic, editing-aware decoding process. This conceptual connection and distinction are now explicitly discussed in revised Sec. 2 (lines 145–154).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers for their constructive feedback and for recognizing the practical value of DynEdit's training-free framework (KSHD, 6EAi, XJRx), its coordinated progressive localization, source guidance, and decoding design (6EAi, XJRx), and the convincing comparisons under identical external masks (KSHD). We appreciate the thoughtful suggestions, which have helped us strengthen the experimental validation, clarify the method, and improve the presentation.

In particular, we have addressed the shared concerns regarding the effectiveness and robustness of dynamic edit localization, the coupled contributions of the three modules, and whether the improvements extend beyond a change in the editing–preservation operating point. We also provide additional analyses of the residual–Gumbel weighting, irreversible token fixation, and relevant baselines.

Our major revisions include:

  1. Direct localization validation and robustness analysis: We added ground-truth-mask IoU curves with a matched random baseline (Appendix I, Fig. 17), along with failure-case visualizations and discussion of irreversible token fixation (Appendix J, Fig. 18).

  2. Stronger evidence for module effectiveness: We added CLIP-E-matched ablations (Sec. 4.5, Table 4) to compare source preservation at similar editing strength, and clarified the functional coupling of the modules in Sec. 4.5.

  3. Expanded baseline and trade-off comparisons: We included native Lumina-DiMOO comparisons (Appendix G, Table 9 and Fig. 15) and guidance-sweep comparisons with GIDE (Appendix H, Fig. 16).

  4. Methodological and implementation clarifications: We added the discussion of DiffEdit (Sec. 2), clarified minimum guidance constraints, residual–noise fusion, and baseline configurations (Appendix B), corrected Fig. 1(a), Fig. 3(b), and added a visualization of edit-aware decoding (Appendix K, Fig. 19).

We have also reorganized and refined the Appendix layout to improve readability and make the additional experiments and analyses easier to follow. All revisions are highlighted in blue in the revised manuscript. We hope these additions and clarifications address the reviewers' concerns and make the contributions and limitations of DynEdit clearer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.