acceptodds
Under review as a conference paper at ICLR 2027

SliderEdit++: Improved Slider Trajectories in Instruction-Based Image Editing

Abstract

Slider-based control aims to give users continuous, fine-grained control over the intensity of individual edits in instruction-based image editing. However, existing approaches fail to deliver reliable continuous controllability: traversing a slider can produce abrupt visual jumps and non-monotonic changes, while jointly controlled sliders can interfere with one another. To systematically study these limitations, we introduce *SlideBench*, a comprehensive benchmark for continuous image editing spanning diverse editing categories and equipped with a unified suite of metrics quantifying monotonic progression, perceptual continuity, and inter-slider disentanglement. Building on these observations, we propose SliderEdit++, which explicitly regularizes the intermediate editing trajectory rather than supervising only its endpoints. Our method combines velocity- and semantic-space trajectory regularization with spectral control to encourage smooth, predictable slider traversal, together with a disentanglement objective that reduces interference between independently controlled edits. On SlideBench, SliderEdit++ achieves state-of-the-art performance, improving the overall score by 18.5% relative to the strongest baseline. At the same time, using a few-step editing backbone makes our method approximately to faster than existing slider-based approaches, enabling substantially more responsive interactive editing.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Controllable Image Editing, Instruction-based Image Editing, Continuous Edit Strength Control, Sliders for Image Editing

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 2 reviews, above 99% of the others.
Reviewer_FW7HOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper focuses on two key limitations of existing slider-based image editing methods: the lack of truly continuous control along a single slider and cross-attribute interference when using multiple sliders. To address these issues, the authors propose SliderEdit++, which builds upon the original SliderEdit framework with four regularizers: velocity-space midpoint regularization, semantic-space midpoint regularization, spectral-norm regularization, and disentanglement regularization. Together, these objectives aim to improve the continuity and independence of editing sliders.

The authors also introduce SlideBench, a new benchmark for evaluating slider-based image editing along three key dimensions: monotonic progression, frame uniformity, and multi-slider disentanglement.

Critical strengths
  1. Clear motivation and well-written presentation. The paper is very well written and easy to follow. In particular, I appreciate how clearly the authors identify the limitations of existing slider-based editing methods. The motivation is convincing, and each component of the proposed method is well explained and naturally follows from the identified problems. The benchmark is also thoughtfully designed.

  2. A useful and intuitive benchmark. I particularly like SlideBench, as it provides a clear and intuitive way to evaluate what slider-based image editing should ideally achieve. Rather than focusing only on editing quality at individual slider positions, it also considers how edits progress throughout the trajectory. I think this is a valuable contribution to the field.

  3. Effective regularization with strong results. The proposed regularizers fit naturally into the original SliderEdit framework and lead to impressive improvements, especially in editing continuity. Both the quantitative and qualitative results are encouraging, and the multi-slider results also look promising.

Critical weaknesses

Overall, I really enjoyed this paper and found the results quite convincing. I only have two minor comments and a few questions that I'd be curious to hear the authors' thoughts on. These could also be interesting to explore in a future version.

  1. A bit more analysis of the individual regularizers.

    I appreciate the ablation study in Table 2, which nicely shows the benefits of each regularizer. One thing I'd love to see is a leave-one-out ablation to better understand how they work together. For example, Figure 6(b) only compares the full model with the version without any regularization. I'm curious what the results would look like if we removed just one regularizer, especially the disentanglement loss in the multi-slider examples (Figure 15).

    I also have two small questions about the disentanglement regularizer:

    • Quantitative improvement. The disentanglement score in Table 3 only improves slightly from 0.9854 to 0.9869. I wonder if this is partly because the metric is already quite saturated or the reason explained under.

    • Velocity vs. semantic space. I really like the motivation behind the semantic-space regularizer, where the authors point out that smoothness in velocity space doesn't necessarily mean smoothness in the final image. This makes me wonder if the same idea also applies to disentanglement. Would it make sense to encourage disentanglement directly in DINOv2 feature space, instead of only enforcing orthogonality in velocity space? I'd be interested to hear the authors' thoughts on this.

  2. The relationship between the training objectives and evaluation metrics.

    I really like SlideBench and think it provides a useful way to evaluate continuous editing. One small thing I noticed is that the semantic-space regularizer uses DINOv2 features to encourage smoother trajectories, while DINOv2 is also part of the continuity evaluation in SlideBench.

    I understand that the benchmark already includes other feature metrics and human evaluation, so this is not a major concern. Still, I'm curious how well uniform changes in DINOv2 feature space actually match what humans perceive as smooth and continuous editing. I think a short discussion on this would be helpful, and perhaps a more targeted human evaluation of intermediate slider transitions could be an interesting addition in a future version.

Questions for authors

Overall, I really enjoyed reading this paper. It is well organized, and I find the motivation and experimental results quite convincing. I would be happy to hear the authors' thoughts on the points raised in the Weaknesses section, and I currently lean toward Weak Accept.

My main hesitation in giving a higher score is that the method still builds heavily on the original SliderEdit framework, particularly its endpoint supervision and LoRA interpolation. While the proposed regularizers clearly improve the editing trajectories, I feel the broader methodological novelty and potential impact on future work are somewhat limited by this reliance. If SliderEdit had not already established this framework, I would probably lean toward Clear Accept.

One thing I'd be curious to hear more about is whether the proposed regularizers could also work with other continuous editing frameworks beyond SliderEdit. I think this would help me better understand the broader potential of this work.

5 comments

We sincerely thank the reviewer for the encouraging feedback and for recognizing the clear motivation and presentation of our work, the usefulness and intuitive design of SlideBench, and the effectiveness of our proposed regularizers. We appreciate the thoughtful suggestions and address each point below.


1. Further analysis of the regularizers

Leave-one-out ablation. We thank the reviewer for this helpful suggestion. While Table 2 already demonstrates the individual contributions of each regularizer, we agree that a leave-one-out ablation would provide further insight into their complementary effects. We therefore extend our ablation by evaluating all pairwise combinations of the regularizers under the same backbone and experimental setup.

ConfigurationLsem\mathcal{L}_{\mathrm{sem}}Lvel\mathcal{L}_{\mathrm{vel}}Lspec\mathcal{L}_{\mathrm{spec}}MP Avg ↑Uniformity Avg ↑Overall ↑
Baseline✗✗✗0.71610.72580.7209
Semantic✓✗✗0.76430.80600.7851
Velocity✗✓✗0.75520.79010.7726
Spectral✗✗✓0.74210.79640.7692
Semantic + Velocity✓✓✗0.77580.82960.8027
Semantic + Spectral✓✗✓0.77210.82510.7986
Velocity + Spectral✗✓✓0.76340.81720.7903
All Regularizers✓✓✓0.78450.83860.8116

The leave-one-out ablation further supports the complementary effects of our proposed regularizers. Combining any two regularizers improves performance over their individual application, while the full objective consistently achieves the best results across all metrics. These findings suggest that each regularizer contributes distinct benefits to trajectory smoothness and continuity, and that their combination is particularly effective. We thank the reviewer for this helpful suggestion, which provides a more comprehensive understanding of how the proposed regularizers interact.

We will incorporate these additional ablations into the revised appendix, together with further qualitative examples illustrating the contribution of each regularizer. We also agree that slider quality is particularly intuitive to assess through an interactive interface, where users can directly observe the continuity of transitions. We plan to make such an interface publicly available upon publication, allowing users to explore and compare different training configurations interactively.

Saturated disentanglement metric. We agree with the reviewer that the disentanglement scores in Table 3 are highly saturated, making them relatively insensitive to meaningful improvements in multi-slider behavior. To better capture these differences, we conducted an extensive additional human evaluation involving 10,000 paired comparisons, where participants selected the model providing a better and more disentangled multi-slider editing experience.

SliderEdit (Baseline)SliderEdit++ (Ours)No Preference
41%56%3%

The results show a clear preference for SliderEdit++, with 56% favoring our method compared to 41% for the baseline. These findings provide further perceptual evidence that our framework improves practical multi-slider controllability beyond what is captured by the saturated automatic metric.

Velocity-space vs. semantic-space disentanglement. We appreciate this insightful suggestion and agree with the reviewer's reasoning. As discussed in our paper, while encouraging orthogonality in velocity space helps reduce interference between different sliders, it may not fully guarantee semantic independence in the final generated images due to the nonlinear generation process. Therefore, extending the disentanglement objective to semantic feature spaces, such as DINOv2, is a well-motivated direction that could further improve the independence of semantic edits.

Our primary focus in this work is improving editing continuity, with disentanglement explored as an additional extension. We believe the reviewer's proposed direction could further strengthen multi-slider control and represents a promising avenue for future work.

2. Relationship between training objectives and evaluation metrics

We thank the reviewer for raising this important point. Since DINOv2 is used in our semantic-space regularizer, we specifically evaluate continuity in multiple independent feature spaces, including LPIPS, CLIP, and SigLIP, rather than relying solely on DINOv2. This helps ensure that the observed improvements are not simply a consequence of optimizing the same representation used for evaluation. In addition, we complement these automatic metrics with human preference studies.

Regarding the relationship between DINOv2 feature distances and perceived smoothness, Figure 3 provides a qualitative analysis illustrating how abrupt semantic transitions correspond to large fluctuations in consecutive-frame DINOv2 distances. In our experiments, we generally observed that more uniform DINOv2 distances correspond to visually smoother editing trajectories. Nevertheless, we acknowledge that DINOv2 is an imperfect proxy for human perception and may not capture every type of semantic transition equally well.

We also believe that more sophisticated formulations of semantic progression could further improve both training and evaluation. For example, Semantic Progress Function (SPF) [1] models cumulative semantic change along a sequence and characterizes deviations from uniform semantic progression. Such approaches provide an interesting alternative for analyzing editing trajectories and could potentially be incorporated into our framework. Importantly, our semantic-space regularization is not inherently tied to DINOv2, and exploring alternative representations or semantic progression objectives is a promising direction for future work.


3. Generalizability beyond SliderEdit

We especially appreciate the reviewer's interest in the broader applicability of our approach. Importantly, our proposed regularizers are not restricted to SliderEdit. They operate on intermediate editing trajectories and can be incorporated into other slider-training objectives.

To demonstrate this, we conducted an additional experiment applying our regularizers to KontinuousKontext, which adopts a different slider-training objective from SliderEdit. We report the results below alongside the original SliderEdit ablation:

MethodGPT-5 ↑Opus-4.8 ↑MP Avg ↑LPIPS ↑CLIP ↑SigLIP ↑DINOv2 ↑Uniformity Avg ↑Overall ↑
SliderEdit0.71000.72220.71610.70300.73100.73500.73420.72580.7209
SliderEdit + Ours0.78010.78900.78450.82720.84310.84430.83990.83860.8116
KontinuousKontext0.70730.69920.70330.62230.66260.66300.65850.65160.6774
KontinuousKontext + Ours0.74120.74380.74250.73160.75890.76140.75490.75170.7471

As shown, our regularizers consistently improve performance across both frameworks. Specifically, the overall score increases from 0.7209 to 0.8116 for SliderEdit and from 0.6774 to 0.7471 for KontinuousKontext. Both monotonic progression and frame uniformity also improve.

These additional results demonstrate that our approach is not merely a refinement of SliderEdit, but offers a more general strategy for improving continuous editing trajectories across different training objectives. We will include this experiment in the revised manuscript.


We sincerely thank the reviewer again for the positive assessment, thoughtful questions, and constructive suggestions, which have helped us further strengthen our work. We hope our responses and additional experiments adequately address the reviewer's concerns.


[1] Metzer, Gal, et al. "Video Analysis and Generation via a Semantic Progress Function."

We thank you again for your valuable feedback and comments which have helped to strengthen our paper. As the discussion period is ending soon, we would really appreciate if you could let us know if our responses have addressed your concerns. We will be happy to answer any further questions and address any remaining concerns to the best of our abilities in the remaining time!

Reviewer_2269OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper studies continuous control in instruction-based image editing, focusing on smooth editing trajectories and independent control of multiple attributes.

The authors first introduces SlideBench, a benchmark evaluating monotonic progression, frame uniformity, and multi-slider disentanglement.

They then introduce their method, SliderEdit++, which trains suppression-based LoRA sliders with velocity-space and semantic-space midpoint regularization, a spectral-norm penalty, and an orthogonality objective for multi-slider control. ]

Critical strengths

The paper addresses a practical limitation of existing editing sliders: intermediate settings can produce abrupt and unpredictable changes even when the endpoints behave reasonably. Explicitly regularizing intermediate trajectories is a well-motivated extension of the existing suppression-based framework. The proposed new benchmark is also useful.

Critical weaknesses

My main concern is that the evaluation, and the model itself does not distinguish improved smoothness over the same editing range from a loss of low-strength editing capability. That is to say, the reported gains do not establish that smoother trajectories are obtained while preserving the range of edits available to users.

According to Section 4.1, the LoRA is trained to suppress the editing instruction. At alpha = 0, the LoRA is disabled and the frozen backbone produces the full edit. At alpha = 1, the LoRA is trained to suppress that edit and recover a neutral state. Weakening the adapter can therefore weaken edit suppression while leaving the fully edited endpoint unchanged.

For example, a "beard" slider should allow users to move from no beard, through light stubble and progressively stronger edits, to a full beard. If suppression becomes weaker, its lowest-edit setting may already contain a substantial beard. The remaining trajectory may look smoother, but users have lost the ability to produce no beard or light stubble. Keeping the numerical slider interval at [0,1] does not guarantee that this editing range is preserved.

And the metrics used by the benchmark in Section 3.2 do not adequately account for this issue:

  • "Monotonic progression": checks whether successive frames apply the edit more strongly. A trajectory restricted to strong edits can still be strictly monotonic.

  • "Continuity": defined as U=exp(−std(d1,...,dT−1))U = exp(-std(d_1, ..., d_{T-1})) , where the distances measure changes between consecutive frames. Multiplying all distances by a factor rr between 0 and 1 reduces their standard deviation and increases UU whenever the original steps are nonuniform. Their relative unevenness remains unchanged.

  • The edit-strength weight ee: measures how strongly the final frame exhibits the intended edit relative to the source image. It does not measure the editing range covered between the slider’s own endpoints. An unchanged fully edited endpoint can therefore retain the same ee even when the nominally neutral endpoint is already strongly edited.

A simple numerical example illustrates the problem. In an idealized one-dimensional attribute model, let 0 denote the unedited source and 1 the full edit, with feature distance equal to the absolute difference in attribute strength. Consider these trajectories, ordered from weaker to stronger edits: A=[0,0.1,0.2,0.3,0.4,0.5,0.6,1]A = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 1] B=[0.8,0.82,0.84,0.86,0.88,0.9,0.92,1]B = [0.8, 0.82, 0.84, 0.86, 0.88, 0.9, 0.92, 1]

B covers only 20% of A’s editing range. Both trajectories still have a final step four times larger than their other steps, but B’s inter-frame distances are five times smaller. Assuming the changes remain distinguishable to the VLM judge, both trajectories can obtain MP = 1 and e = 1. Using population standard deviation, their uniformity scores are approximately 0.9003 and 0.9792. The trajectory that loses 80% of the editing range receives the higher score.

The training objective itself provides a concrete mechanism through which editing range can be lost. Consider scaling every effective LoRA update by a constant cc between 0 and 1. At each layer: Wbase+α∗(c∗ΔW)=Wbase+(c∗α)∗ΔWW_{base} + \alpha * (c * \Delta W) = W_{base} + (c * \alpha) * \Delta W

The new slider therefore traverses only the original parameter interval [0,c]. Its fully edited endpoint remains unchanged, while its nominally neutral endpoint may retain substantial editing. This identity follows directly from the parameterization and does not require the network to be linear.

PPS could somehow suppress this issue, but it is a soft loss and does not guarantee preservation of the neutral endpoint.

Some qualitative results are consistent with this concern. In the 2nd (from top to bottom) example of Figure 13, the leftmost SliderEdit++ output already has a visible beard, while the baseline (SliderEdit) remains clean-shaven. In Figure 14, the leftmost SliderEdit++ outputs in the 2nd and 3rd (from top to bottom) examples already show darkened hair and visible ivy, respectively.

Questions for authors
  • Could you explain the issues described in the critical weaknesses part? This is my critical concern for this submission.
  • In line 161, you said DD (Disentanglement) is the lower the better, but in Table 3 the value for Disentanglement is the large the better. Could you please explain for that?
4 comments

We thank the reviewer for recognizing the practical importance of our work, the motivation behind explicitly regularizing intermediate editing trajectories in sliders, and the value of the proposed benchmark. We appreciate the thoughtful analysis and constructive feedback. Below, we address the concerns individually.


1. Preservation of the editing range

We absolutely agree with the reviewer that improved smoothness should not come at the expense of the available editing range, and that our evaluation should explicitly account for this possibility.

In fact, this concern closely aligns with the motivation behind incorporating VLM Edit Strength into our Frame Uniformity metric. During the development of SlideBench, we observed that certain sliders or editing backbones produce weaker edits, resulting in smaller inter-frame differences and potentially higher uniformity scores without necessarily providing better continuous control. We therefore introduced Edit Strength to penalize such behavior.

However, as the reviewer correctly points out, Edit Strength alone does not guarantee preservation of the entire editing range, since a slider could produce a strong final edit while failing to recover the neutral endpoint. For suppression-based sliders such as SliderEdit, our initial assumption was that the trained suppression LoRA at α=1\alpha=1 effectively removes the instruction, making the original Edit Strength sufficient. While this behavior was consistently observed in our training experiments, we agree that the benchmark metric itself should not rely on this assumption.

Revised continuity metric. Following the reviewer's suggestion, we revise the continuity metric to explicitly account for the editing range. The original metric is

Uˉϕ=1M∑i=1MUiϕei,\bar{U}^{\phi}=\frac{1}{M}\sum_{i=1}^{M}U_i^{\phi}e_i,

where eie_i measures the edit strength at the fully edited endpoint. We replace eie_i with the Edit Range rir_i, defined as the difference between the edit strengths at the fully edited and neutral endpoints:

ri=eifull−eineutralr_i = e_i^{\mathrm{full}} - e_i^{\mathrm{neutral}}

The revised metric becomes

Uˉϕ=1M∑i=1MUiϕri\bar{U}^{\phi} = \frac{1}{M}\sum_{i=1}^{M}U_i^{\phi}r_i

Here, eifulle_i^{\mathrm{full}} and eineutrale_i^{\mathrm{neutral}} are evaluated using the same VLM-based edit-strength evaluator. This modification explicitly penalizes sliders with incomplete instruction suppression, preventing a restricted editing range from being rewarded solely for producing smaller, more uniform transitions. We thank the reviewer for highlighting this limitation and suggesting an important improvement to the evaluation framework.

Experimental verification of instruction suppression. To test whether our regularizers compromise the neutral endpoint, we take 100 SlideBench source images (sampled across all categories) and generate each edit at the neutral endpoint (α=1\alpha=1) with (i) SliderEdit++ and (ii) the same model trained with PPS only (no regularizers). We compare each output to its source image using LPIPS, DINOv2, CLIP-I, PSNR and SSIM. We also report the residual edit: the CLIP text-alignment gain toward the target edit, as a percentage of the full edit (0% = fully suppressed, 100% = full edit). As references, we include the backbone with the instruction "Keep the image the same" (its reconstruction fidelity) and the backbone with the full edit (α=0\alpha=0).

Output vs. sourceLPIPS ↓DINOv2 ↑CLIP-I ↑PSNR ↑SSIM ↑Residual edit ↓
Backbone, "Keep the image the same"0.1100.9960.99231.30.924~1%
PPS only, α=1.00.1150.9920.98730.30.9176%
PPS + Regularizers (SliderEdit++), α=1.00.1470.9710.97529.10.88211%
Backbone, full edit (α=0)0.3770.8340.85916.50.645100%

The regularizers only slightly reduce the editing range compared to PPS only: at α=1\alpha=1, SliderEdit++ remains close to the source (LPIPS 0.147 vs. 0.115, DINOv2 0.971 vs. 0.992), with a residual edit of 11% vs. 6%, and far from the full edit (LPIPS 0.377). Thus, SliderEdit++ does not collapse the editing range and preserves most of it. In practice, this small residual is acceptable: users apply a slider because they want the edit, so they rarely need a fully unedited image at its lowest setting.

Results using the revised metric. We re-evaluate all methods on SlideBench using the revised metric, Frame Uniformity weighted by the Edit Range rir_i. SliderEdit++ still achieves the best score in every feature space, confirming that its improvement in continuity does not come from a reduced editing range.

MethodLPIPSCLIPSigLIPDINOv2Avg
KontinuousKontext0.59430.63230.63270.62820.6219
SDE-Edit0.39240.40080.40110.39660.3977
VeloEdit0.50250.54020.54180.53700.5304
Explicit CFG0.58770.62510.62650.62010.6149
SliderEdit0.49880.54100.54310.53790.5302
Ours0.64260.68380.68400.68440.6737

2. Disentanglement metric

Thank you for catching this. Dk→jD_{k→j} as defined in Sec. 3.2 is a mean absolute deviation (lower is better). For consistency with the other columns of Table 3, which are all on a higher-is-better (0,1](0,1] scale, we report exp(−D)exp(−D) in the table. We used the same mapping for frame uniformity, where U=exp(−std)U = exp(−std). Higher values in Table 3 therefore mean better disentanglement, and SliderEdit++ improves over SliderEdit in all three feature spaces. We will state this mapping explicitly in Sec. 3.2 and in the Table 3 caption.

Furthermore, we observed that the disentanglement metric is highly saturated and does not adequately capture the improvements introduced by our method. To provide a more reliable evaluation, we conducted an extensive human study involving 10,000 paired comparisons, where participants were asked to select the multi-slider model that provides a better and more disentangled editing experience. The results are summarized below:

SliderEdit (Baseline)SliderEdit++ (Ours)No Preference
41%56%3%

These results further demonstrate that our proposed regularizers improve the practical quality of multi-slider control, providing users with a smoother and more disentangled editing experience.


We sincerely thank the reviewer for the insightful comments and constructive suggestions, which have helped us further strengthen our work. We hope our responses and additional experiments adequately address the reviewer's concerns.

We thank you again for your valuable feedback and comments which have helped to strengthen our paper. As the discussion period is ending soon, we would really appreciate if you could let us know if our responses have addressed your concerns. We will be happy to answer any further questions and address any remaining concerns to the best of our abilities in the remaining time!

We thank you again for your valuable feedback. As the discussion period is nearing its end, we would be happy to clarify any remaining questions or concerns you may have. Thank you again for your time and consideration!

Comments on the paper

4 comments

We thank the reviewer for recognizing the clear motivation of our work, the importance of regularizing intermediate editing trajectories in sliders, and the value of our controlled ablations, comprehensive evaluations, and human studies. We appreciate the constructive feedback and address the concerns below.


1. Methodological contribution and generalizability

We respectfully believe that the contributions of our work extend beyond a local refinement of SliderEdit. Specifically, our work (i) introduces SlideBench, a comprehensive benchmark for evaluating continuous image editing across diverse editing categories and multiple complementary metrics; (ii) identifies and systematically investigates the discontinuity and disentanglement issues in existing editing sliders; and (iii) proposes well-motivated regularizers that explicitly address these limitations, achieving state-of-the-art performance with an 18.5% relative improvement over the strongest baseline.

Importantly, our proposed regularizers are not specific to SliderEdit or a particular training objective. They operate on intermediate editing trajectories and can be incorporated into other slider-training frameworks. To directly address the reviewer's concern regarding generalizability, we conducted an additional experiment applying our regularizers to KontinuousKontext, which uses a different training objective from SliderEdit.

The results are summarized below, alongside our original SliderEdit ablation:

MethodGPT-5 ↑Opus-4.8 ↑MP Avg ↑LPIPS ↑CLIP ↑SigLIP ↑DINOv2 ↑Uniformity Avg ↑Overall ↑
SliderEdit0.71000.72220.71610.70300.73100.73500.73420.72580.7209
SliderEdit + Ours0.78010.78900.78450.82720.84310.84430.83990.83860.8116
KontinuousKontext0.70730.69920.70330.62230.66260.66300.65850.65160.6774
KontinuousKontext + Ours0.74120.74380.74250.73160.75890.76140.75490.75170.7471

Our regularizers consistently improve both training frameworks, increasing the overall score from 0.7209 to 0.8116 on SliderEdit and from 0.6774 to 0.7471 on KontinuousKontext. Improvements are also observed in both monotonic progression and frame uniformity.

These results demonstrate that the proposed regularizers are not merely an optimization of a particular SliderEdit training setup, but provide a more general approach for improving intermediate editing trajectories across different slider-training objectives. We will include this additional evaluation in the revised manuscript.


2. Multi-slider disentanglement and saturated metrics

We agree with the reviewer that the automatic disentanglement scores in Table 3 are highly saturated (approximately 0.98–0.99), making them relatively insensitive to meaningful differences in multi-slider editing behavior. As a result, these metrics alone do not adequately reflect the practical improvements introduced by our method.

To provide a more reliable assessment, we conducted an extensive additional human evaluation involving 10,000 paired comparisons, where participants were asked to select the model providing a better and more disentangled multi-slider editing experience.

SliderEdit (Baseline)SliderEdit++ (Ours)No Preference
41%56%3%

The results show a clear preference for SliderEdit++, with 56% favoring our method compared to 41% for the baseline. These findings provide additional perceptual evidence that the proposed training framework improves multi-slider controllability beyond what is captured by the saturated automatic metrics.

3. Controlled ablations and source of the efficiency gains

We thank the reviewer for highlighting the importance of separating the contributions of the proposed regularizers from those of the underlying backbone.

First, all experiments in Table 2 use the same FLUX-Klein-4B backbone, training configuration, and evaluation protocol. The only difference is which regularization terms are enabled. This controlled setting directly isolates the impact of the proposed regularizers, independently of any change in the underlying model or inference configuration.

For clarity, we reproduce the results below:

MethodMP ↑Uniformity ↑Overall ↑
Baseline (no regularizers)0.71610.72580.7209
+ Semantic regularizer0.76430.80600.7851
+ Velocity regularizer0.75520.79010.7726
+ Spectral regularizer0.74210.79640.7692
+ All regularizers0.78450.83860.8116

As shown, each regularizer independently improves performance, with the complete objective achieving the best results. In particular, the overall score increases from 0.7209 to 0.8116 under the same backbone and training setup, demonstrating that the improvements are attributable to the proposed regularization rather than a change in model architecture.

Second, we clarify that the reported 3–20× inference speedup comes from adopting the few-step FLUX-Klein-4B backbone, not from the regularizers themselves. Our regularizers improve slider quality during training without introducing additional inference steps. Thus, our contribution is improving the quality of continuous control while retaining the computational efficiency of the underlying few-step model.

Furthermore, Table 4 in the appendix demonstrates that our method remains effective even with very few generation steps, including single-step inference, further supporting its suitability for fast, interactive image editing.

StepsMP Avg ↑LPIPS ↑CLIP ↑SigLIP ↑DINOv2 ↑Uniformity Avg ↑Overall ↑
80.77940.79770.81300.81460.81040.80890.7942
40.77890.81410.82750.82870.82440.82370.8013
20.76220.82910.83960.84110.83780.83690.7995
10.72880.83040.83690.83870.83610.83550.7822

These results demonstrate that SliderEdit++ maintains strong editing performance even with a single generation step, with only a modest reduction in overall score compared to the eight-step setting.

We will make this distinction more explicit in the revised manuscript to avoid conflating improvements in editing quality with those arising from the choice of backbone.


We sincerely thank the reviewer for the insightful comments and constructive suggestions, which have helped us further strengthen our work. We hope our responses and additional experiments adequately address the reviewer's concerns.

We thank you again for your valuable feedback and comments which have helped to strengthen our paper. As the discussion period is ending soon, we would really appreciate if you could let us know if our responses have addressed your concerns. We will be happy to answer any further questions and address any remaining concerns to the best of our abilities in the remaining time!

We thank you again for your valuable feedback. As the discussion period is nearing its end, we would be happy to clarify any remaining questions or concerns you may have. Thank you again for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.