acceptodds
Withdrawn from ICLR 2027

UniSlider: Perceptually Uniform Sliders for Continuous Image Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sliders provide an intuitive interface for continuous image editing. In current generative approaches, however, the slider is simply a rescaling of the method's strength parameter, such as an adapter coefficient, a prompt weight, or an interpolation factor. This strength relates poorly to perceptual change. The image can partially revert as the slider moves, long stretches of the range produce no visible difference, and short intervals transform the image abruptly. Remapping the strength could fix this uneven pace, but only if the trajectory is monotone, which current methods do not enforce. We therefore distinguish the slider from the strength, and require perceptual distance from the input to grow linearly with the slider value. We introduce UniSlider, a lightweight LoRA trained on a few-step editing backbone so that its strength approximates this ideal slider. Few-step sampling lets us impose this objective in pixel space without intermediate ground truth, and the backbone's output is preserved at full strength. However, a low-rank adapter cannot make the strength fully uniform. Our slider is thus an inference-time remapping of the strength, obtained by adaptive sampling. Since training optmizes to make the trajectory monotone, this remapping closes the remaining gap without extra training or parameters. On a new benchmark of 300 continuous edits evaluating uniformity, monotonicity, edit fidelity, and identity preservation, UniSlider outperforms all prior methods and is preferred in a user study.

Primary area:
generative models
Keywords:
Continuous image editing, Controllable diffusion models, Perceptual uniformity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_HucnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UniSlider focus non-uniform perceptual changes in continuous instruction-based image editing. It trains a LoRA on a few-step editing backbone using strength-dependent perceptual distances to the source image and the backbone’s full edit. Budgeted adaptive sampling then remaps internal editing strength to a user-facing slider. The paper introduces a 300-example benchmark and evaluates editing fidelity, preservation, endpoint consistency, quality, monotonicity, and uniformity through quantitative comparisons, ablations, qualitative results, and user studies.

Critical strengths

The paper studies the practically meaningful problem of non-uniform slider behavior in continuous image editing and extends evaluation beyond final outputs to the full editing trajectory, considering uniformity, monotonicity, edit fidelity, and content preservation. Training and remapping ablations, the independent perceptual metric TPIPS, and user studies provide complementary evidence beyond the optimized DreamSim metric.

Critical weaknesses
  1. Missing discussion of closely related perceptual-uniformity work [1,2] leaves the novelty boundary unclear. The submission neither cites nor discusses these works, leaving unclear whether its distinctive contribution is the dual-endpoint trajectory training, inference-time remapping, or their combination.

  2. The closest baseline comparison contains an unresolved protocol inconsistency. The appendix maps SliderEdit strength using ($\alpha=1-2$s), yet also states that its maximum-strength output equals the frozen-backbone endpoint shared with UniSlider. These statements are inconsistent. Differences in training configurations and adaptive-sampling endpoint handling also require clarification to attribute the reported gains reliably.

[1] IMPUS: Image Morphing with Perceptually-Uniform Sampling Using Diffusion Models. ICLR, 2024.

[2] AdaptiveSliders: User-aligned Semantic Slider-based Editing of Text-to-Image Model Output. CHI, 2025.

Questions for authors
  1. Please discuss the following works and clarify differences in problem setting, methodology, and contribution.

[1] IMPUS: Image Morphing with Perceptually-Uniform Sampling Using Diffusion Models. ICLR, 2024.

[2] AdaptiveSliders: User-aligned Semantic Slider-based Editing of Text-to-Image Model Output. CHI, 2025.

  1. Under the same backbone, matched endpoints, and equal generation budgets, how much additional improvement does dual-endpoint trajectory training provide over an existing continuous editor with perceptual remapping?

  2. Which SliderEdit strength mapping was actually evaluated? Are all methods assessed using consistent user-facing slider coordinates and equivalent endpoint handling?

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

Unislider investigates the problem of slider control based image editing where artists want a uniformly monotonous increasing edit capability. Their main position against prior work is the uniformity of the edit that is included in training using a DreamSim loss objective as well as slider rescaling during inference.

The primary claim of this work is that this training objective that ensures monotonous increase of edit strength along with rescaling during inference provides a more continuous edit slider missing in prior works that often violate this capability.

In implementation, the edit strength (s) is modelled as a scalar slider for LoRA parameters where higher strength uses higher scaling of the lora param weights vs lower strength reduces the lora param weights. At s=0 it should ideally provide the input image as is.

Critical strengths
  1. This problem of continuously increasing the edit strength slider that needs to monotonously increase/decrease is important to avoid dead zones and improper edits.
  2. The reviewer appreciates the capability of training this model without using intermediate images and instead just uses perceptual distance to train the model. This general formulation can in principle be extended to edits that are not only photometric but also geometric and so on. This could be highlighted better in the paper in future versions.
  3. The reviewer also appreciates the explanation for the need for adaptive sampling with figure 4 that shows distance vs slider strength vs N samples.
  4. The metrics are complete and the ablations about lora rank, the need for adaptive slider sampling with pseudo code make the method more easy to understand and reason about the design choices of the architecture.
Critical weaknesses
  1. Generalization & Scale of Data: The reviewer’s first worry is about the training dataset scale (300 image) and capability of the model. It is claimed that works such as Kontinuous Kontext train on held out data from its own curated dataset but actually use Piebench that is publicly available (that is not curated by them). Does training on 300 images perform well against other works that train on thousands of images? The reviewer appreciates if the paper shows more in the wild image edits and some small evaluation on 100-200 images from Piebench (not the full set given the rebuttal window is fine). Please add more images that show varying the slider but for piebench/in-the wild examples and compare against other works visually. The paper rules out Piebench/external benchmarks by stating that these include edits with no meaningful intermediate strengths but none of the metrics need these intermediates and can be calculated for such datasets like prior works (V_VQA, V_LPIPS, CV-DSim, strength monotonicity).
  2. It is also worth showing more carefully about deadzone vs jump slider artifacts in most models including Kontinuous Kontext that is claimed to be uniformly edited but the paper misses many visual comparisons against it. The paper does show quantitative numbers but does not show 3-4 visual images that makes their story stronger.
  3. One interesting direction that is not highlighted here is perhaps the capability to learn without intermediate slider data. Are there some easy examples that can show the benefit of this capability against prior work? Visually showing some structural/perceptual edits that other works can not perform will help here.
  4. Training is time consuming as future works will need to generate the image and then score the perceptual distance. Is it possible to do one step generation or perform this perceptual distance in feature space without having to regenerate the image at training step? One way could be to do one step generation instead of 4 steps. During training the signal is important and high fidelity image may not be? But this can be done as a future work optimization not necessary for this rebuttal.
Questions for authors

I am inclined to weak reject this work and urge authors to perform the following:

Rebuttal

  1. Show capability on dataset not curated in this work, this can be Piebench or any other slider publicly accepted slider benchmark shown in prior works. This is to ensure that there is no benchmark selection bias. Answering this raises the reviewers score
  2. Show 3-4 side by side visual examples that show continuous slider against Kontinuous Kontext as that is the prior work that tries to monotonously vary the strength of edit.

Minor

  1. Tables 1, 4 show VQA_{edit} to be a lower arrow but higher is actually better (UniSlider highlighted). Please fix the arrow direction inconsistency across all tables. There is also a typo in line 50 of abstract optmizes -> optimizes

Future Work

  1. can this be extended to geometric edits that others can not do for now? For example rotating objects, moving items etc? There may be a need for changing the distance function from DreamSim to maybe a geometric/rotational/translational distance?
Reviewer_YewtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposed a simple and intuitive approach to improve the perceptual smoothness for continuous image editing. The method involves fine-tuning an instruction-driven image editing model with LoRA. To achieve continuous control through a strength parameter s a loss based on perceptual linearity is proposed that enforces the following constraint with a distance metric: D(x_src, x_s) + D(x_s, x_edit) = D(x_src, x_edit). The LoRA is trained with a small set of 300 image-instruction pairs with this objective. An additional inference-time adaptation step is proposed that makes the edit trajectory uniform in terms of perceptual change through an adaptive subdivision algorithm, where the strength regions with large changes are divided further. The proposed method has impressive qualitative results where the transitions are smooth across different edit strengths and follow the edit instructions closely.

Critical strengths
  • The paper tackles an important problem of adding perceptual uniformity in slider-based continuous image editing
  • The paper is nicely written, with a good explanation of the model components and extensive ablation experiments are provided
  • The method is simple and intuitive; both the contributions - smoothness objective and the inference-time adaptive sampling are effective in improving perceptual smoothness of sliders and making the slider transformations monotonic, as shown in the qualitative analysis
Critical weaknesses

Generalization - The proposed method, though convincing, is implemented on a single backbone (Flux2Klien) and tested on the proposed benchmark of 300 images and instructions. The method should be tested on other base images to show generalization: QwenImage, Z-Image and Sana. Additionally, the method should be evaluated on other standard image editing benchmarks like PieBench without object addition and removal categories for completeness.

Dataset Scaling - The model is fine-tuned with only 300 examples from MIT5K which is rather small number of samples for training a general behavior of continuous control. Are there edit categories that the model has not learned during training and struggles to achieve continuous control for them during inference? How does the model improve by increasing the number of training samples? Authors should provide a data scaling experiments and do a separate evaluation for unseen categories of edits to validate the generality of learned continuous control.

Training objective - The training objective proposed in eq.4 is implemented only using the distance between the source and intermediate and final edited images. It does not explicitly ensure that the edits follow the edit instructions. This can lead to drift of the edit trajectory, where the intermediate samples mayn’t follow the edit text instruction well. Does the training also involve regular flow matching objective along with the proposed objective?

Questions for authors

Multiple steps during training - To compute the loss in eq.4 the model needs to generate the image pixels after performing 4 steps of denoising. Are the gradients backpropagated through the chain of 4 denoising steps for each step of optimization? This detail should be provided in hte main paper.

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

The paper proposes UniSlider, a method for perceptually uniform strength control in continuous image editing. The approach trains a shared LoRA on a pretrained image editing backbone, using perceptual distances between intermediate outputs, the source image, and the backbone's default edit. The objective encourages uniform progression without requiring ground-truth intermediate images. An adaptive sampling strategy further remaps the strength parameter to produce perceptually uniform slider values. The method is evaluated on a new benchmark of 300 image–instruction pairs, showing improvements in uniformity, monotonicity, and human preference.

Critical strengths

The paper addresses a relevant problem in continuous image editing. The distinction between editing strength and perceived magnitude of change is interesting and well motivated.

The training objective is simple and intuitive. Using perceptual distances to supervise intermediate strengths without requiring ground-truth intermediate images is a useful contribution.

The approach is lightweight and uses a shared LoRA across different editing instructions. The combination of intermediate supervision and adaptive sampling improves the uniformity of the resulting trajectories.

The experimental evaluation is fairly comprehensive, including comparisons with existing slider methods, perceptual metric ablations, and human studies. The improvements in uniformity and user preference are encouraging.

Critical weaknesses

Limited novelty relative to existing slider methods. Concept Sliders already introduces LoRA-based continuous attribute control, while SliderEdit learns a shared adapter for instruction-based editing. Perceptually uniform sampling has also been explored in IMPUS. The contribution appears to be the particular perceptual-distance objective and its combination with adaptive sampling rather than a fundamentally new editing mechanism.

Unclear concept-specific control. The method relies on the pretrained backbone to interpret and execute different editing concepts, while a single LoRA controls the overall edit strength. The 2D editing examples use sequential application of the method rather than jointly learned controls. It is unclear how well the approach handles instructions involving multiple independent attributes.

Perceptual distance does not necessarily correspond to semantic progression. The objective constrains distances to the source and final edited images but does not directly supervise whether the requested concept changes consistently. For example, increasing DreamSim distance does not necessarily correspond to gradually increasing age or changing an expression. The method could satisfy the distance objective through unrelated visual changes, and the paper does not sufficiently isolate instruction-specific progression.

Dependence on the perceptual metric. Both training and adaptive sampling rely on DreamSim, which measures general visual similarity rather than instruction-conditioned editing magnitude. Although the human studies support its use, the metric does not guarantee that equal distances correspond to equal semantic changes. The failure cases also demonstrate that unrelated visual modifications can affect the measured progression.

Baseline comparisons. Several methods use different editing backbones, making it difficult to fully attribute improvements to the proposed approach. Although the appendix includes a matched-backbone SliderEdit comparison, a more direct comparison controlling for training budget and inference settings would strengthen the claims.

Generalization and scope. Training uses 300 image–instruction pairs, and the benchmark is restricted to edits selected for meaningful continuous progression. It remains unclear how well the learned strength control generalizes to unseen editing concepts, more complex multi-attribute instructions, or edits without a natural continuous interpretation.

Questions for authors

How does UniSlider provide consistent control across different editing concepts? How does perceptual uniformity relate to semantic progression? Can the authors evaluate whether attributes such as age, expression, or object size progress uniformly, rather than simply producing equal amounts of visual change?

How does the method handle instructions involving multiple independent attributes? T How sensitive are the results to the choice of perceptual metric? Would an instruction-aware metric better capture the desired editing progression? Can the authors provide a more controlled comparison with SliderEdit? In particular, can the results be compared under matched backbones, training budgets, and inference settings to isolate the benefit of intermediate perceptual supervision and adaptive sampling?

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

Related papers

Discussion (0)

Sign in to comment.