PaperlayerView exact revision on arXiv

Source-linked AI summary

UniSlider: Perceptually Uniform Sliders for Continuous Image Editing

David Serrano-Lozano, Duygu Ceylan, Yannick Hold-Geoffroy, Iliyan Georgiev, Javier Vazquez-Corral, Anna Frühstück

arXiv:2610.06831v1cs.CVcs.AIcs.GR

TL;DR

Generative sliders often map an internal strength directly to the user control, despite uneven or reversing perceptual change. UniSlider trains a lightweight LoRA for perceptual progress and adaptively remaps strength; it ranks highly on the benchmark and wins most user-study comparisons.

  • Problem

    Generative sliders often produce non-uniform perceptual change and reversals because the user-facing slider is tied directly to the method’s internal strength.

    Source: §1 · p. 2

  • Method

    UniSlider trains a lightweight LoRA on a few-step editing backbone for perceptually uniform progress, then adaptively remaps strength at inference.

    Source: §1 · p. 2, §4 · p. 5

  • Results

    UniSlider ranks first or second across benchmark metrics and wins 82% of user-study comparisons for both overall quality and evenness of change.

    Source: §6 · p. 7

  • Takeaways & Limitations

    The benchmark confirms a slider whose value has the same meaning across images and edits, measured jointly with edit fidelity and identity preservation.

    Source: §7 · p. 9

    Uniformity is defined under perceptual metric D rather than human perception, and adaptive sampling approaches but does not exactly reach uniformity under D.

    Source: §7 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

Generative sliders tie the user-facing control to an internal strength that often produces uneven, non-monotone perceptual change. UniSlider trains for perceptual uniformity, then adaptively remaps strength; on a 300-case benchmark, it performs consistently best across the metrics.

1 INTRODUCTION

  • Existing sliders can have dead zones, abrupt jumps, and reversals, leaving only a few effective settings that depend on the edit.Perceptual uniformity had not, to the authors’ knowledge, been posed as an objective for generative image-editing sliders.Source: §1 · p. 2
  • UniSlider trains a lightweight LoRA on a few-step editing backbone so perceptual change grows with strength, then adaptively remaps strength into a more uniform user-facing slider.The few-step backbone allows pixel-space training without ground-truth intermediate images; the low-rank adapter alone cannot reach full uniformity.Source: §1 · p. 2
  • The benchmark addresses the lack of shared evaluation for continuous editing, which reference-image metrics and general editing benchmarks do not adequately cover.Figure 2 illustrates the method’s progression from SliderEdit through the loss-only variant to UniSlider with strength reparametrization.Source: §1 · p. 2, Figure 2 · p. 3
  • UniSlider performs consistently best across metrics on a 300-case benchmark spanning three editing categories and evaluating uniformity, edit fidelity, and identity preservation.The benchmark covers 300 continuous text-based image edits.Source: §1 · p. 2

2 RELATED WORK

Prior generative sliders expose a continuous strength axis, either from the inference process or from a learned adapter. Their intermediate trajectories are generally not constrained to behave predictably, which is the gap UniSlider targets.

2 RELATED WORK

  • Training-free methods expose inference coefficients without training data, while training-based methods learn a strength axis, often through a LoRA adapter.Training-free examples include classifier-free-guidance scale, semantic-space interpolation, and velocity-field decomposition.Source: §2 · p. 3
  • Prior training-based sliders use varied supervision, including contrastive pairs, synthetic material variations, morphed intermediate edits, or neutral-instruction self-supervision.Concept Sliders [Gandikota et al., 2024], Alchemist [Sharma et al., 2024], Kontinuous Kontext [Parihar et al., 2026], and SliderEdit [Zarei et al., 2026] exemplify these approaches.Source: §2 · p. 3
  • Most methods linearly rescale learned strength for the user, leaving intermediate behavior unconstrained and allowing non-uniformity to persist.Kontinuous Kontext is a partial exception, using morphed trajectories; its output quality is tied to that morphing procedure.Source: §2 · p. 3

3 METHOD

UniSlider extends a few-step editor into a continuous family between the input and its default edit. Its perceptual objective encourages uniform progress along that path, while training and spatial masking support the method’s implementation.

3 METHOD

  • The continuous family f(xsrc, c, s) runs from the input at s=0 to the backbone’s default edit at s=1.The intermediate outputs should remain on the perceptual path and avoid changes unrelated to the instruction.Source: §3 · pp. 3–4
  • The objective requires the input-to-output perceptual distance to grow linearly with s while the input-to-output and output-to-default-edit distances sum to the endpoint distance.The paper uses DreamSim as D; the symmetric constraints promote uniformity, monotonicity, and path consistency without intermediate ground-truth images.Source: §3 · p. 4
  • A LoRA on the DiT projection layers scales with slider strength, and its contribution vanishes at s=1, preserving the frozen backbone’s default edit.The method uses few-step text-based diffusion transformer backbones with four inference steps or fewer.Source: §3 · p. 4
  • Training samples s uniformly, runs the full sampling process, and evaluates decoded pixels because the backbone uses few inference steps.The default edit is cached from the frozen backbone at s=1.Source: §3 · p. 4, Figure 3 · p. 4
  • The LoRA training set contains 300 MIT5K images with image-specific continuous-edit instructions generated by Qwen3-VL-30B.Examples are provided in the supplementary material.Source: §3 · p. 5
  • For spatially aligned latent backbones, a soft mask blends edited regions with preserved input features to allow smooth spatial transitions.Mask values of 1 denote edited regions and values of 0 denote preserved regions.Source: §3 · p. 5

4 IMPROVING PERCEPTUAL UNIFORMITY WITH ADAPTIVE SAMPLING

UniSlider’s training encourages smooth, monotone trajectories, but limited adapter capacity prevents perfect uniformity across arbitrary image–instruction pairs. Adaptive sampling remaps strength at inference to make perceptual distance grow approximately linearly, without retraining or adding parameters.

Training objective

  • The training objective encourages endpoint consistency, adherence to the editing path, and smooth, monotone changes, but low-rank capacity cannot ensure perfect uniformity for arbitrary pairs.Figure 4 measures normalized DreamSim distance from input to output; the diagonal represents the ideal uniform slider.Source: §4 · p. 5

Capacity limitation

  • Increasing the number of training examples eventually reduces uniformity; higher adapter rank delays this transition but costs parameters and computation.Figure 4 evaluates ranks r ∈ {4, 8, 16, 32} and overfitting-set sizes N ∈ {1, 2, 4, 8, 16, 32}.Source: §4 · p. 5

Adaptive sampling

  • Adaptive sampling recursively splits the interval with the largest perceptual distance, then remaps strength to slider values using the learned monotone trajectory.The mapping u = g(s) makes perceptual distance from the input grow approximately linearly with the user-facing slider value.Source: §4 · p. 5
  • ×3.7 closer to the ideal diagonal in RMSE, adaptive sampling improves uniformity on 32 unseen examples without extra training or parameters.The full-dataset adapter and sampling budget are unchanged; the evaluation uses the N = 32 examples shown in Figure 4’s rightmost column.Source: §4 · p. 5, Figure 4 · p. 6

5 BENCHMARK

The benchmark targets continuous image editing, which existing editing benchmarks are not designed to assess. It comprises 300 curated image–instruction pairs spanning three edit categories and evaluates six slider properties.

Benchmark

  • The benchmark contains 300 curated image–instruction pairs spanning Global or Stylization, Attribute or Local, and Structural edits.The categories cover consistent image-wide changes, specific objects or regions, and semantic content changes, respectively.Source: §5 · p. 6

Benchmark construction

  • Image–instruction pairs use diverse Adobe-MIT-FiveK images and edits with meaningful intermediate strengths, excluding inherently binary operations such as object removal.Pairs are manually curated for meaningfulness and diversity across the three categories.Source: §5 · p. 6

Metrics

  • Six properties are evaluated: edit fidelity, identity preservation, endpoint consistency, image quality, strength monotonicity, and perceptual uniformity.The benchmark averages its metrics over 300 examples; the table marks each column’s best and second-best methods.Source: §5 · p. 6, Table 1 · p. 7

6 EXPERIMENTS

On the 300-example benchmark, UniSlider ranks first or second across metrics, with its largest gains in monotonicity and uniformity; it also wins 82% of user-study comparisons on both overall quality and evenness. Analyses find that training and adaptive sampling complement one another, while qualitative examples show evenly spaced and composable edits.

Benchmark analysis

  • UniSlider ranks first or second across benchmark metrics, with its largest gains in monotonicity and uniformity; it preserves full-strength editing capability alongside SliderEdit.The comparison averages metrics over 300 benchmark examples; other methods degrade the base model’s full-strength editing capability.Source: §6 · p. 7

User study

  • 82% of user-study comparisons preferred UniSlider for both overall quality and evenly spread change, outperforming the second-best method by a large margin.The 2AFC study involved 25 observers and 881 slider pairs from seven methods.Source: §6 · p. 7

Qualitative results

  • In qualitative examples, intermediate strengths are perceptually evenly spaced, and composed masked and global edits remain approximately uniform and largely independent across both axes.The composed edits use no composed or masked training examples; moving along one axis neither undoes nor amplifies the other edit.Source: §6 · p. 8, Figure 6 · p. 8

Perceptual metric analysis

  • DreamSim and other perceptual metrics correlate best with human judgments of edit magnitude (PLCC 0.81–0.86), while pixel RMSE scores 0.74 and color histograms 0.51.In the metric ablation, DreamSim performs best on endpoint consistency and monotonicity; pixel-space RMSE and color-histogram distance have three to four times larger error at s = 0.Source: §6 · pp. 8–9, Table 2 · p. 9

Training analysis

  • Endpoint-only supervision yields sliders worse by a factor of 3 in monotonicity and uniformity, even when combined with adaptive sampling.Constraining only the lower endpoint leaves intermediate edit strengths unconstrained.Source: §6 · p. 9

Adaptive sampling analysis

  • Training and adaptive sampling are complementary: UniSlider without remapping matches or outperforms every remapped baseline, and their combination performs best overall.Under TPIPS, remapping improves UniSlider by 9%; it reduces UniSlider’s VVQA by 13%, while SliderEdit’s VVQA increases by 34%.Source: §6 · p. 9, Table 3 · p. 9

7 CONCLUSIONS

UniSlider separates the slider value from editing strength, combining perceptual training with inference-time remapping to approximate uniformity. The conclusion also notes that its behavior remains bounded by the editing backbone, the perceptual metric, and the rendering budget.

7 CONCLUSIONS

  • UniSlider trains a lightweight LoRA with a pixel-space perceptual objective, then remaps strength at inference to compensate for non-uniformity the adapter cannot represent alone.The benchmark measures uniformity jointly with edit fidelity and identity preservation.Source: §7 · p. 9
  • The method cannot correct backbone failures, and metric-based uniformity may differ from human perception; fixed-budget adaptive sampling also does not reach exact uniformity.The conclusion identifies training a strength-perceptually-uniform editing model, without remapping, as a next step.Source: §7 · p. 9

A LLM USAGE STATEMENT

  • The authors report using LLMs for manuscript editing and code assistance, and Qwen3-VL-30B to generate training instructions; they reviewed and verified outputs.Source: A LLM USAGE STATEMENT · p. 12

B FURTHER RELATED WORK DISCUSSION

  • Prior text editors moved from inversion-based editing to instruction-tuned feed-forward models, then to few-step rectified-flow editors that make pixel-space training affordable.Instruction-tuned editors still produce one edit rather than a family of edit strengths.Source: B FURTHER RELATED WORK DISCUSSION · p. 12
  • Adaptive sampling starts from trajectory endpoints and a midpoint, then uses a monotone PCHIP fit to place new renders at estimated perceptual midpoints of the largest gaps.Frames closer than ε to a displayed neighbor are omitted from display but retained as measurements.Source: B FURTHER RELATED WORK DISCUSSION · p. 13
  • Spatial, mask-based, point-based, and attribute controls address edit location, geometry, or attributes, while UniSlider targets edit magnitude.These interactive approaches generally apply a selected region, point, or attribute all-or-nothing.Source: B FURTHER RELATED WORK DISCUSSION · p. 13

C FURTHER METHOD DETAILS

Adaptive sampling remaps strength using perceptual-distance measurements, while the method preserves the backbone’s default edit at full strength. Its practical behavior is bounded by render budgets, backbone quality, and the chosen perceptual metric.

C FURTHER METHOD DETAILS

  • The adapter vanishes at strength s = 1, leaving the frozen backbone’s default edit as the target; supervision uses image-instruction pairs without intermediate targets.The implementation trains a single LoRA on attention projections; training uses 2,000 steps and four-step inference.Source: C FURTHER METHOD DETAILS · p. 13
  • Adaptive sampling measures the fraction of total DreamSim distance covered at each strength, then uses a monotone PCHIP fit to place renders at perceptual midpoints of the largest gaps.A uniform slider targets ϕ(s) = s; the method begins with the input, default edit, and a rendered midpoint.Source: C FURTHER METHOD DETAILS · pp. 13–14
  • Frames within ε = 0.03 of a displayed neighbor in normalized perceptual distance are omitted from display but retained as measurements for the spline.Source: C FURTHER METHOD DETAILS · p. 14
  • Adaptive sampling assumes monotone ϕ; rendering 11 Flux-Klein-4B frames averaged 7.02 s without and 7.16 s with adaptive sampling, a 2% increase on an NVIDIA RTX Pro 6000.Source: C FURTHER METHOD DETAILS · p. 14
  • UniSlider cannot correct errors in the backbone’s default edit, so an unrequested color change can persist along the slider trajectory.Source: C FURTHER METHOD DETAILS · p. 14, Figure 7 · p. 15
  • Uniformity under D is only a proxy for human perception, and DreamSim may count an unrelated color change as progress before the requested edit is visible.The authors note that text-conditioned TPIPS could isolate instruction-related change, but its 14 billion parameters make it too expensive for interactive use.Source: C FURTHER METHOD DETAILS · p. 14
  • With 11 renders, adaptive sampling brings the perceptual-distance curve much closer to the diagonal than Vanilla, and its frames are visually similar to a 60-render dense reference.The dense reference matches the diagonal almost exactly; adaptive sampling uses less than one-fifth as many renders.Source: C FURTHER METHOD DETAILS · p. 14, Figure 8 · p. 15

D FURTHER TRAINING DATASET DETAILS

The training dataset comprises 300 image-instruction pairs, with no overlap between its source photographs and benchmark images. Instructions omit edit intensity because strength is carried by the slider parameter.

D FURTHER TRAINING DATASET DETAILS

  • Training uses 300 image-instruction pairs from MIT-Adobe-FiveK, excluding every image used in the benchmark; the backbone’s default edit supplies the full-strength target, so no intermediates are needed.The training-data figure likewise describes the image-instruction pairs as the complete supervision.Source: D FURTHER TRAINING DATASET DETAILS · p. 15, Figure 9 · p. 16
  • A three-stage VLM pipeline describes and screens each photograph, pairs it with a concept seed, then writes and judges an instruction.Seeds cover global styles and tones, object or person attributes, and structural edits such as additions or resizing.Source: D FURTHER TRAINING DATASET DETAILS · p. 15
  • Instructions cannot specify degree, range, percentage, or mechanism because strength is carried entirely by s; their median length is eight words.Source: D FURTHER TRAINING DATASET DETAILS · p. 15
  • The dataset pipeline generated 540 candidates to retain a target of 300 after filtering.Source: D FURTHER TRAINING DATASET DETAILS · p. 15

E FURTHER BENCHMARK DETAILS

The benchmark is designed for continuous edits, which existing method-specific or one-shot benchmarks do not assess consistently. It evaluates six complementary properties using metrics that judge images or trajectories directly rather than comparing them with a nonexistent ground-truth intermediate.

Benchmark scope

  • The benchmark covers 300 examples across global, attribute, and structural edits, with both prompt formats supplied to support comparison across trained and training-free methods.The examples use disjoint, aspect-preserving source photographs; instructions are more specific than the training instructions.Source: E FURTHER BENCHMARK DETAILS · pp. 16–17

Metrics

  • Six properties are assessed: edit fidelity, identity preservation, endpoint consistency, image quality, strength monotonicity, and perceptual uniformity.Because continuous edits have no unique ground-truth trajectory, metrics judge image pairs or trajectories rather than distance to a fixed target.Source: E FURTHER BENCHMARK DETAILS · p. 17

Edit fidelity

  • CLIP directional similarity measures whether edits move in the direction requested by text, but is insensitive to whether they are localized or plausible.It is reported at full strength for comparability and across strength values to assess monotonicity.Source: E FURTHER BENCHMARK DETAILS · p. 17

Other metrics

  • The benchmark also uses VLM judgments of edit success and collateral change, pixel error at zero strength, image-quality measures, and metric-trajectory violation rates.Collateral change is framed as evidence of damage; the positive-form preservation question saturates, while the violation-rate measure uses a dead-band so noise around a flat segment is not counted.Source: E FURTHER BENCHMARK DETAILS · pp. 17–18

Perceptual uniformity

  • Perceptual uniformity is measured with the coefficient of variation of consecutive-frame perceptual distances; the cumulative distance curve should follow a diagonal.Flat segments indicate dead zones, while steep segments indicate jumps; the scale-free coefficient allows comparison across methods with different total edit magnitudes.Source: E FURTHER BENCHMARK DETAILS · p. 18

Metric interpretation

  • Uniformity scores are meaningful only alongside edit fidelity, since a method that barely changes the image can achieve even increments without performing the edit.The benchmark therefore presents all six properties together.Source: E FURTHER BENCHMARK DETAILS · pp. 17–18

F ADDITIONAL RESULTS

Across the benchmark, UniSlider performs consistently near the top, including when results are split by edit category and when additional metrics are considered. It also leads both user-study judgments, for overall preference and perceived uniformity.

Benchmark results

  • UniSlider is best or second-best on every edit category and metric except two, and remains consistently at the top across the benchmark.The per-category comparison includes a SliderEdit implementation on the same FLUX.2-Klein-4B backbone; Table 4 separates edit types and marks same-backbone methods with K.Source: F ADDITIONAL RESULTS · p. 19, Table 4 · p. 20

Additional metrics

  • UniSlider performs best on all additional metrics, including Q-Eval, its violation rate, and the combined VQAedit × (1 − VQAid) score.The combined score rewards requested edits while penalizing collateral changes.Source: F ADDITIONAL RESULTS · p. 19, Table 5 · p. 20

Qualitative comparisons

  • In qualitative comparisons, UniSlider transitions smoothly where Firefly Tune has a late VQAedit jump and color shift, and where other baselines show artifacts or fail to interpolate.UniSlider also stays close to the ideal perceptual-distance diagonal while Firefly Tune and SliderEdit show dead zones and jumps.Source: F ADDITIONAL RESULTS · pp. 19–20

User study

  • UniSlider leads the user study on uniformity and overall preference, with Thurstone Case V scores of +0.88 and +0.82, respectively.The next method scores +0.23 on uniformity; the two judgment scales correlate at r = 0.92. Figure 10 summarizes the pairwise judgments.Source: F ADDITIONAL RESULTS · p. 21, Figure 10 · p. 21

Additional examples

  • Additional examples show UniSlider across stylization, attribute, facial-expression, and structural edits, while spatial masks support localized edits and smooth transitions between edited and preserved regions.The mask can also define arbitrary shapes.Source: F ADDITIONAL RESULTS · p. 21

Composed edits

  • Sequentially composed 2D edits remain approximately uniform along both axes and progress largely independently, although reversing their order is not generally interchangeable.The first instruction spans the horizontal axis and the second the vertical axis; composition requires no additional training.Source: F ADDITIONAL RESULTS · pp. 21–22

G FURTHER ANALYSIS

The authors validate perceptual-distance metrics against human judgments of edit magnitude before using them to assess slider uniformity. DreamSim and TPIPS agree best with those judgments and are effectively tied.

Study design

  • In a study of 15 scenes and 538 pairwise judgments, metric distances were correlated with Bradley–Terry scales using Pearson and Spearman correlations.SRCC measures agreement in ranking edits, while PLCC measures whether metric values scale linearly with perceived change.Source: G FURTHER ANALYSIS · p. 22

Metric validation

  • DreamSim and TPIPS agree best with human judgments of perceived edit magnitude and are effectively tied, followed by LPIPS and CLIP.Pixel-space RMSE performs clearly worse, and color-histogram distance performs worse still.Source: G FURTHER ANALYSIS · p. 22
  • For every tested metric, PLCC and SRCC were within 0.01, indicating that metrics ranking edits like humans also scaled close to linearly with perceived magnitude.The authors use this result to justify perceptual distance as a measure of strength, not merely a ranking signal.Source: G FURTHER ANALYSIS · p. 22

H ABLATION STUDIES

The backbone and LoRA-rank ablations show that slider-specific behavior varies across backbones, while increasing adapter rank yields only marginal uniformity gains. Flux-Klein-4B is selected for its balance of uniformity and efficiency.

Backbone ablation

  • Both Flux-Klein backbones lead on monotonicity and uniformity, while FireRed-1.0 is weakest and fails to return the input at zero strength.The authors attribute FireRed-1.0's weaker results to its eight-step schedule, which makes decoded-pixel optimization harder than with four-step models.Source: H ABLATION STUDIES · p. 22
  • Flux-Klein-4B is selected over 9B because 9B is slightly better on fidelity and image quality but worse on uniformity, while twice as large and slow.The 4B model is the most efficient of the four tested backbones: 8B parameters and 7.2 seconds for 11 frames.Source: H ABLATION STUDIES · p. 23

LoRA rank

  • Across LoRA ranks 4, 8, 16, and 32, fidelity and identity preservation are unchanged, while uniformity improves only marginally with rank.The authors use rank 32; residual non-uniformity on the full dataset is recovered by adaptive sampling rather than additional adapter capacity.Source: H ABLATION STUDIES · p. 23

Additional examples

  • Additional examples illustrate UniSlider outputs for visual styles, attributes, structural edits, spatial masking, and 2D edits.The ablation section’s figures include examples across these edit settings.Source: Figure 13 · p. 25, Figure 14 · p. 26, Figure 15 · p. 27
Loading 2610.06831v1…