PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Arman Zarei, Sayan Nag, Bryan Wang, Franck Dernoncourt, Varun Manjunatha, Nedim Lipka, Soheil Feizi, Samyadeep Basu

arXiv:2610.14253v2cs.CVcs.GR

TL;DR

Existing image-editing sliders can produce abrupt or uneven changes and interfere with one another. The paper introduces SlideBench and trains SliderEdit++ with trajectory and disentanglement regularizers. It reports an 18.5% overall-score improvement over the strongest baseline and approximately 3–20× faster operation.

  • Problem

    Existing sliders can produce abrupt semantic transitions, while shared LoRA weights and cross-attention mixing can cause one instruction to shift another attribute.

    Source: §4 · p. 4, §4.3 · p. 6

  • Method

    SliderEdit++ regularizes slider trajectories during training and penalizes overlap between the velocity shifts induced by distinct instructions.

    Source: §1 · p. 2, §4.3 · p. 6

  • Results

    18.5% overall-score improvement over the strongest baseline, with SliderEdit++ operating approximately 3× to 20× faster.

    Source: §6 · p. 9

  • Takeaways & Limitations

    SliderEdit++ improves smoothness, uniform progression, and disentanglement, supporting more reliable and responsive continuous image editing.

    Source: §6 · p. 9

    Performance remains constrained by the underlying editing model; Camera has lower absolute performance across all methods.

    Source: E LIMITATIONS · p. 14

Abstract

from arXiv · show

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 3x to 20x faster than existing slider-based approaches, enabling substantially more responsive interactive editing.

1 INTRODUCTION

Slider-based editing offers gradual control over instruction strength, but existing methods can produce uneven perceptual changes and cross-slider interference. The paper introduces SlideBench and SliderEdit++, reporting stronger benchmark performance and faster inference.

1 INTRODUCTION

  • Existing slider evaluations are fragmented, and common metrics often miss the fine-grained perceptual behavior needed to assess continuous control.The paper introduces SlideBench, spanning diverse editing categories and fine-grained control axes, with metrics for monotonic progression, perceptual continuity, and disentanglement.Source: §1 · p. 1
  • Slider-based editing can yield highly uneven perceptual changes, including abrupt visual jumps, while adjusting one control can affect others.The paper motivates its approach with these discontinuities and multi-slider interference.Source: §1 · pp. 1–2
  • SliderEdit++ regularizes intermediate slider trajectories and encourages different controls to act along orthogonal edit directions.It builds on SliderEdit and uses step-distilled editing backbones for efficient inference.Source: §1 · p. 2
  • 18.5% overall-score improvement over the strongest baseline, with SliderEdit++ operating approximately 3–20× faster.The paper presents these as its headline SlideBench results.Source: §1 · p. 2

2 RELATED WORKS

Prior work explores continuous semantic control through latent directions, adapters, and interventions in text or velocity spaces. More recent methods extend slider control to natural-language image-editing instructions.

2 RELATED WORKS

  • Earlier work studied disentangled representations and meaningful latent directions for continuously manipulating visual attributes.Later approaches use attribute-specific LoRA directions, text-embedding directions, or semantically interpretable representations in rectified-flow transformers.Source: §2 · p. 2
  • Instruction-based sliders modulate the strength of arbitrary edits through learned adapters, scalar conditioning, velocity-field interventions, or text-embedding interpolation.The section also notes that SDEdit can trade off source preservation and edit strength without dedicated slider training.Source: §2 · p. 2
  • SliderEdit++ targets smoothness, consistency, and independence across controls in the instruction-based editing setting.This positions the method as addressing limitations of existing sliders.Source: §2 · p. 2

3 SLIDEBENCH EVALUATION BENCHMARK

SlideBench is designed to evaluate continuous image editing across varied attributes and control behaviors. Its dataset covers diverse editing contexts, and its metrics assess progression, continuity, and multi-slider disentanglement.

3 SLIDEBENCH EVALUATION BENCHMARK

  • SlideBench is introduced as a benchmark for evaluating continuous control in image editing with metrics for fine-grained editing behavior.Figure 2 compares categories spanning diverse continuous editing scenarios.Source: §3 · p. 2, Figure 2 · p. 3

3.1 DATASET DESCRIPTION

  • The dataset spans seven categories and 57 fine-grained subcategories, with five samples per subcategory.Each sample pairs a source image and an instruction defining a transformation between semantic endpoints; the same editing axis appears across visual contexts to test generalization.Source: §3.1 · pp. 2–3
  • The benchmark includes naturally gradual attributes, content-strength changes, and fine-grained character edits such as facial attributes.These categories support comparison across methods; Figure 2 showcases the benchmark categories.Source: §3.1 · p. 3, Figure 2 · p. 3

3.2 EVALUATION METRICS

  • The evaluation measures monotonic progression, continuity, and, for multi-slider models, whether changing one slider preserves other attributes.Monotonic progression averages VLM judgments of whether later consecutive frames show stronger edits; pairwise judgments are used to capture fine-grained perceptual differences.Source: §3.2 · p. 3
  • Continuity evaluates the uniformity of consecutive-frame changes in feature space, with the score weighted by overall edit strength to avoid rewarding no-op sliders.The final continuity score averages over feature representations.Source: §3.2 · p. 3
  • Multi-slider disentanglement measures how much non-varying attributes shift while one slider changes; lower deviation indicates better disentanglement.The measure averages alignment-score deviations across slices and slider pairs.Source: §3.2 · p. 3

4 METHOD: IMPROVING CONTINUITY AND DISENTANGLEMENT IN SLIDERS

SliderEdit++ regularizes intermediate slider trajectories to make semantic changes smoother and more predictable, while reducing interference between independently controlled edits. Its method adds velocity-, semantic-, and spectral-space constraints to the SliderEdit baseline and an orthogonality objective for multi-slider editing.

4.1 PRELIMINARIES: CONTINUOUS EDIT SLIDER USING SLIDEREDIT

  • SliderEdit scales low-rank adapters by α: α=0 leaves the instruction-following edit untouched, α=1 is trained to suppress it, and intermediate controls come from weight scaling without direct supervision.Source: §4.1 · pp. 4–5

4.2 TOWARDS SOLVING CONTINUITY ISSUES IN SLIDERS

  • The velocity-space midpoint regularizer trains predictions at sampled slider midpoints to match the average of neighboring predictions, encouraging equal slider steps to yield predictable output steps.This addresses the mismatch between linear LoRA-weight scaling and potentially nonlinear output changes.Source: §4.2 · p. 5
  • The semantic feature-space regularizer applies midpoint consistency to decoded image features, targeting perceptual jumps that velocity-space smoothness may miss.It uses a frozen DINOv2 feature encoder on decoded one-step clean-image estimates.Source: §4.2 · p. 5
  • The spectral-norm penalty constrains each layer’s effective adapter update, bounding sensitivity to α and allowing training to reach full convergence instead of relying on early stopping.The method targets large operator norms that can turn small slider changes into disproportionate output changes.Source: §4.2 · pp. 5–6

4.3 TOWARDS SOLVING DISENTANGLEMENT ISSUES IN MULTI-SLIDERS

  • For multi-instruction editing, SliderEdit++ penalizes overlap between distinct sliders’ velocity shifts to promote orthogonal effects and reduce attribute leakage from shared adapter weights.The framework builds on token-level gating, but shared LoRA weights and cross-attention mixing can still couple instructions.Source: §4.3 · p. 6

5 EXPERIMENTS

Experiments evaluate SliderEdit++ on SlideBench through comparisons, ablations, multi-slider tests, and mask-conditioned editing. Results show improved single- and multi-slider control, with an extension for localized edits.

5.1 SETTING

  • Evaluations use a FLUX-Klein 4B backbone, eight denoising steps, and 285 benchmark samples spanning seven categories.Slider positions are rendered at eight uniformly spaced values; the study includes quantitative and qualitative comparisons, speed-controllability analysis, and a human study.Source: §5.1 · p. 6, §5 · p. 6

5.2 SINGLE SLIDER

  • Each regularizer improves slider quality over the unregularized baseline, while combining all three yields the strongest performance.The ablation reports complementary effects; qualitative results show full-model edits emerging progressively rather than abruptly at isolated slider positions.Source: §5.2 · p. 7, Table 2 · p. 8
  • 18.5% relative improvement over the strongest prior method, with the best performance in every SlideBench category and a 3× to 20× speed advantage.The human study also finds SliderEdit++ preferred over baselines; Figure 6(a) shows smoother, more gradual trajectories than competing methods.Source: §5.2 · p. 7, Table 1 · p. 8, Figure 6 · p. 8

5.3 MULTI SLIDER

  • In multi-slider evaluation, SliderEdit++ raises the overall score from 0.7097 to 0.8811 and reduces cross-slider interference relative to SliderEdit.Figure 9 illustrates smoother, more disentangled controls, with changes to one slider having minimal effect on others.Source: §5.3 · p. 8, Table 3 · p. 9, Figure 9 · p. 9

5.4 EXTENDING TO MASK-CONDITIONED SLIDERS

  • Mask-conditioned editing confines continuous changes to a user-specified region while preserving surrounding content.The method replaces out-of-mask latents with a re-noised source copy during the final two denoising steps; Figure 7 shows a qualitative example.Source: §5.4 · pp. 8–9, Figure 7 · p. 9

6 CONCLUSION

The paper introduces SlideBench and SliderEdit++, reporting stronger benchmark performance and faster editing. For multi-slider editing, Selective-Token LoRA provides independently addressable controls through shared weights.

6 CONCLUSION

  • SliderEdit++ improves smoothness, uniform progression, and disentanglement, achieving an 18.5% overall-score gain over the strongest baseline while operating approximately 3× to 20× faster.The conclusion frames these results as enabling more reliable and responsive continuous image editing.Source: §6 · p. 9

B.1 CONTINUOUS MULTI-SLIDER EDIT USING SLIDEREDIT

  • Selective-Token LoRA uses instruction-token spans to apply each slider’s adapter contribution only to its corresponding instruction tokens.The mechanism supports K independently addressable sliders with one shared set of weights in a single forward pass.Source: §B.1 · p. 12
  • Restricting each instruction’s LoRA activation to its own tokens provides a structural bias toward disentanglement, while retaining the single-slider continuity treatment.Source: §B.1 · p. 12

C.1 DATASET

The dataset spans diverse continuous editing attributes and visual contexts. Its categories are designed to test gradual variation, progressive content changes, and fine-grained character control.

C.1 DATASET

  • The dataset covers seven categories, 57 fine-grained subcategories, and five samples per subcategory, with each subcategory defining a continuous editing axis.Source: §C.1 · p. 13
  • Each sample is a source caption, source image, and editing instruction defining a transformation toward the opposite endpoint of an editing axis.Source: §C.1 · p. 13
  • Category design covers gradual attributes, edits that progressively introduce visual content, and fine-grained character changes where small variations can matter perceptually.Examples include lighting and camera variation, crowd density or aging, and facial attributes.Source: §C.1 · p. 13

C.2 METRICS

The section evaluates slider trajectories through monotonic progression, frame uniformity, edit strength, and disentanglement. It uses vision-language model assessments where text-image similarity models were too insensitive to fine-grained changes, and notes limits in feature-based disentanglement measures.

C.2 METRICS

  • Text-image similarity models produced noisy or inconsistent scores for small attribute-strength changes, so the authors use stronger vision-language models for monotonic-progression evaluation.They report more consistent assessments from GPT-5 and Claude Opus 4.8 than from LLaVA and Qwen.Source: §C.2 · p. 13
  • Frame uniformity is assessed from consecutive-frame perceptual and semantic changes using LPIPS, CLIP, SigLIP, and DINOv2, while edit strength uses an aggregated VLM-based evaluator.The aggregation is intended to provide a more robust assessment.Source: §C.2 · p. 13
  • CLIP, SigLIP, and BLIP feature spaces provide only limited proxies for disentanglement, so the authors complement them with human evaluation.The authors identify more robust feature spaces for fine-grained editing effects as future work.Source: §C.2 · p. 13

D.1 RESULTS

Qualitative results show that SliderEdit++ regularizers improve semantic continuity over the unregularized objective and support smooth control across single- and multi-slider examples. Fewer generation steps retain strong slider performance while reducing inference latency.

D.1 RESULTS

  • The regularizers yield substantially more semantically continuous trajectories than the base objective, which often has abrupt transitions between neighboring slider values.The examples span global edits and fine-grained local changes.Source: §D.1 · p. 13
  • Figures 14–17 show smoother comparisons with existing methods, largely disentangled multi-slider controls, text-rich design edits, and mask-conditioned localized editing.The design examples preserve underlying content and layout, while the mask-conditioned extension edits within a user-selected region.Source: §D.1 · p. 14

D.2 ABLATING ON FEW-STEP GENERATION

  • Up to 8× lower inference latency is suggested by reducing generation steps from the default N = 8 to N = 4, 2, or 1 while maintaining strong performance.Table 4 reports minimal impact on slider performance with fewer steps and significantly improved generation speed.Source: §D.2 · p. 14, Table 4 · p. 14

D.3 ABLATING THE UNDERLYING SLIDER TRAINING OBJECTIVE

At comparable data budgets on the few-step Klein-4B backbone, SliderEdit outperforms Kontinuous-Kontext and avoids paired training data.

D.3 ABLATING THE UNDERLYING SLIDER TRAINING OBJECTIVE

  • On the few-step Klein-4B backbone, SliderEdit outperforms Kontinuous-Kontext at comparable data budgets while requiring ≤10k unpaired samples rather than approximately 110k paired samples.The authors adopt SliderEdit as the base objective because it offers stronger performance without paired training data.Source: §D.3 · p. 14

E LIMITATIONS

SliderEdit++'s performance remains constrained by the capabilities of its underlying image-editing model, with challenging categories such as Camera showing lower absolute performance across methods.

E LIMITATIONS

  • SliderEdit++ performance remains constrained by the underlying image-editing model, and Camera has lower absolute performance across all methods.The passage attributes some limitations in challenging categories to the base model's editing ability rather than slider control itself.Source: E LIMITATIONS · p. 14
  • Figure 11 presents strong continuity across different slider transition levels.Source: Figure 11 · p. 15
  • Figure 14 compares continuous edit-control methods and identifies SliderEdit++ as demonstrating superior performance.Source: Figure 14 · p. 18

Related papers

Loading 2610.14253v2…