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

cs.CVOct 9, 2026

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

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

Organizations: Adobe Research · University of Maryland

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

Figures & tables

Panel of Figure 3 — Figure 3: Slider Discontinuity Analysis . Baseline image-editing sliders often exhibit abrupt transitions during interaction. The semantic change distance on the right quantifies this instability. In contrast, SliderEdit++ produces smoother sliders by addressing the issue during training.

Figure 4: Overall pipeline of SliderEdit++ for improving the continuity in image editing sliders.

Figure 5: SliderEdit++ displays strong continuity across different transition levels of the slider.

Figure 6: Qualitative comparison with baselines and the effect of our proposed regularizers. (a) demonstrates that SliderEdit++ produces smoother trajectories than competing baselines, while (b) illustrates the impact of our regularizers encouraging significantly smoother semantic transitions.

	Monotonic Progression			Frame Uniformity × VLM Edit Strength					Overall
	GPT-5	Opus-4.8	Avg	LPIPS	CLIP	SigLIP	DINOv2	Avg	Score
KontinuousKontext	0.7073	0.6992	0.7033	0.6223	0.6626	0.6630	0.6585	0.6516	0.6774
TextSteer	0.4980	0.5120	0.5050	0.6020	0.6320	0.6510	0.6530	0.6345	0.5690
SDE-Edit	0.5363	0.5820	0.5591	0.6528	0.6630	0.6634	0.6590	0.6596	0.6094
VeloEdit	0.6837	0.7098	0.6967	0.6381	0.6845	0.6868	0.6816	0.6727	0.6847
Explicit CFG	0.6381	0.6406	0.6393	0.6400	0.6795	0.6810	0.6746	0.6688	0.6541

Table 1: Comparison on monotonic progression and frame uniformity. Best and second-best results are shown in bold and underline ; the final row reports our relative gain over the strongest baseline.

Method	$\mathcal{L}_{\mathrm{sem}}$	$\mathcal{L}_{\mathrm{vel}}$	$\mathcal{L}_{\mathrm{spec}}$	$MP_{avg}$	$\bar{U}_{avg}$	Overall
Baseline	✗	✗	✗	0.7161	0.7258	0.7209
+ $\mathcal{L}_{\mathrm{sem}}$	✓	✗	✗	0.7643	0.8060	0.7851
+ $\mathcal{L}_{\mathrm{vel}}$	✗	✓	✗	0.7552	0.7901	0.7726
+ $\mathcal{L}_{\mathrm{spec}}$	✗	✗	✓	0.7421	0.7964	0.7692
+ All Regularizers	✓	✓	✓	0.7845	0.8386	0.8116

Table 2: Effect of each regularizer on SliderEdit.

\mathcal{L}_{\mathrm{sem}}

: DINO-based semantic midpoint regularizer,

\mathcal{L}_{\mathrm{vel}}

: latent velocity midpoint regularizer,

\mathcal{L}_{\mathrm{spec}}

: spectral norm regularizer on the LoRA matrices.

	Monotonic Progression			Frame Uniformity × VLM Edit Strength					Disentanglement				Overall
	GPT-5	Opus-4.8	Avg	LPIPS	CLIP	SigLIP	DINOv2	Avg	CLIP	SigLIP	BLIP	Avg	Score
SliderEdit	0.6511	0.6687	0.6599	0.7294	0.7638	0.7708	0.7740	0.7595	0.9930	0.9939	0.9694	0.9854	0.7097
SliderEdit++	0.7810	0.7918	0.7864	0.9683	0.9747	0.9782	0.9818	0.9757	0.9932	0.9957	0.9719	0.9869	0.8811

Table 3: Comparison of multi-slider baselines on monotonic progression, VLM-weighted frame uniformity, and disentanglement.

Figure 9: Qualitative results of SliderEdit++ showing strong disentanglement in multi-slider setting.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

	Monotonic Progression			Frame Uniformity × VLM Edit Strength					Overall
	GPT-5	Opus-4.8	Avg	LPIPS	CLIP	SigLIP	DINOv2	Avg	Score
$T=8$	0.7589	0.8000	0.7794	0.7977	0.8130	0.8146	0.8104	0.8089	0.7942
$T=4$	0.7629	0.7950	0.7789	0.8141	0.8275	0.8287	0.8244	0.8237	0.8013
$T=2$	0.7398	0.7845	0.7622	0.8291	0.8396	0.8411	0.8378	0.8369	0.7995
$T=1$	0.6937	0.7639	0.7288	0.8304	0.8369	0.8387	0.8361	0.8355	0.7822

Appendix

Table 4: Ablating the number of generation steps for slider quality. Results show that using fewer generation steps has minimal impact on slider performance while significantly improving generation speed.

Figure 10: At similar data-budgets, SliderEdit on a stronger base model such as Klein-4B outperforms Kontinuous-Kontext on the same Klein-4B model.

Panel of Figure 11 — Figure 11: SliderEdit++ displays strong continuity across different transition levels of the slider.

Panel of Figure 12 — Figure 12: Qualitative comparison demonstrating the effect of our regularizer relative to the baseline SliderEdit method under the same backbone.

Panel of Figure 13 — Figure 13: Qualitative comparison demonstrating the effect of our regularizer relative to the baseline SliderEdit method under the same backbone.

Panel of Figure 14 — Figure 14: Qualitative comparison of continuous edit-control methods. SliderEdit++ demonstrates superior performance.

Panel of Figure 15 — Figure 15: Qualitative Results on Multi-Slider Outputs. Best viewed zoomed in.

Figure 16: Qualitative Results on Design Images.

Panel of Figure 17 — Figure 17: Qualitative Results on Mask-Conditioned SliderEdit++.

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 302 papers

UniSlider: Perceptually Uniform Sliders for Continuous Image Editing
Oct 5, 2026David Serrano-Lozano, Duygu Ceylan, Yannick Hold-Geoffroy +3Image Editing Evaluation Diffusion-Based Image Editing
Computer Vision Center · Universitat Autònoma de Barcelona · Adobe Research
FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition
Apr 2, 2026Taichi Endo, Guoqing Hao, Kazuhiko SumiTraining-Free Image Editing Image Editing
Aoyama Gakuin University
Measured Sliders: Learning Continuous Controls from Differentiable Image Measurements
Sep 4, 2026Yijia Chen, Boyu Wei, Xuanhua YinControllable Image Generation Diffusion-Based Image Generation
School of Computer Science, The University of Sydney

Oct 5, 2026J/K move · Enter open · S save

UniSlider: Perceptually Uniform Sliders for Continuous Image Editing

David Serrano-Lozano, Duygu Ceylan, Yannick Hold-Geoffroy, Iliyan Georgiev+2

Computer Vision Center · Universitat Autònoma de Barcelona · Adobe Research

Image Editing Evaluation Diffusion-Based Image Editing Image Editing

Open paper AskPDF arXiv

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.

Abstract

Image Editing

Explore similar work

UniSlider: Perceptually Uniform Sliders for Continuous Image Editing

FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition

Measured Sliders: Learning Continuous Controls from Differentiable Image Measurements