Source-linked AI summary
SliderEdit: Continuous Image Editing with Fine-Grained Instruction Control
Arman Zarei, Samyadeep Basu, Mobina Pournemat, Sayan Nag, Ryan Rossi, Soheil Feizi
TL;DR
Instruction-based image editors apply each instruction at a fixed strength, limiting fine-grained control over individual edits. SliderEdit trains low-rank adapters to provide continuous instruction-specific sliders, and the paper reports improved controllability, visual coherence, and flexibility. It supports continuous, compositional instruction-driven editing, though some attribute entanglement remains inherent in the base model.
Problem
Instruction-based image editors apply edits at fixed strengths, limiting continuous adjustment of the intensity of individual instructions.
Source: §1 · p. 2Method
SliderEdit trains low-rank adapters with a suppression objective and scales their updates to provide continuous control over individual instructions.
Source: §1 · p. 2, §3.3 · p. 5Results
SliderEdit reports superior controllability, visual coherence, and flexibility when integrated with instruction-based editing models.
Source: §5 · p. 8Takeaways & Limitations
The framework lays a foundation for interactive instruction-driven editing with continuous and compositional control.
Source: §5 · p. 8Some attribute entanglement persists across models, including the base model, due to inherent coupling in the generative model.
Source: §8.3.1 · p. 15
Abstract
from arXiv · showhide
Instruction-based image editing models have recently achieved impressive performance, enabling complex edits to an input image from a multi-instruction prompt. However, these models apply each instruction in the prompt with a fixed strength, limiting the user's ability to precisely and continuously control the intensity of individual edits. We introduce SliderEdit, a framework for continuous image editing with fine-grained, interpretable instruction control. Given a multi-part edit instruction, SliderEdit disentangles the individual instructions and exposes each as a globally trained slider, allowing smooth adjustment of its strength. Unlike prior works that introduced slider-based attribute controls in text-to-image generation, typically requiring separate training or fine-tuning for each attribute or concept, our method learns a single set of low-rank adaptation matrices that generalize across diverse edits, attributes, and compositional instructions. This enables continuous interpolation along individual edit dimensions while preserving both spatial locality and global semantic consistency. We apply SliderEdit to state-of-the-art image editing models, including FLUX-Kontext and Qwen-Image-Edit, and observe substantial improvements in edit controllability, visual consistency, and user steerability. To the best of our knowledge, we are the first to explore and propose a framework for continuous, fine-grained instruction control in instruction-based image editing models. Our results pave the way for interactive, instruction-driven image manipulation with continuous and compositional control.
1. Introduction
Instruction-based editors produce capable local and global changes, but apply each instruction at a fixed strength, limiting precise, continuous control. SliderEdit assigns each instruction a slider and trains low-rank adapters with a suppression objective, reporting improved controllability and semantic disentanglement.
1. Introduction
- Fixed-strength, all-or-nothing edits prevent systematic adjustment of individual instruction intensity, limiting user flexibility and interpretability.The paper frames continuous control as important for interactive editing.Source: §1 · p. 2
- SliderEdit gives each instruction its own slider for smooth adjustment from suppression through full application to amplification, without per-instruction fine-tuning.The framework targets continuous, disentangled, interpretable control of multi-instruction prompts.Source: §1 · p. 2
- Low-rank adapters act on instruction-relevant token embeddings and learn suppression by matching full-prompt output to output with the target instruction removed.Scaling the learned adapter weights then yields continuous adjustment of that instruction's influence.Source: §1 · p. 2
- SliderEdit integrates with FLUX-Kontext and Qwen-Image-Edit and reports superior edit controllability and semantic disentanglement across quantitative and qualitative evaluations.The introduction also describes control across subtle attributes, style refinements, and complex multi-object scenes.Source: §1 · p. 2
- The stated contributions are continuous instruction-level modulation, PPS-trained adapters, and integration with foundation editing models with improvements in consistency and controllability.The paper describes this as the first framework for continuous instruction-based image editing.Source: §1 · p. 2
2. Related Works
Prior editing work progressed from diffusion-based steering to natural-language instruction editing, while continuous sliders largely targeted generation rather than real-image editing. SliderEdit extends slider-based control to instruction editing and aims to avoid per-attribute retraining while supporting simultaneous edits.
2.1. Image Editing
- Instruction-based foundation editors support local and global changes, but lack continuous control over the strength of individual edits.Earlier editing methods steered diffusion trajectories through optimization or conditioning while preserving image fidelity.Source: §2.1 · p. 2
2.2. Continuous Attribute Slider
- Prior continuous-control methods often train per-attribute LoRAs or embedding directions for text-to-image generation, with entanglement, multi-edit degradation, or limited real-image applicability.The cited approaches include Concept Sliders and Baumann et al..Source: §2.2 · p. 3
- SliderEdit generalizes slider control to instruction-based image editing, avoiding per-attribute retraining and supporting simultaneous edits and unseen attributes.The passage also reports robustness across editing scenarios and better performance on real-image editing tasks.Source: §2.2 · p. 3
3. SliderEdit: Continuous Image Editing
SliderEdit assigns independently adjustable strengths to instructions in a multi-instruction prompt. The method builds on localized instruction-token effects in MMDiT representations, then trains selective adapters to suppress a target instruction and scales their updates for continuous control.
3. SliderEdit: Continuous Image Editing
- For each instruction P_i in a multi-instruction prompt, SliderEdit defines a strength β_i that suppresses, fully applies, or exaggerates that edit.β_i = 0 suppresses the instruction, β_i = 1 fully applies it, and β_i > 1 exaggerates it.Source: §3 · p. 3
3.1. MMDiT Architecture
- MMDiT jointly processes image latents and prompt-derived text embeddings through shared attention layers, enabling cross-modal information exchange.The text tokens are produced by encoding the prompt with a pretrained language model such as T5.Source: §3.1 · p. 3
3.2. Instruction-Level Interpretability Analysis
- Target-instruction token interventions show localized effects and intermediate edit strengths, but direct embedding interpolation offers only limited, discontinuous modulation.The analyzed embeddings correspond to the target instruction, and interpolation replaces them toward padding-token embeddings.Source: §3.2 · pp. 3–4, Figure 2 · p. 3
3.3. Fine-Grained Control of Edit Instructions
- PPS trains an adapter so the full-prompt adapted model matches the frozen model's denoising direction when the target instruction is removed.STLoRA applies low-rank updates only to target-instruction tokens, while leaving other text tokens and image tokens unmodified.Source: §3.3 · pp. 4–5
- Scaling trained STLoRA updates provides a smooth continuum from suppression through full application to exaggerated edits.The adapter scale α and instruction scale β are related by α = 1 − β.Source: §3.3 · p. 5
- GSTLoRA applies updates across image and text tokens for single-instruction edits, leveraging global context for stronger control and often higher-fidelity edits.STLoRA is described as handling both single- and multi-instruction prompts through selective token modulation.Source: §3.3 · p. 5
4. Experiments
SliderEdit is evaluated for smooth, fine-grained control across local and global edits, multi-instruction scenarios, and quantitative measures of continuity, extrapolation, and disentanglement. Results report strong single-instruction continuity and different strengths for FLUX and Qwen across multi-instruction metrics.
4.1. Implementation details
- Training uses lightweight rank-16 adapters and 1k–8k GPT-Image-Edit samples; STLoRA trains for 1,000 iterations and GSTLoRA on FLUX-Kontext for 300.STLoRA models converged around 400 iterations, and the training process is described as computationally lightweight and data-efficient.Source: §4.1 · p. 5
4.2. Qualitative Results
- GSTLoRA produces smooth, continuous trajectories for local edits such as makeup and car age, and global edits such as season and lighting.STLoRA also provides smooth transitions across a 2D space of two edit directions, with additional examples spanning three directions.Source: §4.2 · p. 5, Figure 4 · p. 6, Figure 6 · p. 7
- With Qwen-Image-Edit, STLoRA supports zero-shot multi-subject personalization followed by scene editing, producing coherent image sequences that evolve with slider adjustments.The examples are presented as having potential for story-like visual editing.Source: §4.2 · p. 6, Figure 5 · p. 7
4.3.1. Evaluation Set
- The facial-editing benchmark varies the number of instructions and each instruction’s strength across a structured space of images.It uses subjects with diverse genders, ages, and ethnicities, and starts from images where the target attributes are absent.Source: §4.3.1 · p. 6
4.3.2. Metrics
- Evaluation measures edit execution with vision-language image–text similarity, extrapolation with maximum similarity, continuity with a chi-squared uniformity test, and disentanglement through identity and perceptual distances.Identity is measured in ArcFace space; perceptual distances use LPIPS and DINOv2.Source: §4.3.2 · p. 7
4.3.3. Baselines
- Single-instruction baselines include Explicit and Implicit CFG, Concept-Slider, and Continuous Attribute Control; for multiple instructions, STLoRA is compared with CFG variants.The CFG baselines cannot independently control multiple edit directions, while STLoRA enables per-instruction control.Source: §4.3.3 · p. 7
4.3.4. Results
- GSTLoRA achieves the highest continuity among single-instruction baselines while maintaining strong disentanglement and satisfactory extrapolation, and both SliderEdit variants significantly outperform Explicit CFG.Figure 7 also shows gradual similarity changes and lower identity drift for GSTLoRA than the CFG variants.Source: Table 1 · p. 8, §4.3.4 · p. 8, Figure 7 · p. 8
- In multi-instruction settings, FLUX performs better on continuity for single edits and on identity preservation, while Qwen performs better on continuity for two- and three-instruction edits and on extrapolation.The results report a trade-off among continuity, extrapolation, and disentanglement across configurations.Source: Table 2 · p. 8, §4.3.4 · p. 8
5. Conclusion
SliderEdit uses lightweight low-rank adapters and a novel loss to provide continuous, fine-grained instruction control in image-editing models. It achieves superior controllability, visual coherence, and flexibility with FLUX-Kontext and Qwen-Image-Edit.
5. Conclusion
- SliderEdit trains lightweight low-rank adapters with a novel loss to disentangle and modulate instruction effects, achieving continuous control in FLUX-Kontext and Qwen-Image-Edit.The conclusion describes the framework as supporting interactive editing with continuous and compositional control.Source: §5 · p. 8
Supplementary Material
SPPS applies the PPS suppression objective to an entire edit prompt treated as one instruction. Its training includes a neutralized-image forward pass and is described as generalizing to multi-instruction editing.
Supplementary Material
- SPPS treats the full edit prompt as one instruction and adds a forward pass with a neutralized image to suppress undesired edit effects.The neutralized image uses either an empty prompt or a neutral instruction such as “keep the image the same.”Source: Figure 8 · p. 11
6. Related Works
The section outlines stochastic diffusion models and deterministic flow matching as generative approaches that transform noise toward data. It contrasts their sampling trajectories and training or inference characteristics.
6.1. Diffusion Models and Flow Matching
- Diffusion models gradually corrupt data with noise toward a Gaussian distribution, then learn a reverse process that reconstructs the data from noise.The reverse-time SDE uses the score function to guide noisy samples toward the data distribution.Source: §6.1 · p. 11
- Flow matching learns a deterministic vector field and maps an initial noise distribution to the target data distribution by integrating an ODE.Training aligns the learned field with a predefined target vector field.Source: §6.1 · pp. 11–12
- Unlike stochastic SDE trajectories in diffusion, flow matching uses deterministic ODE trajectories, eliminating sampling stochasticity and generally enabling faster, more efficient training and inference.The section characterizes flow matching as a computationally efficient deterministic counterpart to diffusion models.Source: §6.1 · p. 12
7. SliderEdit: Continuous Image Editing
The section presents SPPS as a streamlined alternative to PPS that trains on single-instruction examples while retaining compositional generalization. PPS provides finer supervision for complex multi-instruction control.
7.1. Simplified Partial Prompt Suppression Loss
- SPPS trains on a single instruction by matching the adapted model's denoising prediction to the frozen base model's prediction with the instruction removed or replaced by a null instruction.This is intended to neutralize the edit induced by that instruction and isolate its representation.Source: §7.1 · p. 12
- SPPS avoids parsing multi-instruction prompts or finding token boundaries, so it can use general editing datasets that contain only single-instruction pairs.Its adapters also generalize compositionally to multi-instruction edits at inference.Source: §7.1 · p. 12
- PPS gives finer-grained supervision and better disentangled, localized adaptations for complex multi-instruction control, while SPPS is simpler and achieves comparable results in most cases.Both objectives support smooth, continuous interpolation along edit directions.Source: §7.1 · p. 12
8. Experiments
The experiments describe the training setup, evaluation criteria, and baselines, then report qualitative evidence of continuous and disentangled editing across image and text tasks. They also compare PPS with SPPS and SliderEdit with other approaches, while noting persistent attribute entanglement.
8.1. Implementation Details
- Training uses FLUX-Kontext and Qwen-Image-Edit, with 1k–8k GPT-Image-Edit-1.5M samples and lightweight LoRA adaptation; STLoRA runs 1,000 iterations and GSTLoRA runs 300 iterations on FLUX-Kontext.The reported setup uses one NVIDIA H100-SXM GPU; the authors defer comprehensive study of block and timestep efficiency choices.Source: §8.1 · p. 12
8.2.1. Metrics
- Continuity is assessed by how smoothly and uniformly similarity scores vary as edit strength increases, using a chi-squared statistic; disentanglement is assessed with identity and perceptual or semantic feature distances.The metrics include ArcFace identity distance, LPIPS with AlexNet and VGG backbones, and DINOv2 feature distance.Source: §8.2.1 · pp. 12–13
8.2.2. Baselines
- For multiple instructions, Explicit CFG, Implicit CFG, and GSTLoRA cannot independently control edit directions, whereas STLoRA supports per-instruction control; Concept-Slider and Continuous Attribute Control are omitted from this setting after limited single-instruction effectiveness.For single-instruction comparisons, the baselines also include Explicit and Implicit CFG on FLUX-Kontext, Concept-Slider, and Continuous Attribute Control.Source: §8.2.2 · pp. 13–14
8.3. Qualitative Results
- Qualitative examples show smooth, coherent transitions for local and global edits, continuous face and text attribute changes, and disentangled multi-instruction control across two- and three-instruction edits.The multi-instruction examples include separate instruction axes and text attributes such as font style and color.Source: §8.3 · pp. 14–15
8.3.1. PPS vs SPPS
- PPS yields a more disentangled latent space and finer, more independent instruction control than SPPS, while SPPS is simpler and comparable in most cases.Some attribute entanglement remains across all models, including the base model; the paper attributes correlated changes to inherent coupling in the generative model.Source: §8.3.1 · p. 15
8.3.2. Comparison with other baselines
- SliderEdit and Explicit Guidance produce smooth, coherent, faithful edits, while Concept-Slider and Continuous Attribute Control often fail to preserve image fidelity or follow the target modification on real-image editing tasks.The paper links the latter methods' poor real-image-editing performance to their indirect adaptation from text-to-image generation.Source: §8.3.2 · p. 15