Source-linked AI summary
FlowSlider: Training-Free Continuous Image Editing via Fidelity-Steering Decomposition
Taichi Endo, Guoqing Hao, Kazuhiko Sumi
TL;DR
Continuous image editing needs slider control that preserves source fidelity and edit direction, but existing learned approaches require extra supervision and training. FlowSlider decomposes FlowEdit into fidelity and steering terms, scales only steering, and achieves stable, high-quality control across tasks and Rectified Flow backbones without post-training.
Problem
Continuous editing requires smooth, direction-consistent strength control that preserves source identity and structure, while existing learned sliders rely on extra synthetic or proxy supervision.
Method
FlowSlider decomposes FlowEdit’s update into a fixed source-conditioned fidelity term and a steering term scaled by a single strength scalar s.
Results
FlowSlider consistently achieves strong prompt alignment, high fidelity, and improved monotonic and smooth slider behavior across FLUX.1-dev and SD3 without post-training or curated data.
Takeaways & Limitations
Approximately orthogonal fidelity and steering components enable predictable strength changes while preserving source identity and structure across diverse editing tasks.
Takeaways & Limitations
The proposed slider behavior is scoped to training-free editing built on FlowEdit and evaluated with specified Rectified Flow backbones and benchmarks.
Abstract
from arXiv · showhide
Continuous image editing aims to provide slider-style control of edit strength while preserving source-image fidelity and maintaining a consistent edit direction. Existing learning-based slider methods typically rely on auxiliary modules trained with synthetic or proxy supervision. This introduces additional training overhead and couples slider behavior to the training distribution, which can reduce reliability under distribution shifts in edits or domains. We propose \textit{FlowSlider}, a training-free method for continuous editing in Rectified Flow that requires no post-training. \textit{FlowSlider} decomposes FlowEdit's update into (i) a fidelity term, which acts as a source-conditioned stabilizer that preserves identity and structure, and (ii) a steering term that drives semantic transition toward the target edit. Geometric analysis and empirical measurements show that these terms are approximately orthogonal, enabling stable strength control by scaling only the steering term while keeping the fidelity term unchanged. As a result, \textit{FlowSlider} provides smooth and reliable control without post-training, improving continuous editing quality across diverse tasks.
1 Introduction
FlowSlider addresses the need for smooth, fidelity-preserving, direction-consistent edit-strength control without post-training. It decomposes FlowEdit’s update into fixed fidelity stabilization and scalable semantic steering, yielding reliable slider behavior across benchmarks and backbones.
- Continuous editing requires smooth strength control that balances semantic change with source-image fidelity and maintains direction consistency.
- FlowSlider decomposes FlowEdit’s update into a source-conditioned fidelity term and a semantic steering term, scaling only the latter with scalar s.The fidelity term preserves source fidelity as strength increases, while the steering term drives the target transition.
- Strength control is training-free and implemented at inference time, requiring no additional data or post-training.The mechanism is not tied to a specific training distribution.
- Naive FlowEdit scaling introduces artifacts at s = 2, whereas FlowSlider maintains stable fidelity and matches FlowEdit at s = 1.
- The fidelity and steering components are approximately orthogonal, allowing steering-strength changes while largely preserving stabilization.
- FlowSlider achieves the best slider behavior while maintaining superior edit quality on continuous-editing and PIE-Bench evaluations.It also generalizes across FLUX.1-dev and Stable Diffusion 3 Medium without post-training.
2 Related Work
Prior continuous-editing methods commonly learn slider behavior from domain-specific, synthetic, or proxy supervision, while heuristic post-hoc controls can entangle strength with direction and fidelity. FlowSlider instead derives slider control directly from a pretrained Rectified Flow editor without post-training.
- Earlier GAN-based methods learned domain-specific manipulation directions to obtain continuous control.
- Recent slider methods learn strength control from synthetic or proxy supervision, adding training requirements and coupling behavior to the training distribution.
- Kontinuous Kontext trains a lightweight projector that maps scalar strength to modulation offsets using synthesized editing triplets and intermediate images.
- FlowSlider targets slider-style control directly from pretrained Rectified Flow editing dynamics without post-training.
- Heuristic inference-time controls can entangle edit magnitude with direction and source fidelity, producing non-monotonic behavior, semantic drift, or degraded preservation.
- Naively scaling FlowEdit’s velocity update can destabilize the ODE path by amplifying residual non-cancelled components.
3 Method
FlowSlider enables training-free continuous editing by decomposing FlowEdit’s update into fidelity and steering terms, then scaling only steering to control semantic strength. The method preserves source-conditioned stabilization while avoiding the drift and artifacts caused by naive scaling.
- FlowEdit setup: FlowEdit constructs an editing path from noisy source states and a velocity difference between target- and source-conditioned predictions.Rectified Flow uses a learned velocity field and backward ODE integration; FlowEdit avoids ODE inversion for prompt-pair editing.
- Motivation: Naively scaling the full update amplifies semantic drive, source-conditioned stabilization, and residual non-cancelled components, increasing drift and artifacts at large strengths.Residuals accumulate during integration, weaken source coupling, and can produce unrealistic outputs, including ringing or over-sharpening.
- Fidelity–steering decomposition: The exact decomposition writes the FlowEdit update as V∆(t) = Vfid(t) + Vsteer(t), separating source stabilization from prompt-driven semantic steering.Vsteer is a same-state, cross-prompt difference, whereas Vfid is a same-prompt, cross-state difference; the identity introduces no approximation.
- Decomposed strength modulation: FlowSlider controls strength with V∆_s(t) = Vfid(t) + sVsteer(t), leaving the fidelity term fixed while varying semantic intensity.Setting s = 1 recovers FlowEdit, while changing s modulates semantic strength without changing source-conditioned stabilization.
4 Experiments
FlowSlider is evaluated against learning-based and training-free baselines on edit quality and slider behavior. It achieves stable, source-faithful control across continuous and general editing evaluations, while steering-only modulation outperforms alternative ablations.
- Experimental setup: The experiments compare FlowSlider with learning-based sliders, heuristic inference-time controls, and two training-free FlowEdit ablations.Evaluations cover a continuous-editing benchmark, PIE-Bench general editing, and qualitative slider results.
- Evaluation metrics: Evaluation measures intended semantic change, source preservation, monotonicity, and smoothness across increasing strength values.CLIP-dir measures prompt-aligned change, DreamSim measures perceptual source similarity, and slider metrics assess ordering and gradual transitions.
- Qualitative results: FlowSlider achieves stable continuous control while preserving source identity, whereas Kontinuous Kontext and SliderEdit show minimal changes, fidelity loss, or synthetic-looking outputs in qualitative examples.The comparison covers time and expression changes as slider strength increases.
- Quantitative results: FlowSlider attains the best slider behavior and source fidelity simultaneously, with highest monotonicity, lowest smoothness error, lowest DreamSim, and relatively high CLIP-dir.FLUX-Kontext and SliderEdit obtain higher CLIP-dir but have worse source preservation and less stable trajectories.
- Quantitative results: PIE-Bench results show strong FlowSlider performance consistent with the continuous-editing benchmark, despite PIE-Bench not being designed for continuous-strength control.The general-editing evaluation excludes Add and Remove edits but still includes tasks that are not naturally slider-continuous.
- Ablation study: Naive global scaling improves CLIP-dir but severely degrades fidelity and slider behavior, while interpolation smooths trajectories but under-edits; steering-only scaling performs best overall.The ablation attributes FlowSlider’s advantage to scaling semantic steering while keeping the source-conditioned stabilizer fixed.
5 Conclusion
FlowSlider provides training-free slider control for Rectified Flow editing by separating fidelity preservation from semantic steering. Across FLUX.1-dev and SD3, it combines prompt alignment, source fidelity, and improved monotonic and smooth behavior without post-training or curated data.
- Conclusion: FlowSlider decomposes FlowEdit’s update into fidelity and steering terms to control edit strength while preserving source identity and structure.The method scales only the steering term and keeps the source-conditioned fidelity component fixed.
- Conclusion: Across FLUX.1-dev and SD3, FlowSlider achieves strong prompt alignment, high fidelity, and improved monotonic and smooth slider behavior without post-training or curated data.The conclusion attributes predictable strength changes to weak coupling between the decomposed components.
A Evaluation Metric Definitions
The metric definitions establish the source image, edited image, prompt pair, and ordered strength values used for evaluation.
- Notation: The evaluation defines xsrc as the source image, xedit_s as the edited image at strength s, and (csrc, ctar) as the source–target prompt pair.It evaluates K discrete strengths ordered as s1 < s2 < · · · < sK.
A.1 Edit quality metrics
Edit quality is assessed through prompt-direction alignment and perceptual source preservation. CLIP-dir rewards alignment with the intended prompt change, while DreamSim quantifies perceptual distance from the source.
- CLIP-dir: CLIP-dir measures cosine similarity between image-space and text-space edit directions.Higher values indicate better alignment between the edit effect and intended prompt change.
- DreamSim: DreamSim measures perceptual distance between source and edited images in a learned feature space.Lower values indicate that the edited image remains more perceptually faithful to the source.
A.2 Slider behavior metrics
Slider behavior is evaluated by whether increasing strength produces progressively stronger, target-aligned changes and by how smoothly edited images traverse perceptual space.
- Monotonicity: Monotonicity measures the fraction of consecutive strength pairs where both CLIP-T and DreamSim are non-decreasing.Higher scores indicate that larger slider values tend to increase target alignment and distance from the source image together.
- Smoothness: Smoothness evaluates consecutive edited-image triplets using a normalized triangle deficit in perceptual distance.The deficit compares the two adjacent distances with the endpoint distance.
- Smoothness: Lower smoothness values indicate smoother transitions, with zero corresponding to a geodesic trajectory in perceptual space.The metric is defined as the worst-case deficit across triplets.
B Benchmark Specification
The evaluation uses a 250-sample benchmark spanning five categories and 13 editing tasks, with qualitative and trade-off analyses comparing FlowSlider against prior slider methods.
- Benchmark construction: The continuous editing benchmark contains 250 samples from PIE-Bench and Pixabay images annotated with source–target prompt pairs and editing instructions.It includes 31 PIE-Bench samples and 219 Pixabay images.
- Benchmark scope: The benchmark spans five categories and 13 editing tasks.The category breakdown is reported in Table 4, with representative examples shown in Figure 7.
- Evaluation protocol: Qualitative comparisons show original images followed by outputs at increasing slider strengths across all five editing categories.Figures 8–11 provide additional comparisons with Kontinuous Kontext and SliderEdit.
- Evaluation findings: FlowSlider occupies the upper-left trade-off region across all four preservation metrics, combining higher CLIP-T with better source identity and structure preservation.At s=5, it preserves fidelity better than the baselines at their weakest settings.
E Failure Cases
FlowSlider has two task-dependent failure modes: extreme strengths can cause structural degradation, while discrete concept changes may saturate before higher strengths remain meaningful.
- Discrete concept changes: For discrete changes such as “rabbit → cat,” the target concept is fully realized at s=1, so further scaling produces artifacts instead of coherent semantic change.Observed artifacts include color shifts and over-sharpening.
- Task dependence: The effective strength range depends on the task: gradual attribute edits tolerate larger values, whereas discrete concept replacements saturate early.Selecting an appropriate s range is important for reliable results.
F Reverse Editing with Negative Strength
Negative strength reverses FlowSlider’s editing direction, enabling bidirectional control from one prompt pair, although reverse-edit quality depends on the task and is generally less stable.
- Bidirectional control: Setting s < 0 reverses the semantic transition driven by the steering term, enabling reverse edits without retraining.Examples include de-aging and reversing seasonal changes from the same prompt pair.
- Qualitative behavior: Negative strengths produce semantically coherent reverse edits across four diverse tasks, providing a continuous slider from −s to +s without additional supervision.The demonstrated reversals include hair color, saturation, surface age, and facial expression.
- Mechanism: Reverse behavior arises because the prompt-pair steering vector approximately aligns with a meaningful semantic axis in the pretrained model’s latent space.Negating the vector traverses the opposite direction along that axis.
- Caveat: Reverse editing is task-dependent and tends to be less stable when the implicit semantic opposite is ambiguous or poorly defined.The reverse direction is not explicitly specified by the prompt pair.