acceptodds
Withdrawn from ICLR 2027

RewardSlider: Continuous Image Editing via Reward-Guided Trajectory Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing instruction-based image editing models typically produce fully edited outputs, making it difficult to continuously and finely control the degree of editing. We present **RewardSlider**, a continuous image editing framework based on reward-guided trajectory control. We first observe that the editing dynamics of FLUX-Kontext exhibit pronounced spatiotemporal plasticity: early generation steps are substantially more sensitive to interventions, and this sensitivity is concentrated in instruction-relevant regions. Building on this observation, RewardSlider keeps the original editing model frozen and introduces lightweight velocity controls only at early timesteps and within edit-relevant regions. Multiple complementary reward models are combined into a terminal optimization objective to assess the generated result from different perspectives, including semantic edit satisfaction, content preservation, and visual quality. Different target strengths are associated with distinct reward constraints, and the terminal rewards are backpropagated through the complete subsequent generation dynamics to jointly optimize a shared set of early control variables, thereby guiding the model to produce outputs of varying strengths along a coherent and continuous editing trajectory. Without requiring intermediate-strength training data or additional model fine-tuning, RewardSlider turns a pretrained instruction-based image editor into a continuously adjustable, fine-grained editing system.

Primary area:
generative models
Keywords:
Continuous Image Editing, Reward-Guided Optimization, Trajectory Control

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_PjZrOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

RewardSlider works on controlling continuous edit strength on a frozen FLUX.1 Kontext through per-instance test-time optimization. It first builds a candidate trajectory by locally interpolating between source-preservation and instruction following velocities in the early sampling steps. It then calibrates the active strength range and slider control points for each pair using DreamSim, and optimizes early-step velocity residuals with SigLIP, DreamSim and DINOv2 rewards. On a 540-image subset of PIE-Bench++, it achieves the best smoothness metrics but not the best edit alignment or monotonicity, at roughly 10× the runtime of VeloEdit.

Critical strengths
  1. The model requires no additional heavy model training to make slider positions map to perceptually uniform changes.
  2. It reports runtime and memory transparently, which makes the cost of its test-time optimization clear.
Critical weaknesses
  1. There is no real novelty. This paper stitches together VeloEdit's velocity interpolation, generic perceptual re-spacing of slider positions, and standard reward-gradient correction. Its key component, the reward correction, is never ablated.
  2. The results do not support the claim. RewardSlider wins only on smoothness metrics, which its own test-time calibration directly optimizes. It trails SliderEdit on edit fidelity and Kontinuous Kontext on monotonicity, produces weaker full-strength edits, runs about 10× slower than VeloEdit, and has no user study.
Questions for authors
  1. If the same DreamSim-based range calibration and control-point re-spacing is applied to VeloEdit, SliderEdit and Kontinuous Kontext, does RewardSlider still outperform them?
  2. What is the contribution of the multi-reward correction alone (an ablation without it)?
Reviewer_FxkpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RewardSlider, an instance-specific framework for continuous instruction-based image editing with a frozen FLUX.1-Kontext model. The method first constructs candidate edit trajectories by interpolating source-preserving and instruction-conditioned velocity fields within an early, spatially localized control window. It then queries generated outputs to determine an active strength range and redistribute slider control points according to DreamSim distances. Finally, it optimizes per-strength velocity residuals at the first four sampling steps using SigLIP, DreamSim, and DINO-based rewards. RewardSlider achieves the lowest reported second-order DreamSim deficit and maximum adjacent LPIPS distance, as well as slightly better preservation outside the edit mask. However, it does not achieve the best edit alignment, endpoint text similarity, or monotonicity, and is substantially slower than all compared methods.

Critical strengths
  1. The spatial and temporal analyses provide some motivation for restricting interventions to early sampling steps and instruction-relevant regions.
  2. The paper observes that linear changes in an internal control value do not necessarily produce uniformly spaced perceptual changes. Calibrating the effective range and control points for each image–instruction pair is therefore better aligned with practical continuous editing.
  3. Activation-range calibration and adaptive control points improve trajectory smoothness, while the spatial mask better preserves instruction-irrelevant regions. The full method consistently outperforms uniform velocity control on CLIP-Dir, second-order DreamSim, and mask-out LPIPS.
Critical weaknesses
  1. RewardSlider performs best on second-order DreamSim and maximum adjacent LPIPS, but its CLIP-Dir is only 0.258, substantially below SliderEdit’s 0.447. Its final CLIP-T is also lower than those of SliderEdit and Kontinuous Kontext. Thus, the trajectory is smoother, but the requested edit may be less fully realized. Since weaker edits naturally produce shorter perceptual paths and smaller adjacent changes, it is unclear whether the improvement comes from better calibration or simply more conservative editing. A comparison at matched endpoint strength or text alignment would be more informative.
  2. DreamSim is used to identify the active range, adjust control-point spacing, and define part of the preservation objective. The paper then evaluates its main smoothness advantage using a second-order DreamSim metric. This creates a risk that the reported improvement is specific to the metric being directly optimized rather than reflecting general perceptual smoothness.
  3. RewardSlider requires 337.43 seconds to generate five strength levels, compared with approximately 32–100 seconds for the baselines. The additional cost comes from instance-specific range probing, control-point refinement, reward evaluation, and velocity-residual optimization.
Questions for authors

My main concern is the substantial inference cost. RewardSlider requires 337.43 seconds to generate five strength levels, compared with 32.13–100.16 seconds for the evaluated baselines. This gap seems difficult to reconcile with the intended use as an interactive editing slider.

Reviewer_9zucOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposed an inference time approach for continuous control for image editing with calibration. The method builds on Flux.Kontext that is an instruction driven image editing flow model and modifies the intermediate velocity to achieve continuous control. The key idea is that perturbing velocity estimates at early stage of denoising (t<4) leads to large changes in the terminal state. Leveraging this information the paper proposed a two step method where first from a given uncaliberated interpolation range, a caliberated scale consisting of discrete interpolants is estimated. Next, for the early timesteps an optimization is performed to learn the velocity residual for smoothness of the trajectory and the preserving text alignment. The paper use off-the-shelf models such as SigLIP for text-alignment and Dino for computing the Dreamsim metric for smoothness. The optimization is done for each image and edit instruction to obtain the smooth edited image trajectory. The qualitative results looks reasonable with identity change in some examples but the method requires optimization and reliance on other model during inference making it impractical for interactive slider based controls.

Critical strengths
  • Having calibrated sliders for continuous control for image editing can be useful and is not explored well in the community; most of the existing slider methods assume the uniform change in sliders will result in uniform output changes.
  • The proposed method is inference time and in principle seems generalizable to other flow models. Though the paper focuses on only one image editing model Flux-Kontext.
Critical weaknesses

[Presentation]

The paper is poorly written with difficult to parse paragraphs, introducing several notations in the method section without clear explanation. This makes the paper very difficult to read. Some major concerns about the presentation are below:

  • Fig.4 (left + centre) - The distinction in different branch is barely visible even at the early tilmestep t-0,t-1 making it difficult to understand the key insight that the paper is referring to. (right) - All the three rows for different regions perturbation appear to generate very similar looking terminal state. The key insight could have been validate with multiple examples and clear distinction between the terminal samples as discussed in the text.
  • Method section has several things going on, it has flow velocity formulation, selection of intermediate calibration states and finally the reward based optimisation. The section 3.2 and 3.3 has several notations without clear explanation or incorrect ordering of notation explanation.
  • The tables are oddly spaced with a large variation in the font size.
  • Eq.6 - What is c’? What is the intuition behind this normalized similarity, the denominator has a term of |V_edit - V_keep|, how is it normalizing the similarity.
  • Eq.11 - What is I_const
  • L362 - H200?

[Method]

  • Extremely slow method as it has two stages making it infeasible for practical slider applications that desire to be close to real time for interactive usecases. The method first finds calibration of the intermediate strengths and then perform optimization of the intermediate velocities at these obtained states.
  • Further performing optimisation on these states using external reward models Dino and SigLIP makes the system dependent on the external reward models during inference whereas in comparison the baseline methods do not expect additional models.

[Experiments]

  • Ideally the setup of slider based control makes intuitive sense for real time applications, where the users can change the slider value and visualise the outputs. The current methods is extremely slow due to the RL optimisation during the generation process making it infeasible for real-time and interactive use cases.
  • Evaluation is shown on a single model Flux-Kontext, that is a previous generation of image editing model, the paper should implement and evaluate on more recent Flux2 and Qwen-Image models. As the method is inference based it can be also integrated with video editing model as is done in Velo-edit. The paper should present some preliminary results on it.
  • Why is there a step for reward optimisation over the intermediate states if we are already doing the calibration. The design choices should be compared as a detailed ablation.
  • Experiments are shown only on a single dataset Piebench is provided, it should be evaluated on other benchmarks for image editing.
Questions for authors
  • How does the mask prediction in Eq.7 play out for global edits, global edits change the complete image in that case the mask M_edit will the complete image?
  • How does the method ensures that the rewards are truly learning a smooth transition of the edit sequency. It enforces text alignment and similarity between current prediction and the source image.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.