acceptodds
Under review as a conference paper at ICLR 2027

StableCtrl: One-Step Stable Image Editing via Optimal Control

Abstract

Recent one-step text-to-image models enable real-time image editing by completing semantic transformations in a single denoising step. However, existing training-free editors often become unstable in this setting. Source-target residuals can act as high-energy control signals, leading to structural distortion and background inconsistency. To address this issue, we introduce StableCtrl, a model-agnostic, training-free, and inversion-free framework for stable one-step image editing. StableCtrl casts one-step editing as a regularized optimal control problem and treats source-target residuals as noisy control cues instead of directly applying them as editing directions. This formulation yields an elliptic control equation with an explicit spectral solution. The solution admits efficient one-step optimization without iterative refinement. It regularizes the control field while retaining the dominant semantic direction. The resulting operator constrains control energy, enforces spatial smoothness, and preserves temporal reference consistency. It therefore balances semantic editing with structural preservation while suppressing unstable high-frequency perturbations through spectral low-pass filtering. StableCtrl consequently produces stable control fields that preserve scene geometry and background consistency without additional model training or latent inversion. Extensive experiments on PIE-Bench show that StableCtrl consistently improves background preservation and editing robustness.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-Guided Image Editing, One-Step Image Editing, Optimal Control, Training-Free Image Editing, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_553rOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

StableCtrl aims to reduce the structural damage that can occur in one-step image editing. It treats the source–target prediction residual as a noisy editing signal, then regularizes its magnitude, spatial variation, and distance from a nearby-time reference. This gives a quadratic objective with a closed-form spectral solution. The authors evaluate it mainly on PIE-Bench with SD-Turbo, and add component ablations, simple regularization baselines, and tests on three other backbones. The main gain is better preservation of structure and background, at some cost to target-prompt alignment.

Critical strengths
  1. The preservation gains are large. In Table 1, PSNR rises from 23.80 to 27.46 and scaled LPIPS drops from 89.20 to 55.55 compared with ChordEdit without the proximal step. Table S1(a) is particularly helpful here: it compares scalar shrinkage, Gaussian smoothing, their combination, and reference blending at roughly matched CLIP-E. StableCtrl still has the best reported preservation metrics. That makes it harder to explain the gain simply as weaker editing.

  2. I like the simplicity of the method. It has an explicit objective and needs neither training nor iterative optimization. Table S1(b) also shows preservation gains on InstaFlow, SwiftBrush-v2, and SDXL-Lightning, in addition to SD-Turbo. Those comparisons are against direct residual editing, so they don't establish superiority over the best editor on each backbone, but they do suggest that the regularizer is useful across generators.

Critical weaknesses
  1. I'm not yet convinced about the contribution of the temporal reference. In Table 2, removing the reference term drops PSNR from 27.46 to 25.23. But in Table 3, the reference-free version has slightly better PSNR and LPIPS than the adopted previous-residual reference, with only a small CLIP-E difference. The authors do say that the settings differ, so I wouldn't call this a contradiction. I would still like to see exactly what changed between the two tables. Without that, it's hard to tell how much the reference contributes, even though it is one of the method's three main components.

  2. I couldn't reconstruct the full inference procedure from the description. Section 3.2 defines a generic perturbation kernel and projection operator, and Section 5.1 lists the regularization weights. The evaluation time, the construction and batching of the nearby-time reference, and the discrete spectral operator still need to be spelled out. The text discusses both Neumann and periodic boundaries, which give different finite-grid implementations. It would also help to report the number of condition–time branches alongside the batched-call NFE count. These details are needed to reproduce the result and check the efficiency comparison.

  3. Minor presentation issues: the equally spaced y-axis ticks in Fig. 6(d) jump from 26 to 28, which appears to be a labeling error. In Fig. 3, please correct “Spatial Smothness” inside the Optimal Control Solver to “Spatial Smoothness” and “Before Row R” to “Before Raw R.”

Questions for authors
  1. What exactly does “reference-free” mean in Table 3, and which settings differ from Table 2? Under the final configuration, does the previous-residual reference improve the preservation–alignment trade-off?
  2. Could you add inference pseudocode? In particular, please specify the evaluation time, perturbation and response-to-control conversion, how both residuals are formed and batched, and the discrete spectral denominator and boundary condition.
2 comments
AuthorsOpenReview ↗

[W1/Q1] Clarifying the Temporal Reference and Its Contribution to the Preservation–Alignment Trade-off

[W1] I'm not yet convinced about the contribution of the temporal reference. In Table 2, removing the reference term drops PSNR from 27.46 to 25.23. But in Table 3, the reference-free version has slightly better PSNR and LPIPS than the adopted previous-residual reference, with only a small CLIP-E difference. The authors do say that the settings differ, so I wouldn't call this a contradiction. I would still like to see exactly what changed between the two tables. Without that, it's hard to tell how much the reference contributes, even though it is one of the method's three main components.

[Q1] What exactly does “reference-free” mean in Table 3, and which settings differ from Table 2? Under the final configuration, does the previous-residual reference improve the preservation–alignment trade-off?

Response: We thank the reviewer for the careful comparison of Tables 2 and 3. We agree that the original presentation did not clearly distinguish the two reference ablations. We clarify their settings and report controlled experiments to assess the contribution of the temporal reference.

[W1-Q1-1] Reference Definitions and Controlled Ablation

The two ablations remove different components:

  • Reference-free (Table 3): Sets the temporal reference to zero but retains its weight (μ=0.25\mu=0.25). The reference penalty becomes a control-energy term.
  • w/o Reference Regularization (Table 2): Sets μ=0\mu=0 and removes the entire reference penalty.

The component-removal variants in Table 2 used (λ,γ,μ,η)=(0.05,0.10,0.20,1.0)(\lambda,\gamma,\mu,\eta)=(0.05,0.10,0.20,1.0) with cleanup, whereas the full method used the final configuration (0.08,0.15,0.25,0.9)(0.08,0.15,0.25,0.9) without cleanup. These differences confounded the original comparison. We have clarified the configuration differences in Implementation Details, the ablation definitions in Ablation of Objective Terms, and the reference-free setting in Reference-Prior Design.

We reran all variants under the final configuration on 700 PIE-Bench images with SD-Turbo (FP32). We used three seeds (42, 43, 44), MC=1, ts=0.90t_s=0.90, δ=0.15\delta=0.15, one editing step, and no cleanup.

Table R1. Controlled reference ablation (mean ± std over three seeds).

MethodPSNR ↑LPIPS ×10³ ↓SD ×10³ ↓CLIP-E ↑
w/o Reference Regularization26.909 ± 0.16963.7 ± 2.55.1 ± 0.521.265 ± 0.017
Reference-free27.505 ± 0.15356.0 ± 2.14.1 ± 0.421.217 ± 0.015
StableCtrl (Full)27.289 ± 0.15257.9 ± 2.24.4 ± 0.421.270 ± 0.019

Removing the entire reference penalty lowers PSNR by 0.380 dB and raises the reported LPIPS from 57.9 to 63.7. Removing only temporal information improves PSNR by 0.216 dB but reduces CLIP-E by 0.053. Thus, the reference penalty helps preserve source structure, while temporal information favors semantic alignment at the default operating point.

[W1-Q1-2] Motivation, Mechanism, and Matched-Alignment Comparison

A control field based only on the current residual depends on a single-time model response. It may therefore retain local variations that are not essential to the edit. We introduce a nearby-time residual from the same source–target task to provide an additional reference.

From the spectral solution, the difference between Full and Reference-free is

u^Full(ω)−u^RF(ω)=μR^ref(ω)1+λ+μ+γ∥ω∥2.\hat{u}_{\mathrm{Full}}(\omega)-\hat{u}_{\mathrm{RF}}(\omega)=\frac{\mu\hat{R}_{\mathrm{ref}}(\omega)}{1+\lambda+\mu+\gamma\|\omega\|^2}.

This term depends on the reference residual at each spatial frequency. It can change the control direction and spatial distribution, unlike a uniform scale applied to the current control. This motivates our use of temporal information rather than magnitude adjustment alone.

To test whether the benefit comes only from stronger editing, we compare Full with Reference-free at matched CLIP-E. We test five global scales (1.10, 1.12, 1.14, 1.16, 1.20). The closest aggregate CLIP-E occurs at 1.16 for seed 42 and 1.12 for seeds 43 and 44. Each scale is fixed across all 700 images.

Table R2. Preservation at matched CLIP-E (mean ± std over three seeds).

MethodPSNR ↑LPIPS ×10³ ↓CLIP-E ↑
Scaled Reference-free26.824 ± 0.07163.3 ± 1.421.269 ± 0.017
StableCtrl (Full)27.289 ± 0.15257.9 ± 2.221.270 ± 0.019

At nearly identical CLIP-E, Full improves PSNR by 0.465 ± 0.099 dB and reduces the reported LPIPS by 5.4 ± 0.9. The preservation gains hold across all three seeds. These results support the value of the temporal reference beyond uniform control scaling. They are consistent with the proposed mechanism, although they do not directly establish which spatial changes account for the gains.

The matching scales were selected on the evaluation set. This comparison assesses the trade-off against uniform scaling, not held-out generalization.

AuthorsOpenReview ↗

[W2/Q2/W3] Clarifying Inference Details, Computational Cost, and Presentation Issues

[W2] I couldn't reconstruct the full inference procedure from the description. Section 3.2 defines a generic perturbation kernel and projection operator, and Section 5.1 lists the regularization weights. The evaluation time, the construction and batching of the nearby-time reference, and the discrete spectral operator still need to be spelled out. The text discusses both Neumann and periodic boundaries, which give different finite-grid implementations. It would also help to report the number of condition–time branches alongside the batched-call NFE count. These details are needed to reproduce the result and check the efficiency comparison.

[Q2] Could you add inference pseudocode? In particular, please specify the evaluation time, perturbation and response-to-control conversion, how both residuals are formed and batched, and the discrete spectral denominator and boundary condition.

Response: We agree that the original description lacked several details needed to reproduce the inference procedure and assess its cost. We clarify the implementation below and identify the spectral formulation used in our code.

[W2-Q2-1] Inference Procedure and Temporal Reference

StableCtrl evaluates the main residual at ts=0.90t_s=0.90 and the reference residual at tref=ts−δ=0.75t_{\mathrm{ref}}=t_s-\delta=0.75, with δ=0.15\delta=0.15. The nearby-time reference provides a second response from the same editing task without an additional sequential denoising step.

At each time, the source latent is perturbed as

zt=αtxsrc+σtϵ,z_t=\alpha_t x_{\mathrm{src}}+\sigma_t\epsilon,

where ϵ\epsilon is Gaussian noise. The SD-Turbo prediction is converted to a clean-latent estimate:

x^0(zt,t,c)=zt−σtϵθ(zt,t,c)αt.\hat{x}_{0}(z_t,t,c)=\frac{z_t-\sigma_t\epsilon_\theta(z_t,t,c)}{\alpha_t}.

The semantic residual is the difference between target and source responses:

Rt=x^0(zt,t,ctar)−x^0(zt,t,csrc).R_t=\hat{x}_{0}(z_t,t,c_{\mathrm{tar}})-\hat{x}_{0}(z_t,t,c_{\mathrm{src}}).

The reference residual RrefR_{\mathrm{ref}} is computed at treft_{\mathrm{ref}} using the same procedure. With MC=1, the four source/target and time branches are processed in one batched UNet call.We have specified the evaluation times, temporal-reference construction, and four-branch batching in the revised Implementation Details. The perturbation and residual formulations are provided in the method formulation.

Algorithm R1. StableCtrl one-step inference.

Input: I_src, c_src, c_tar, SD-Turbo
       t_s=0.90, delta=0.15, MC=1
       lambda=0.08, gamma=0.15, mu=0.25, eta=0.90

1: x_src <- Encode(I_src)
2: t_ref <- t_s - delta
3: Sample epsilon; construct z_t and z_ref
4: Batch four condition-time branches:
     (z_t, t_s, c_src), (z_t, t_s, c_tar)
     (z_ref, t_ref, c_src), (z_ref, t_ref, c_tar)
5: Evaluate the batched SD-Turbo UNet
6: Convert predictions to clean-latent estimates
7: R_t <- x0_hat(t_s,c_tar) - x0_hat(t_s,c_src)
8: R_ref <- x0_hat(t_ref,c_tar) - x0_hat(t_ref,c_src)
9: N <- FFT2(R_t) + mu * FFT2(R_ref)
10: D <- 1 + lambda + mu
         + gamma * (omega_x^2 + omega_y^2)
11: u <- IFFT2(N/D).real
12: I_edit <- Decode(x_src + eta*u)

Output: I_edit

The control is applied once to the source latent. Thus, one-step editing denotes a single latent update, not a single condition–time evaluation.

[W2-Q2-2] Spectral Implementation and Computational Cost

For a latent field of size H×WH\times W, the frequency coordinates are

ωk=2π fftfreq(H)k,ωl=2π fftfreq(W)l.\omega_k=2\pi\,\mathrm{fftfreq}(H)_k,\qquad \omega_l=2\pi\,\mathrm{fftfreq}(W)_l.

The implemented spectral solution is

u^k,l=R^t,k,l+μR^ref,k,l1+λ+μ+γ(ωk2+ωl2).\hat{u}_{k,l}=\frac{\hat{R}_{t,k,l}+\mu\hat{R}_{\mathrm{ref},k,l}}{1+\lambda+\mu+\gamma(\omega_k^2+\omega_l^2)}.

The implementation uses a two-dimensional FFT with periodic boundary conditions and a denominator based on squared FFT frequencies, rather than Neumann boundaries or finite-difference Laplacian eigenvalues. We have corrected the corresponding formulation in the revised Elliptic Optimality Condition and Spectral Filtering subsections.

For MC=1, inference evaluates four condition–time branches in one batched UNet call. The temporal reference adds two branches but no sequential denoising step. Batched-call NFE counts model invocations, while branch count reflects the number of evaluated condition–time pairs. Although batching reduces sequential calls, all four branches incur computation. We therefore distinguish these quantities rather than treat them as equivalent costs.

[W3] Presentation Corrections. We have corrected the y-axis ticks and spacing in Fig. 6(d) and both typos in Fig. 3 (“Spatial Smothness” → “Spatial Smoothness”; “Before Row R” → “Before Raw R”). These corrections do not alter the underlying experimental results.

Reviewer_hQXnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

StableCtrl addresses training-free, inversion-free one-step text-guided editing with distilled generators. Building on ChordEdit, it regularizes the source–target residual through magnitude, spatial smoothness, and nearby-timestep consistency penalties, yielding a screened-Poisson solution computed via FFT. The formulation is motivated by optimal control and supported by well-posedness and energy-bound analyses. Experiments on PIE-Bench report stronger background preservation than competing one-step editors at low runtime, but lower semantic alignment. The evaluation includes component ablations, matched-semantic filtering comparisons, and experiments with additional backbones.

Critical strengths

S1. A simple, efficient stabilizer with consistent preservation gains. The closed-form spectral solve adds no network evaluations and negligible latency. Background preservation improves across four one-step backbones and all ten PIE-Bench categories, supporting the practical value of regularizing a single editing update without training.

S2. An interpretable operator supported by relevant controls. The formulation provides clear spectral interpretations and an explicit energy bound, with empirical checks of the predicted behavior. Comparisons against shrinkage, smoothing, and reference blending at approximately matched semantic alignment support preservation gains beyond simply weakening edits.

Critical weaknesses

W1. Unclear improvement in the fidelity–semantics trade-off and inconsistent efficiency comparisons. StableCtrl has the weakest semantic alignment in Table 1 (CLIP-W 23.34; CLIP-E only ≈0.25 above a near-identity update, Fig. 5a), and fidelity depends strongly on update magnitude, so weaker editing remains a plausible explanation for its preservation gains. Table S1(a) provides a useful matched-CLIP comparison, but only at one operating point. Missing controls include a plain Gaussian or Butterworth low-pass of the same residual mixture, a timestep shift, and additional Monte Carlo samples under matched compute. Performance frontiers are needed to establish an advantage. Runtime comparisons also mix hardware settings, while ChordEdit without its proximal stage is reported slower than the full method, leaving the claimed speedup unsupported.

W2. Limited theoretical novelty and insufficient justification of the control formulation. The reduction removes the pretrained dynamics, leaving classical quadratic H¹/Tikhonov regularization solved by a fixed isotropic filter. Standard well-posedness and energy bounds do not establish what the optimal-control view contributes beyond residual smoothing. Connections to frequency-aware editing (e.g., FDS, CVPR 2025) and ChordEdit’s residual averaging are insufficiently distinguished. Moreover, uniform filtering may suppress fine target details while retaining low-frequency background leakage, yet per-category semantic results are absent.

Questions for authors

Q1. Fidelity and semantics frontier. Please compare preservation versus semantic alignment using shared seeds for StableCtrl, ChordEdit without and with the proximal stage, and SwiftEdit, sweeping their respective strength parameters. For the direct residual, also vary the evaluation timestep with δ = 0. Include unedited-source scores. Does StableCtrl improve the frontier or merely occupy a more conservative operating point?

Q2. Strong simple alternatives. At matched semantic alignment, how does StableCtrl compare with Gaussian or Butterworth filtering of the same residual mixture, additional Monte Carlo samples under matched runtime, and a pure timestep shift?

Q3. Result provenance and timing. Which Table 1 results were rerun versus taken from ChordEdit? Please benchmark all one-step methods under identical hardware, resolution, and precision, and report conditional forward passes and FLOPs. Why is ChordEdit without the proximal stage reported slower than the full method and its original implementation?

Q4. Implementation details. Please specify the evaluation timestep, Laplacian discretization, frequency units and grid spacing, boundary conditions, per-channel treatment, and filtering order relative to B_t. What is the effective cutoff at γ = 0.15 on a 64 × 64 latent?

5 comments
AuthorsOpenReview ↗

W1–Q1: Matched-Strength Evaluation and Experimental Evidence

Q1. Fidelity and semantics frontier.

We thank the reviewer for these concerns. Preservation must be evaluated alongside semantic alignment. We compare methods at matched CLIP-E levels and report default operating points separately. The efficiency concern is addressed in Q3 through unified runtime measurements.

W1. Unclear improvement in the fidelity–semantics trade-off and inconsistent efficiency comparisons.

Why is matched semantic alignment necessary? Reducing residual magnitude can improve preservation simply by weakening the intended edit. An unchanged source even achieves perfect reconstruction without editing. StableCtrl instead regularizes residuals through spatial smoothness and temporal-reference consistency, rather than uniform magnitude reduction. We evaluate three matched CLIP-E levels to test whether this structured control preserves more content at comparable editing strength.

Table R1. Preservation–Semantics Trade-off at Matched CLIP-E.
PSNR ↑ / LPIPS × 10³ ↓.

MethodLow (21.30)Moderate (21.60)High (21.80)
ChordEdit Full27.21 / 53.8425.83 / 70.2624.91 / 93.47
ChordEdit w/o Proximal27.56 / 48.4025.62 / 65.8424.37 / 81.20
SwiftEdit26.48 / 64.7224.91 / 79.3622.16 / 92.84
StableCtrl27.38 / 56.2725.97 / 63.4125.38 / 78.63

Analysis. At low editing strength, ChordEdit without proximal refinement achieves better preservation. At moderate and high strengths, StableCtrl achieves the highest PSNR and lowest LPIPS, exceeding the best competing PSNR by 0.14 and 0.47 dB. ChordEdit uses low-energy transport and proximal refinement, whereas StableCtrl jointly regulates residual fidelity, control magnitude, spatial variation, and nearby-time consistency. This difference may help explain StableCtrl's stronger preservation at higher editing strengths, although it does not dominate the entire trade-off.

Q1. Preservation–semantic alignment trade-off.

Why do default settings differ? Each method uses its own default editing configuration, which may produce different semantic strengths. Consequently, preservation metrics measured at these settings cannot isolate the effect of the editing algorithm. We report default operating points separately to clarify this distinction.

Table R2. Comparison at Default Operating Points.
PSNR ↑ / LPIPS × 10³ ↓.

MethodCLIP-E ↑PSNR ↑LPIPS × 10³ ↓
Unedited Source (η=0\eta=0)21.18∞0.00
SwiftEdit21.8521.7191.22
ChordEdit Full22.9622.20128.25
ChordEdit w/o Proximal21.8723.8988.36
Direct Residual21.9920.57139.70
StableCtrl21.2627.4655.55

Unedited Source. Setting η=0\eta=0 leaves the source unchanged, yielding infinite PSNR and zero LPIPS. Its nonzero CLIP-E reflects metric similarity rather than successful editing, highlighting the need to evaluate preservation alongside semantic alignment.

Analysis. Table R2 shows that StableCtrl achieves the strongest preservation among editing methods at default settings, although its CLIP-E is lower. Since default editing strengths differ, this comparison alone cannot establish a superior trade-off. Crucially, Table R1 shows that StableCtrl retains higher PSNR and lower LPIPS at matched moderate and high semantic strengths. Thus, its preservation advantage cannot be explained solely by weaker editing.

Why does joint control help? One-step editing applies the residual without later refinement, making unwanted changes difficult to correct. Increasing residual strength can amplify both intended edits and spatial artifacts, while uniform scaling suppresses both. StableCtrl instead combines residual fidelity, control regularization, spatial smoothness, and temporal-reference consistency. Spatial regularization provides frequency-dependent attenuation, while the temporal reference incorporates nearby-time residual information. This structured control helps explain the preservation gains at matched semantic strengths.

What does this comparison establish? The matched-strength results support the benefit of joint residual control beyond simply weakening the edit. StableCtrl achieves stronger preservation at moderate and high semantic strengths, while ChordEdit remains competitive at low strength. We therefore claim a more favorable trade-off in the evaluated regimes, not universal dominance.

AuthorsOpenReview ↗

[W2/Q2] Theoretical Motivation, Joint Control Formulation, and Comparison with Simple Alternatives

W2. Limited theoretical novelty and insufficient justification of the control formulation.

We agree that quadratic regularization and FFT solvers are classical. Our contribution lies in formulating residual control for one-step editing, where semantic fidelity, update magnitude, spatial variation, and temporal consistency must be balanced within a single update.

Why is control needed? One-step editing relies on a single residual RtR_t without subsequent correction. This residual may contain both desired edits and unwanted source changes. Uniform scaling cannot distinguish them: smaller updates preserve more content but may weaken the edit. This motivates controlling residual structure rather than merely adjusting its magnitude.

Why joint constraints? Frequency alone does not determine semantic relevance: high-frequency responses may contain target details, while low-frequency changes may disrupt the background. FDS [1] employs spatially localized frequency optimization, while ChordEdit [2] uses a transport-derived editing field with optional proximal refinement. However, neither formulation explicitly combines spatial regularization with nearby-time residual consistency in the same control objective. StableCtrl addresses this distinction by jointly balancing residual fidelity, control energy, spatial smoothness, and temporal-reference consistency. The fidelity term retains the editing signal, energy regularization limits excessive updates, spatial smoothness penalizes abrupt variations, and the temporal reference introduces nearby-time information. Their joint optimization provides structured residual control beyond uniform scaling or frequency suppression alone.

Q2. Strong simple alternatives.

To test simpler alternatives, we compare Gaussian filtering, Butterworth filtering, reference averaging, and removal of reference regularization at matched CLIP-E. We further examine whether additional Monte Carlo sampling or timestep shifting can explain the gains. These comparisons control for editing strength and are included in Appendix D.1, Tables S1(a)–S1(b) of the revised supplementary material.

Table R3. Comparison with Simple Filtering and Residual Alternatives.
PSNR ↑ / LPIPS × 10³ ↓ at matched CLIP-E.

MethodLow (21.30)Moderate (21.60)High (21.80)
Gaussian Filtering26.38 / 66.2125.31 / 73.2624.86 / 87.41
Butterworth Filtering26.24 / 68.4725.18 / 76.5324.69 / 91.27
Reference Averaging26.87 / 61.4325.24 / 75.1824.77 / 89.63
No Reference Regularization26.91 / 60.7625.46 / 70.3524.91 / 85.74
StableCtrl27.38 / 56.2725.97 / 63.4125.38 / 78.63

Table R4. Comparison with Additional Monte Carlo Sampling and Timestep Shifting.
PSNR ↑ / LPIPS × 10³ ↓ at matched CLIP-E.

MethodLow (21.30)Moderate (21.60)High (21.80)
Direct Residual (MC=1)26.71 / 55.3823.94 / 87.1622.02 / 114.31
Direct Residual (MC=4)27.12 / 54.8325.42 / 69.7624.86 / 85.93
Direct Residual + Timestep Shift (δ=0\delta=0, MC=1)27.06 / 55.1425.51 / 68.4224.97 / 84.61
StableCtrl (MC=1)27.38 / 56.2725.97 / 63.4125.38 / 78.63

Table R5. Runtime Comparison under Different Monte Carlo Sample Counts.

ConfigurationRuntime (s/image) ↓
StableCtrl (MC=1)0.231
StableCtrl (MC=2)0.325
StableCtrl (MC=4)0.522
Direct Residual (MC=1)0.243
Direct Residual (MC=4)0.532

Monte Carlo Sampling. Table R4 shows that StableCtrl (MC=1) achieves the highest PSNR at all three matched CLIP-E levels and the lowest LPIPS at moderate and high strengths. At low strength, Direct Residual (MC=4) has slightly lower LPIPS (54.83 vs. 56.27) but takes 0.532 s/image versus 0.231 s/image (Table R5). Increased sampling alone therefore does not explain the preservation gains.

Why are simple alternatives insufficient? Filtering controls spatial variation, reference averaging introduces temporal information, and additional sampling reduces variability. StableCtrl instead jointly balances residual fidelity, control energy, spatial smoothness, and temporal consistency. Tables R3–R4 show that these individual alternatives do not fully explain its gains.

What does the formulation provide? StableCtrl integrates current and nearby-time residuals into a unified control objective. Its closed-form spectral solution requires no iterative optimization or model changes. Tables R3–R4 support its preservation advantages over the tested alternatives, highlighting the value of joint residual control.

References

[1] FDS: Frequency-Aware Denoising Score for Text-Guided Latent Diffusion Image Editing. CVPR 2025.

[2] ChordEdit: One-Step Low-Energy Transport for Image Editing. CVPR 2026.

AuthorsOpenReview ↗

Q3–Q4. Runtime Efficiency, Implementation Details, and W2-Category-wise Validation

Q3. Result provenance and timing.

We thank the reviewer for identifying the runtime reporting inconsistency. The multi-step, few-step, and SwiftEdit results in Table 1 follow ChordEdit. We re-evaluated ChordEdit and StableCtrl on 100 images using a single RTX 4090D under identical settings, without GPU contention. The corrected results are shown below.

Unified Runtime and Reproduced Results.

MethodUNet Calls ↓Equivalent Forwards ↓Runtime (s) ↓VRAM (MB) ↓PSNR ↑LPIPS × 1000 ↓CLIP-E ↑
ChordEdit Full250.27591622.20128.2522.96
ChordEdit w/o Proximal140.23591623.8988.3621.87
StableCtrl140.23591427.4655.5521.26

NFE clarification and paper revision. NFE counts batched UNet calls, not equivalent single-image forwards. ChordEdit Full adds one call for proximal refinement. StableCtrl matches ChordEdit without proximal refinement in four UNet branches, adding non-iterative FFT control. Their runtimes are comparable, while StableCtrl achieves better default preservation. Matched-strength comparisons are provided in Q1 to account for CLIP-E differences. We have corrected ChordEdit results and runtime values in Table 1, with revisions highlighted in red.

Q4. Implementation details.

For reproducibility, we use ts=0.90t_s=0.90, tref=0.75t_{\mathrm{ref}}=0.75, and δ=0.15\delta=0.15, with one Monte Carlo sample and fixed random seeds. The control parameters are λ=0.08\lambda=0.08, γ=0.15\gamma=0.15, and μ=0.25\mu=0.25. All latent channels share these parameters, with αc=1\alpha_c=1. The FFT solution is applied independently to each latent channel:

u^c,k,l=αcR^t,c,k,l+μcR^ref,c,k,l1+λc+μc+γc(ωk2+ωl2).\hat{u}_{c,k,l}=\frac{\alpha_c\hat{R}_{t,c,k,l}+\mu_c\hat{R}_{\mathrm{ref},c,k,l}}{1+\lambda_c+\mu_c+\gamma_c(\omega_k^2+\omega_l^2)}.

Spectral operator and boundary conditions. The FFT solution assumes periodic boundary conditions. We use a Fourier spectral discretization of the continuous Laplacian, with angular frequencies ωk=2π fftfreq(64)\omega_k=2\pi\,\mathrm{fftfreq}(64) and unit spatial grid spacing. The frequencies lie in [−π,π)[-\pi,\pi), with Nyquist magnitude π\pi. The denominator uses the Fourier symbol ωk2+ωl2\omega_k^2+\omega_l^2, rather than the five-point finite-difference symbol 4−2cos⁡(ωk)−2cos⁡(ωl)4-2\cos(\omega_k)-2\cos(\omega_l). We have clarified the boundary conditions and spectral discretization in Sections 4.2.2 and 4.3 of the revised paper.

The reference uref=Rt−δu_{\mathrm{ref}}=R_{t-\delta} is computed alongside the current residual in the same batched UNet call. Both residuals enter the spectral solution before the resulting control is applied to the editing update. Spatial regularization attenuates rapidly varying components, while reference consistency reduces dependence on a single evaluation timestep. No additional UNet calls or iterative optimization are required.

Additional Evidence: Category-wise Preservation at Matched Semantic Alignment.

Why is category-wise matching necessary? Aggregate gains may be driven by a few editing categories, while default settings can differ in semantic strength. We compare StableCtrl with ChordEdit without proximal refinement across ten categories at matched CLIP-E to test its preservation advantage across tasks.

Table R6. Category-wise Preservation at Matched Semantic Alignment.
Results at moderate editing strength (seed 42). StableCtrl uses the final configuration with δ=0.15\delta=0.15. PSNR ↑ / LPIPS × 10³ ↓.

CategoryMatched CLIP-EChordEdit w/o Prox.StableCtrl
Object Change21.5424.86 / 79.4225.31 / 73.26
Object Addition21.6225.14 / 71.8525.48 / 68.37
Object Deletion21.4727.18 / 57.6226.93 / 60.14
Content Change21.6825.37 / 68.9125.82 / 64.73
Pose Change21.5924.92 / 76.3824.76 / 78.12
Color Change21.7526.11 / 62.4526.47 / 59.83
Material Change21.6325.76 / 66.7226.14 / 62.19
Background Change21.5124.68 / 81.2625.09 / 76.48
Style Change21.7225.23 / 72.3425.51 / 69.86
Random Editing21.5626.18 / 63.7226.04 / 65.13

Analysis. StableCtrl achieves higher PSNR and lower LPIPS in seven of ten categories, including object changes, material edits, and background changes. ChordEdit performs better in object deletion, pose change, and random editing. Together with the overall matched-strength results in Table R1, these findings show that StableCtrl's preservation advantage extends across diverse editing tasks under comparable semantic alignment, rather than being driven by a few categories.

Supplementary Revision. The complete category-wise matched-strength evaluation is included in Appendix D.2, Table S2 of the revised supplementary material.

Reviewer_KFUrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes StableCtrl, a model-agnostic, training-free, and inversion-free framework for one-step image editing. The core idea is to cast the one-step image editing problem as regularized optimal control problem and treats source-target residuals as noisy control cue. This allows one-step optimization without iterative refinement, constraining control energy, enforcing spatial smoothness, as well as preserving temporal reference consistency. Extensive experiments are conducted on PieBench to demonstrate method’s effectiveness.

Critical strengths

The method is a closed-form FFT-domain filter with no training, no inversion, and no iteration. It drops into any one-step editor that produces a residual field.

Critical weaknesses
  • The background discussion on one-step image editing is hard to follow. Specifically, it lacks intuitive description on how Eq.2 relates to one-step edit.
  • The model outperforms in terms of background preservation but comes at the cost of editing semantic. This is evident through the clip editing score where it underperforms for most of previous methods.
  • The term “reference prior” is discussed frequently (line 247, 268, 456) but no formal definition is provided.
  • The method performs the same model computation as the ChordEdit variant plus an extra filtering step. However, it is reported as faster (0.26 s vs 0.43 s). No explanation is given.
  • There are many inconsistent numbers in the ablation tables.
    • The default method scores 27.46 in the main table but 27.29 in later tables.
    • Removing the “reference” component gives 27.50 PSNR in Tab. 3 but 25.70 in Tab.2
Questions for authors

My main concern is the performance on PieBench. Specifically, the background preservation comes at a significant trade-off in terms of editing semantic.

5 comments
AuthorsOpenReview ↗

W1–W3: Motivation and Mathematical Interpretation of StableCtrl

W1. Unclear connection between Eq. 2 and one-step image editing.

We thank the reviewer for pointing this out. We agree that the connection between Eq. 2 and one-step editing needs clarification. We have explained its continuous-time control interpretation in the revised One-Step Image Editing subsection. Below, we clarify why such control is needed.

Why does one-step editing require control? A one-step editor applies a residual update to the source latent:

zedit=zsrc+ηRt.z_{\mathrm{edit}}=z_{\mathrm{src}}+\eta R_t.

Here, RtR_t denotes the source–target residual. Unlike iterative methods, one-step editing has no subsequent updates to correct unwanted changes. The main difficulty is that the residual may affect not only the intended regions but also source content that should remain unchanged.

Existing methods address this differently. FDS [1] uses frequency-aware denoising but is not designed for one-step editing. SwiftEdit [2] focuses on one-step inversion and localized editing, while Direct Residual applies the update without control. ChordEdit [3] uses low-energy transport and proximal refinement. None jointly optimizes residual fidelity, update magnitude, spatial variation, and nearby-time consistency.

Why is this distinction important? A residual can produce the intended semantic change while also modifying unrelated source content. Applying it directly retains both effects. Reducing the update strength suppresses unwanted changes but also weakens the intended edit. Likewise, low-energy transport or proximal refinement can improve preservation without explicitly regulating all these properties together. StableCtrl addresses this remaining problem: how to retain the editing signal while limiting unnecessary changes in a single update.

Why are joint constraints necessary? We optimize the controlled update through

J(u)=∥u−Rt∥22+λ∥u∥22+γ∥∇u∥22+μ∥u−uref∥22.J(u)=\|u-R_t\|_2^2+\lambda\|u\|_2^2+\gamma\|\nabla u\|_2^2+\mu\|u-u_{\mathrm{ref}}\|_2^2.

The first term keeps the update close to the predicted residual. The remaining terms limit excessive changes, abrupt spatial variations, and deviation from a nearby-time reference. Each addresses a different source of instability. Reducing magnitude alone cannot regulate spatial variation, while smoothing alone cannot account for temporal consistency. Their joint optimization provides a controlled alternative to directly applying or uniformly scaling the residual.

W3. Missing formal definition of the reference prior.

What is the reference prior? We define

uref:=Rt−δ,δ>0.u_{\mathrm{ref}}:=R_{t-\delta},\qquad \delta>0.

Here, Rt−δR_{t-\delta} is the source–target residual evaluated at a nearby timestep using the same pretrained model. It is not an externally learned prior or an additional network.

Why is temporal consistency useful? A residual from a single timestep can contain variations that are not essential to the intended edit. The nearby-time reference provides another response to the same editing task. Encouraging agreement between these responses reduces reliance on one evaluation timestep without assuming that the reference is error-free.

How does the formulation regulate the update? Under periodic boundary conditions, the frequency-domain solution is

u^k,l=R^t,k,l+μR^ref,k,l1+λ+μ+γ(ωk2+ωl2).\hat{u}_{k,l}=\frac{\hat{R}_{t,k,l}+\mu\hat{R}_{\mathrm{ref},k,l}}{1+\lambda+\mu+\gamma(\omega_k^2+\omega_l^2)}.

It can be separated into a reference-guided residual and a frequency-dependent gain:

R^mix,k,l=R^t,k,l+μR^ref,k,l1+μ.\hat{R}_{\mathrm{mix},k,l}=\frac{\hat{R}_{t,k,l}+\mu\hat{R}_{\mathrm{ref},k,l}}{1+\mu}.

H(ωk,ωl)=1+μ1+λ+μ+γ(ωk2+ωl2).H(\omega_k,\omega_l)=\frac{1+\mu}{1+\lambda+\mu+\gamma(\omega_k^2+\omega_l^2)}.

The reference-guided residual combines two model responses, while the frequency-dependent gain regulates their magnitude and spatial variation. Unlike filtering the current residual alone, this solution also uses nearby-time information. The control is computed in closed form, without iterative refinement.

Our contribution lies in formulating one-step residual editing as a joint control problem, rather than introducing a new FFT solver. By combining residual fidelity, control energy, spatial smoothness, and reference consistency, StableCtrl aims to preserve the editing signal while limiting unwanted changes. We have formally defined the reference prior in Optimal Control Formulation and clarified its role in the revised paper. The preservation–semantics trade-off is further evaluated at matched CLIP-E in W2–Q1.

References

[1] Yufan Ren et al. "FDS: Frequency-Aware Denoising Score for Text-Guided Latent Diffusion Image Editing." CVPR, 2025.

[2] Trong-Tung Nguyen et al. "SwiftEdit: Lightning Fast Text-Guided Image Editing via One-Step Diffusion." CVPR, 2025.

[3] Liangsi Lu et al. "ChordEdit: One-Step Low-Energy Transport for Image Editing." CVPR, 2026.

AuthorsOpenReview ↗

W2–Q1: Preservation–Semantics Trade-off at Matched Editing Strength

W2. Background preservation comes at the cost of editing semantics.

Q1. Does improved background preservation come at a significant cost in semantic editing?

We thank the reviewer for raising this concern. We agree that preservation must be evaluated alongside semantic alignment. StableCtrl has a lower default CLIP-E than competing methods, which can partly explain its higher preservation scores. We therefore compare methods at matched CLIP-E and report their default operating points separately.

Why is matched semantic alignment necessary? Higher PSNR can simply result from weaker editing. The unedited source illustrates this limitation: it preserves the image perfectly without performing the intended transformation. To control for editing-strength differences, we evaluate methods at three matched CLIP-E levels.

Table R1. Preservation–Semantics Trade-off at Matched CLIP-E.
PSNR ↑ / LPIPS × 10³ ↓.

MethodLow (21.30)Moderate (21.60)High (21.80)
ChordEdit Full27.21 / 53.8425.83 / 70.2624.91 / 93.47
ChordEdit w/o Proximal27.56 / 48.4025.62 / 65.8424.37 / 81.20
SwiftEdit26.48 / 64.7224.91 / 79.3622.16 / 92.84
StableCtrl27.38 / 56.2725.97 / 63.4125.38 / 78.63

Analysis. Table R1 shows that ChordEdit without proximal refinement performs better at low editing strength. At moderate and high strengths, StableCtrl achieves the highest PSNR and lowest LPIPS, improving PSNR over the strongest competitor by 0.14 and 0.47 dB, respectively. Thus, weaker editing alone cannot explain its preservation gains.

Why does the advantage vary with editing strength? Stronger residual updates can amplify both intended edits and unwanted source changes. Uniform scaling reduces both simultaneously. StableCtrl instead optimizes

J(u)=∥u−Rt∥22+λ∥u∥22+γ∥∇u∥22+μ∥u−uref∥22.J(u)=\|u-R_t\|_2^2+\lambda\|u\|_2^2+\gamma\|\nabla u\|_2^2+\mu\|u-u_{\mathrm{ref}}\|_2^2.

The fidelity term retains agreement with the predicted residual, while the remaining terms regulate update magnitude, spatial variation, and nearby-time consistency. Unlike uniform scaling, which attenuates all residual components equally, StableCtrl applies frequency-dependent regularization and incorporates a nearby-time reference. This allows different residual components to be regulated differently rather than uniformly weakening the entire edit. The results in Table R1 are consistent with this mechanism at moderate and high editing strengths, although the advantage does not extend across the entire trade-off frontier.

Why do default operating points matter? Different methods use different editing strengths. We report their default results to distinguish these differences from the matched-strength comparison.

Table R2. Comparison at Default Operating Points.
PSNR ↑ / LPIPS × 10³ ↓ / SD × 10³ ↓.

MethodCLIP-E ↑PSNR ↑LPIPS × 10³ ↓SD × 10³ ↓
Unedited Source (η=0\eta=0)21.18∞0.000.00
SwiftEdit21.8521.7191.228.22
ChordEdit Full22.9622.20128.257.13
ChordEdit w/o Proximal21.8723.8988.366.56
Direct Residual21.9920.57139.7021.70
StableCtrl21.2627.4655.552.86

Analysis. Table R2 confirms that StableCtrl achieves the strongest preservation among editing methods at its default setting, but also has the lowest CLIP-E. The unedited source (η=0\eta=0) further illustrates why preservation alone is insufficient: identical images yield infinite PSNR and zero LPIPS and SD without performing an edit. Its nonzero CLIP-E reflects metric similarity rather than editing success.

What do these results establish? StableCtrl does not dominate the entire preservation–semantics frontier. Its advantage is more pronounced at moderate and high matched editing strengths, where it achieves better preservation at the same CLIP-E. Together, Tables R1–R2 show that default editing strength partly explains the original preservation gap, but does not fully account for StableCtrl's gains.

AuthorsOpenReview ↗

W4–W5: Runtime and Ablation Consistency

W4. Inconsistent runtime comparison with ChordEdit.

We thank the reviewer for identifying this issue. We agree that the original comparison did not adequately control for implementation and measurement conditions. We conducted a unified runtime evaluation and corrected the reported results in the revised paper.

Why is a unified comparison necessary? Runtime depends on model calls, batching, and implementation overhead. StableCtrl and ChordEdit without proximal refinement both use one batched UNet call with four single-image-equivalent forwards. StableCtrl additionally performs FFT-based control without iterative optimization.

Table R1. Unified Runtime Comparison on RTX 4090D.
Average over 100 images without GPU contention.

MethodUNet CallsEquivalent ForwardsRuntime (s/image)Memory (MB)
ChordEdit Full250.26805916
ChordEdit w/o Proximal140.23125916
StableCtrl140.23115914

Analysis. Table R1 shows comparable runtime for StableCtrl and ChordEdit without proximal refinement (0.2311 vs. 0.2312 s/image), while ChordEdit Full takes 0.2680 s/image. We therefore claim no meaningful speed advantage. Instead, StableCtrl introduces closed-form residual control within the same UNet evaluation budget, improving preservation at comparable inference cost. We have corrected the runtime comparison and clarified the measurement protocol in the revised paper.

W5. Inconsistent numerical results across ablation tables.

We agree that the original presentation did not clearly distinguish seed statistics and reference-ablation definitions. The full-model PSNR differs because Table 2 reports seed 42 (27.46), whereas Table 3 averages seeds 42–44 (27.29). Both use the same configuration.

Why do the reference-ablation results differ? The tables differ in configuration and ablation definition. Table 2 reports 25.70 PSNR under an earlier configuration with cleanup, while Table 3 reports 27.50 under the final configuration without cleanup. Crucially, Reference-free (Table 3) sets uref=0u_{\mathrm{ref}}=0 with μ=0.25\mu=0.25, retaining an additional energy penalty μ∥u∥22\mu\|u\|_2^2. In contrast, w/o Reference Regularization (Table 2) sets μ=0\mu=0, removing this penalty entirely. These are distinct controls, not contradictory results. We have clarified them in Implementation Details, Ablation of Objective Terms, and Reference-Prior Design.

How does the reference contribute? We reran all variants under the final configuration on 700 PIE-Bench images using SD-Turbo (FP32), three seeds (42, 43, 44), MC=1, and no cleanup.

Table R2. Controlled Reference Ablation (mean ± std over three seeds).

MethodPSNR ↑LPIPS × 10³ ↓SD × 10³ ↓CLIP-E ↑
w/o Reference Regularization26.909 ± 0.16963.7 ± 2.55.1 ± 0.521.265 ± 0.017
Reference-free27.505 ± 0.15356.0 ± 2.14.1 ± 0.421.217 ± 0.015
StableCtrl (Full)27.289 ± 0.15257.9 ± 2.24.4 ± 0.421.270 ± 0.019

Removing reference regularization reduces PSNR by 0.380 dB and increases LPIPS by 5.8. Setting uref=0u_{\mathrm{ref}}=0 instead increases PSNR by 0.216 dB but lowers CLIP-E by 0.053, as the remaining penalty also suppresses update magnitude. Thus, default preservation cannot isolate the benefit of nearby-time information.

Why is matched alignment necessary? Reference-free has higher default PSNR but lower CLIP-E. We therefore compare it with Full at matched semantic alignment.

Table R3. Preservation at Matched CLIP-E (mean ± std over three seeds).

MethodPSNR ↑LPIPS × 10³ ↓CLIP-E ↑
Scaled Reference-free26.824 ± 0.07163.3 ± 1.421.269 ± 0.017
StableCtrl (Full)27.289 ± 0.15257.9 ± 2.221.270 ± 0.019

Why does the nearby-time reference help? A single-timestep residual may contain unwanted variations. StableCtrl incorporates a nearby-time response to reduce reliance on one residual estimate. At matched CLIP-E, Table R3 shows gains of 0.465 dB in PSNR and 5.4 in LPIPS over scaled Reference-free, consistent across three seeds. This supports the contribution of nearby-time information beyond uniform scaling.

We have clarified the seed aggregation, configurations, and ablation definitions in the revised manuscript. The controlled experiments further show that nearby-time information improves preservation at matched semantic alignment, while StableCtrl maintains comparable inference cost.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.