- 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
-
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.
-
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
-
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.
-
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.
-
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
- 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?
- 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
[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 and the reference residual at , with . 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
where is Gaussian noise. The SD-Turbo prediction is converted to a clean-latent estimate:
The semantic residual is the difference between target and source responses:
The reference residual is computed at 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 , the frequency coordinates are
The implemented spectral solution is
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.
[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:
The component-removal variants in Table 2 used (λ,γ,μ,η)=(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) 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=0.90, δ=0.15, one editing step, and no cleanup.
Table R1. Controlled reference ablation (mean ± std over three seeds).
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)=1+λ+μ+γ∥ω∥2μR^ref(ω).
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).
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.