acceptodds
Under review as a conference paper at ICLR 2027

SPECTRA: Source-Preserving Editing with Cross-Scale Transform and Representation Adaptation

Abstract

Text-guided image editing must follow the target instruction while retaining source information that should remain unchanged. Existing training-free methods use feature injection, attention control, or predefined frequency regions, but typically apply coarse or fixed preservation policies. We study source trajectories and find that preservation evidence differs across transform scales, is non-uniform across orientations within a scale, and changes over sampling time. SPECTRA therefore defines preservation over scale, direction, space, and time. At each native step, source-only multiscale transform energies determine threshold-free scale– direction weights, while target-token attention supplies spatial evidence through Semantic Structure Mass-Odds (SSMO). The fused clean prediction is returned to the backbone’s native DDIM or FlowMatch scheduler, enabling one interface across Stable Diffusion and transformer flow backbones; appearance-preservation tasks optionally use decoder-cycle residual chroma injection (DCRCI). Across two appearance-changing and two appearance-preserving benchmarks, SPECTRA shifts the preservation–editability operating point toward stronger source retention. Controlled analysis separates a strong cross-scale effect from a smaller conditional directional refinement: directional benefit tends to increase with source anisotropy, while a shuffled-direction control shows that structural preservation depends on matching source-derived weights to their corresponding bands.

Primary area:
generative models
Keywords:
Diffusion Models, Training-Free Image Editing, Source-Preserving Editing, Directional Heterogeneity, Adaptive Frequency Guidance, Multiscale Directional Representations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PMw8OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

SPECTRA is a training-free controller for source-preserving image editing. At every sampling step, it obtains clean predictions from a source-reference branch and a target-conditioned branch, decomposes them into multiscale directional subbands, and uses source statistics plus attention-based spatial gating to decide how much source information to retain. The fused clean prediction is then converted into the update expected by the backbone's native DDIM or FlowMatch scheduler. Across the four main benchmark panels, SPECTRA gives the strongest structural and appearance-preservation scores, but several baselines have higher CLIP-T, making the preservation-versus-editability trade-off central to the result.

Critical strengths
  • Framing preservation as allocation across scale, direction, space, and sampling time is intuitive, and the clean-prediction interface gives a concrete way to use the same controller with both diffusion and flow backbones.
  • The paper evaluates nine baselines, four backbones, two editing objectives, and several complementary preservation metrics. The scale-depth, direction-assignment, transform-family, and DCRCI ablations are more informative than a single end-to-end comparison.
  • Reproducibility is handled carefully. The paper fixes task manifests and paired settings, records model and scheduler configurations, and provides detailed pseudocode and diagnostic logging for the mechanism.
Critical weaknesses
  • SPECTRA has lower CLIP-T than several baselines in every main benchmark panel. This is consistent with stronger source retention, but it also means that part of the apparent gain may come from making a weaker edit. More direct human evaluation of edit success at matched preservation levels would help establish a better trade-off rather than a different operating point.
  • The cross-backbone results are not controlled component comparisons: SD3 and FLUX use strength 0.4 for SPECTRA and 0.6 for Vanilla, while the U-Net variants additionally use Self Attn-Conv. These results support portability of the interface, but not the size of the controller's causal contribution.
  • Within-scale direction adaptation produces modest DINO improvements, while Edge, CSD, and CLIP-T are mixed. The paper acknowledges this, but it makes the broad Scale × Direction framing stronger than the isolated evidence for the Direction component.
  • The method evaluates separate source and target branches and applies directional transforms at every selected step. GPU specifications are provided, but runtime, memory overhead, and the cost relative to the baselines are not.
Questions for authors

Please refer to the Weaknesses section.

4 comments
  1. SPECTRA has lower CLIP-T than several baselines in every main benchmark panel. This is consistent with stronger source retention, but it also means that part of the apparent gain may come from making a weaker edit. More direct human evaluation of edit success at matched preservation levels would help establish a better trade-off rather than a different operating point.

Preservation–editing trade-off. Table 5 in Section 4.9, also discussed in our response to Reviewer cBud, motivates our recommended balanced settings for Color40 and Style-9. Both improve CLIP-T over default SPECTRA through configuration changes while retaining its core approach of selectively preserving source information during editing.

Blinded human evaluation. We compared the recommended Style-9 setting with FIA-Edit because the two settings have nearly identical mean DINO over the full 80-case benchmark (Table 5). Eighteen participants evaluated 20 uniformly sampled PIE-Bench Style-9 cases, yielding 360 paired judgments. Participants saw the source image, the exact target instruction, and two anonymous outputs in randomized left–right order. They selected either output, both, or neither based on which edits clearly applied the requested style while preserving the source image’s shapes, poses, and layout.

Table 19. Blinded human evaluation on Style-9.

Method or contrastJoint edit-and-preservation success95% CI
FIA-Edit112/360 (31.1%)[17.8%, 44.7%]
Balanced SPECTRA135/360 (37.5%)[23.9%, 51.7%]
Difference (SPECTRA − FIA-Edit)+6.4 percentage points[−14.7, +27.5] percentage points

Each selected output counted as a success for its method. The intervals use 10,000 two-way bootstrap resamples over cases and participants, keeping the methods paired. The observed joint success rate is 6.4 percentage points higher for balanced SPECTRA, with a confidence interval that spans zero. Preservation equivalence on the 20 surveyed cases remains open.

Section 4.9 (“Preservation–Editing Trade-off”) summarizes the human result. Appendix M (“Blinded Human Evaluation on Style-9”) gives the survey protocol and reports the success rates and paired difference in Table 19.

  1. The cross-backbone results are not controlled component comparisons: SD3 and FLUX use strength 0.4 for SPECTRA and 0.6 for Vanilla, while the U-Net variants additionally use Self Attn-Conv. These results support portability of the interface, but not the size of the controller's causal contribution.

Matched img2img strength on SD3 and FLUX. We agree that the original strength mismatch complicated the comparison. We reran native SD3-Medium and FLUX.1-dev img2img at strength 0.4, matching the submitted SPECTRA runs within each backbone on the same benchmark cases.

Matched-strength comparisons from Table 3. All rows use img2img strength 0.4. DINO is scaled by 100. Bold marks the better result within each backbone pair.

(a) Color40 (n = 40).

BackboneMethodDINO ↓Edge ↑CLIP-T ↑
SD3-MediumNative img2img0.232800.958030.30236
SD3-MediumSPECTRA0.18450.9633700.299517
FLUX.1-devNative img2img0.214920.958650.29401
FLUX.1-devSPECTRA0.18700.9658610.293674

(b) Style-9 (n = 80).

BackboneMethodDINO ↓Edge ↑CLIP-T ↑CSD ↑ΔE₀₀ ↓LPIPS ↓
SD3-MediumNative img2img0.256940.956360.308290.800774.926910.22454
SD3-MediumSPECTRA0.16460.9698020.2964830.8731073.0540360.159466
FLUX.1-devNative img2img0.224160.957830.291710.881564.778110.22328
FLUX.1-devSPECTRA0.15860.9705010.2920880.9237503.2431540.170879

At matched strength, SPECTRA improves every reported preservation metric on both backbones. DINO decreases by 20.7% on SD3 and 13.0% on FLUX for Color40, and by 35.9% and 29.2%, respectively, for Style-9. CLIP-T is lower in three comparisons and slightly higher for FLUX on Style-9. Table 3 in Section 4.4 reports these comparisons and specifies the shared strength in its caption.

Role of Self Attn-Conv and controller attribution. The U-Net and flow variants share SPECTRA’s core controller, which combines source-derived scale–direction weighting with SSMO-guided fusion of clean predictions. Self Attn-Conv is an additional source-injection mechanism in the U-Net variants. The matched flow results show that SPECTRA improves preservation on both tasks without this injection.

To isolate the controller’s contribution on a fixed backbone, our matched SD2.1 2×2 ablation varies the controller and Self Attn-Conv independently. With injection held on, adding the controller reduces DINO by 25.6% on Color40 and 31.7% on Style-9, improves Edge on both, and lowers CLIP-T by 0.00608 and 0.00811, respectively. The controller also improves DINO and Edge with injection held off. Section 4.8 summarizes this ablation, and Appendix L.1 (“Full-controller factorial ablation”) reports the full results in Table 17.

  1. Within-scale direction adaptation produces modest DINO improvements, while Edge, CSD, and CLIP-T are mixed. The paper acknowledges this, but it makes the broad Scale × Direction framing stronger than the isolated evidence for the Direction component.

Scope of the Scale × Direction contribution. We agree that the isolated benefit of within-scale directional adaptation is modest. Tables 9–11 show small DINO improvements, while Edge, CSD, and CLIP-T results are mixed. In the full Color40 configuration (Table 13), adding directional adaptation to Scale-only slightly improves mean DINO and Edge, with both paired 95% confidence intervals spanning zero. Assigning the same weights to their corresponding bands improves DINO and Edge over shuffling them, with both intervals excluding zero and a small decrease in CLIP-T. Together, these results support directional adaptation as a modest, conditional refinement.

SPECTRA’s contribution is a controller that allocates source preservation across scales, directions, spatial locations, and denoising steps. The components play complementary roles. Scale weighting determines how much source information to retain at each resolution, directional weighting adjusts that contribution among bands within a scale, and SSMO uses target attention to determine where the contribution is applied at each step. Sections 4.6 and 4.7 describe the directional benefit at the magnitude supported by the isolated ablations.

Matched 2×2 controller ablation on SD2.1. To quantify the contribution of the controller as a whole, we independently toggle joint self-attention and convolutional-feature injection (I) and the complete controller (C), with 0 denoting off and 1 denoting on. The controller combines low-frequency fusion, adaptive scale–direction weighting, SSMO gating, and spectral reconstruction. All four conditions share the same source construction and sampling settings. AdaIN and DCRCI are off throughout Color40 and on throughout Style-9. When enabled, injection uses (at,ft)=(0.8,1.0)(a_t,f_t)=(0.8,1.0). Values are dataset means, and bold marks the best mean in each column.

Table 17(a). Color40, appearance changing (n = 40).

ConditionDINO ×100 ↓Edge ↑CLIP-T ↑
Neither (I = 0, C = 0)0.734960.886200.33317
Injection only (I = 1, C = 0)0.199380.963190.31689
Controller only (I = 0, C = 1)0.580960.903620.32918
Full SPECTRA (I = 1, C = 1)0.148320.970600.31081

Table 17(b). Style-9, appearance preservation (n = 80).

ConditionDINO ×100 ↓Edge ↑CLIP-T ↑CSD ↑ΔE₀₀ ↓LPIPS ↓
Neither (I = 0, C = 0)0.558990.917490.331940.606658.955880.39912
Injection only (I = 1, C = 0)0.224860.964300.311750.767925.166150.21300
Controller only (I = 0, C = 1)0.411800.934880.321770.697176.512980.30725
Full SPECTRA (I = 1, C = 1)0.153500.971380.303640.830144.306590.16164

Adding the complete controller to injection lowers mean DINO by 25.6% on Color40 and 31.7% on Style-9 and improves Edge on both datasets, with CLIP-T decreasing by 0.00608 and 0.00811, respectively. The controller also improves DINO and Edge with injection disabled. These gains quantify the contribution of the controller as a whole. The isolated direction ablations characterize the smaller, conditional effect of directional adaptation within it.

Section 4.8 summarizes the controller results, and Appendix L.1 (“Full-controller factorial ablation”) reports both panels in Table 17. Table H.1 reports the paired directional-assignment intervals directly below Table 13 in Appendix H.

  1. The method evaluates separate source and target branches and applies directional transforms at every selected step. GPU specifications are provided, but runtime, memory overhead, and the cost relative to the baselines are not.

Runtime and peak GPU memory. To quantify the cost of evaluating separate source and target branches and applying directional transforms during denoising, we profiled SPECTRA and the baselines on an RTX 3090 at batch size 1. The within-backbone comparisons with Vanilla measure the combined overhead of the SPECTRA pipeline. First-edit latency includes source preparation, editing, and final decoding, excluding model loading and image writing. For SD2.1, we also report inversion time and the latency of a subsequent edit that reuses the cached inversion. A dash indicates that a timing breakdown is unavailable.

Table 15. Computational cost on PIE-Bench Color40. Timings and per-case peak VRAM are averaged over 40 cases. FIA-Edit uses its native SD3.5-Medium backbone at the same resolution as the SD3-Medium runs.

MethodBackboneResolutionFirst edit (s) ↓Inversion (s) ↓Cached edit (s) ↓Peak VRAM (GiB) ↓
SPECTRASD2.1512 × 51296.7880.9715.764.51
VanillaSD2.1512 × 51282.7180.262.453.12
PnPSD2.1512 × 512107.6996.5311.194.51
SPECTRASD3-Medium1024 × 102436.96——18.78
VanillaSD3-Medium1024 × 102411.35——18.76
FIA-EditSD3.5-Medium1024 × 102441.34——20.09

On Color40, relative to Vanilla, SPECTRA adds 14.07 s and 1.39 GiB on SD2.1, and 25.61 s and 0.02 GiB on SD3-Medium. On SD2.1, SPECTRA completes the first edit 10.91 s faster than PnP at the same peak VRAM, primarily due to shorter inversion. At 1024 × 1024, SD3-Medium SPECTRA completes the first edit 4.38 s faster than FIA-Edit and uses 1.31 GiB less peak VRAM.

Table 16. Computational cost on PIE-Bench Style-9. Timings and per-case peak VRAM are averaged over 80 cases under the same profiling protocol. AdaIN and DCRCI are optional appearance components. The AdaIN-only setting disables DCRCI and its iterative VAE encoding and decoding.

MethodBackboneAppearance settingResolutionFirst edit (s) ↓Inversion (s) ↓Cached edit (s) ↓Peak VRAM (GiB) ↓
SPECTRASD2.1AdaIN + DCRCI512 × 512110.5782.3128.304.52
SPECTRASD2.1AdaIN only512 × 51297.5780.1617.454.52
VanillaSD2.1—512 × 51283.6981.262.433.12
PnPSD2.1—512 × 512107.5396.5011.044.51
SPECTRASD3-MediumAdaIN + DCRCI1024 × 102455.73——19.49
SPECTRASD3-MediumAdaIN only1024 × 102437.78——18.78
VanillaSD3-Medium—1024 × 102411.86——18.76
FIA-EditSD3.5-MediumNative1024 × 102442.36——20.09

On Style-9, disabling DCRCI while retaining AdaIN reduces first-edit latency by 13.00 s on SD2.1 and 17.95 s on SD3-Medium. With AdaIN alone, SPECTRA completes the first edit faster than PnP on SD2.1 and faster than resolution-matched FIA-Edit on SD3-Medium. The recommended balanced Style-9 setting in Table 5 of Section 4.9 also disables DCRCI and achieves better edge preservation than FIA-Edit at comparable target-prompt alignment.

The “Computational cost” paragraph in Section 4.9 summarizes the within-backbone overhead. Appendix K.1 reports both computational-cost tables and the profiling protocol.

Reviewer_cBudOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SPECTRA, a training-free image editing method that leverages multi-scale directional energy from the source image alongside attention maps derived from the target text. It adaptively fuses source information with editing-related content and employs an optional chromaticity correction step to enhance the preservation of the original appearance. While the method compatible with both diffusion and flow-matching models excels at retaining source information, it exhibits a slight decline in text alignment.

Critical strengths
  1. The motivation is clear. It shifts from preserving fixed frequency bands to an adaptive control mechanism that varies according to scale, orientation, space, and time.
  2. It achieves superior preservation of structure and appearance on evaluation datasets.
  3. It requires no training, supports various backbone architectures, and includes targeted ablation studies, such as the shuffling of orientation weights, to validate the approach.
Critical weaknesses
  1. Improvements in retention metrics are often accompanied by a decline in textual consistency. The potential advantages stemming from insufficient editing have not been ruled out, and there is a lack of comparison under conditions of equal editing success rates.
  2. The core orientation-adaptive module provides only marginal gains in the final configuration, and the attribution of benefits for the overall method remains insufficiently justified.
  3. The approaches used for SD3 and FLUX employ different image-to-image (img2img) strengths compared to the baselines, which compromises the fairness of the evaluation.
  4. SSMO relies on the total attention weight across all target tokens. The effectiveness of this approach for localizing regions to be preserved has not been sufficiently validated.
  5. Evaluations are primarily limited to color and style modifications. Furthermore, metrics based on source-style similarity may inadvertently reward results where the target style has not been fully realized.
  6. The absence of reports on runtime and VRAM usage makes it difficult to assess the additional overhead incurred by high-step inversion and iterative VAE encoding/decoding.
Questions for authors

Please see the weaknesses.

7 comments
  1. Improvements in retention metrics are often accompanied by a decline in textual consistency. The potential advantages stemming from insufficient editing have not been ruled out, and there is a lack of comparison under conditions of equal editing success rates.

Preservation–editing trade-off. We agree that stronger source retention can come at the cost of a weaker edit. To examine this trade-off, we evaluated several SPECTRA configurations alongside FIA-Edit on the same image–prompt pairs for Color40 and Style-9. We report CLIP-T as a proxy for target-prompt alignment and DINO and Edge as measures of structural preservation.

Color40 results from Table 5 (n = 40).

ConfigurationDINO ↓Edge ↑CLIP-T ↑
FIA-Edit0.177610.958810.31354
Full SPECTRA0.148320.970600.31081
SPECTRA with reduced injection0.246550.947810.31410
SPECTRA with scale-only adaptation0.14850.9700970.310990
SPECTRA with one NSST scale (J = 1) (recommended for balanced editing)0.186840.966020.31661

Style-9 results from Table 5 (n = 80).

ConfigurationDINO ↓Edge ↑CLIP-T ↑
FIA-Edit0.215360.953990.32392
Full SPECTRA0.153500.971380.30364
SPECTRA with reduced injection0.243170.959240.31023
SPECTRA with scale-only adaptation0.157720.970520.30381
SPECTRA with appearance alignment disabled0.162690.964230.31407
SPECTRA with one NSST scale (J = 1) and AdaIN only (recommended for balanced editing)0.215820.961520.32506

Values are dataset means. SPECTRA uses SD2.1, and DINO is scaled by 100. Bold scores mark the best mean in each column. Full SPECTRA uses three NSST scales and attention and convolutional injection windows (at,ft)=(0.8,1.0)(a_t,f_t)=(0.8,1.0). Reduced injection uses (0.5,0.8)(0.5,0.8). On Style-9, full SPECTRA uses AdaIN and DCRCI. AdaIN only disables DCRCI, and disabling appearance alignment turns off both modules. Appendix D specifies these configurations.

Recommended settings. The submitted configuration gives the strongest structural preservation on both datasets. Shorter injection, fewer NSST scales, and changes to appearance alignment offer different ways to favor prompt alignment. We recommend J = 1 on Color40 and J = 1 with AdaIN enabled and DCRCI disabled on Style-9. These settings achieve the highest mean CLIP-T in their respective panels and retain higher Edge than FIA-Edit. DINO distance is higher than FIA-Edit on Color40 and nearly identical on Style-9. Users can select these settings through the existing configuration file while retaining SPECTRA’s scale–direction weighting and SSMO-guided fusion.

Editing success. These results show an Edge advantage at similar or slightly higher mean CLIP-T. Retention at equal edit-success rates remains open. Answering that question directly would require judging whether the requested color or style change is visibly realized, reporting success rates over all cases, and comparing retention on the same cases where both methods succeed.

Section 4.9 (“Preservation–Editing Trade-off”) presents both comparisons in Table 5 and distinguishes prompt alignment from edit success. It also summarizes the blinded Style-9 evaluation, which is reported in Appendix M and Table 19.

  1. The core orientation-adaptive module provides only marginal gains in the final configuration, and the attribution of benefits for the overall method remains insufficiently justified.

Directional adaptation and the full controller. We agree that directional adaptation yields a small incremental gain in the full Color40 configuration. Table 13 shows slightly better mean DINO and Edge than Scale-only, with paired 95% confidence intervals spanning zero. Correct assignment also improves both metrics over shuffling the same weights across directional bands, with both intervals excluding zero. The isolated ablations in Tables 9–11 provide complementary evidence for directional adaptation as a modest, conditional refinement.

SPECTRA addresses the broader question of how to allocate source preservation across scales, directions, spatial locations, and denoising steps. At each step, scale and direction weighting determine how much source information each band contributes, and SSMO uses target attention to determine where that contribution is applied. These components serve complementary roles within the controller.

Matched 2×2 ablation on SD2.1. We independently toggle joint self-attention and convolutional-feature injection (I = 0/1, off/on) and the complete controller (C = 0/1, off/on). The controller combines low-frequency fusion, adaptive scale–direction weighting, SSMO gating, and spectral reconstruction. All four conditions share the same source construction and sampling settings. AdaIN and DCRCI are off throughout Color40 and on throughout Style-9. When enabled, injection uses (at,ft)=(0.8,1.0)(a_t,f_t)=(0.8,1.0). Values are dataset means, and bold marks the best mean in each column.

Table 17(a). Color40, appearance changing (n = 40).

ConditionDINO ×100 ↓Edge ↑CLIP-T ↑
Neither (I = 0, C = 0)0.734960.886200.33317
Injection only (I = 1, C = 0)0.199380.963190.31689
Controller only (I = 0, C = 1)0.580960.903620.32918
Full SPECTRA (I = 1, C = 1)0.148320.970600.31081

Table 17(b). Style-9, appearance preservation (n = 80).

ConditionDINO ×100 ↓Edge ↑CLIP-T ↑CSD ↑ΔE₀₀ ↓LPIPS ↓
Neither (I = 0, C = 0)0.558990.917490.331940.606658.955880.39912
Injection only (I = 1, C = 0)0.224860.964300.311750.767925.166150.21300
Controller only (I = 0, C = 1)0.411800.934880.321770.697176.512980.30725
Full SPECTRA (I = 1, C = 1)0.153500.971380.303640.830144.306590.16164

With injection enabled, adding the full controller lowers mean DINO by 25.6% on Color40 and 31.7% on Style-9, improves Edge on both, and lowers CLIP-T by 0.00608 and 0.00811, respectively. The controller also improves DINO and Edge with injection disabled. This ablation quantifies the contribution of the complete controller under both injection settings, while Table 13 measures the smaller effect of directional weighting within the full Color40 configuration.

Section 4.8 summarizes the controller results, and Appendix L.1 (“Full-controller factorial ablation”) reports both panels in Table 17. The paired directional-assignment intervals discussed in Section 4.6 appear in Table H.1, directly below Table 13 in Appendix H.

  1. The approaches used for SD3 and FLUX employ different image-to-image (img2img) strengths compared to the baselines, which compromises the fairness of the evaluation.

Matched img2img strength on SD3 and FLUX. We reran native SD3-Medium and FLUX.1-dev img2img at strength 0.4 and compared them with the submitted SPECTRA runs at the same strength on the same benchmark cases.

Matched-strength comparisons from Table 3. All rows use img2img strength 0.4. DINO is scaled by 100. Bold marks the better result within each backbone pair.

(a) Color40 (n = 40).

BackboneMethodDINO ↓Edge ↑CLIP-T ↑
SD3-MediumNative img2img0.232800.958030.30236
SD3-MediumSPECTRA0.18450.9633700.299517
FLUX.1-devNative img2img0.214920.958650.29401
FLUX.1-devSPECTRA0.18700.9658610.293674

(b) Style-9 (n = 80).

BackboneMethodDINO ↓Edge ↑CLIP-T ↑CSD ↑ΔE₀₀ ↓LPIPS ↓
SD3-MediumNative img2img0.256940.956360.308290.800774.926910.22454
SD3-MediumSPECTRA0.16460.9698020.2964830.8731073.0540360.159466
FLUX.1-devNative img2img0.224160.957830.291710.881564.778110.22328
FLUX.1-devSPECTRA0.15860.9705010.2920880.9237503.2431540.170879

At matched strength, SPECTRA improves every reported preservation metric on both backbones. CLIP-T is lower in three of the four comparisons and slightly higher for FLUX.1-dev on Style-9. Table 3 in Section 4.4 reports these matched comparisons and specifies the shared strength in its caption.

Reviewer_qUEXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SPECTRA is a training-free image-editing method that fuses source and target clean predictions in a multiscale directional transform. Source energies determine scale/direction weights, while target-token attention controls spatial fusion; the fused prediction is returned to the native diffusion or flow scheduler. Optional appearance components use adaptive normalization and decoder-cycle chroma correction. Experiments span four backbones and style/color benchmark subsets, with method comparisons, transform/component ablations, directional controls, and per-image statistical analyses. The paper reports improved source preservation while documenting preservation/editability trade-offs.

Critical strengths
  1. A clear, broadly instantiated editing interface. Clean source/target predictions are fused and returned to the native scheduler across four backbones, including diffusion and flow models. This demonstrates a useful implementation pattern. The disclosed .4-versus-.6 flow-strength comparison is an interface demonstration and should be interpreted within that stated scope.
  2. Meaningful controls support conditional preservation effects. Directional shuffling, isolated DCRCI, multiple transforms, and per-image sign tests with Wilson intervals provide evidence beyond selected images. These controls and the reported source-retention gains support a useful preservation mechanism, even though its benefit at comparable editing success is unresolved.
Critical weaknesses
  1. The task metrics do not resolve the editing-versus-preservation trade-off. Style-9 requires style changes (p.13), yet source CSD similarity is rewarded (p.14). An unchanged source would score well on source-preservation metrics; its CLIP-T baseline and explicit edit-success measurements are absent. SPECTRA has the lowest CLIP-T among Table 2 methods on both datasets. Qualitative successes do not establish the aggregate trade-off. FGD, Section 5.1, illustrates why source-style-preserving and source-structure-preserving tasks need different evaluation. This is not proof that SPECTRA generally ignores instructions, but it limits the main efficacy conclusion.
  2. The incremental benefit over the closest mechanisms is not sufficiently established. W-Edit, Section 3.3, Eq.10, already uses source-subband energy for adaptive frequency selection/fusion, but is neither discussed nor evaluated. Its thresholded rule differs from SPECTRA's continuous policy. Within the complete Color40 system, directional weighting changes the reported DINO distance (scaled by 100) from .1485 to .1483 and Edge from .970097 to .970599 while reducing CLIP-T (Table 13; scaling convention, p.5, line 252). This modest effect does not refute larger isolated effects, but leaves the complete new controller's practical advantage unclear without a focused matched comparison and paired uncertainty.
Questions for authors
  1. Assessment-critical analysis of existing outputs: Can source-identity baselines and explicit edit-success measurements show better retention at comparable successful editing, particularly on Style-9? Results separating successful style changes from failed edits could change my assessment.
  2. Closest-prior and component evidence: What benefit remains over W-Edit under matched task, backbone, and editing strength? Which existing controls isolate scale/SSMO contributions in the complete system, and what is the paired uncertainty of Table 13's directional effect? A focused matched comparison would be assessment-relevant, rather than a request for broad additional benchmarking.
2 comments
  1. Assessment-critical analysis of existing outputs: Can source-identity baselines and explicit edit-success measurements show better retention at comparable successful editing, particularly on Style-9? Results separating successful style changes from failed edits could change my assessment.

We agree that an unchanged source or a weak edit can score well on preservation metrics. To examine this trade-off, we evaluated several SPECTRA configurations alongside FIA-Edit on the same 80 Style-9 image–prompt pairs. We report CLIP-T as a proxy for target-prompt alignment and DINO and Edge as measures of structural preservation.

Style-9 results from Table 5 (n = 80). Values are dataset means. SPECTRA uses SD2.1. DINO is scaled by 100. Bold marks the best mean in each column.

ConfigurationDINO ↓Edge ↑CLIP-T ↑
FIA-Edit0.215360.953990.32392
Full SPECTRA0.153500.971380.30364
SPECTRA with reduced injection0.243170.959240.31023
SPECTRA with scale-only adaptation0.157720.970520.30381
SPECTRA with appearance alignment disabled0.162690.964230.31407
SPECTRA with one NSST scale (J = 1) and AdaIN only (recommended for balanced editing)0.215820.961520.32506

Users can move toward stronger prompt alignment in several ways. Reducing injection or disabling appearance alignment raises CLIP-T with different costs to DINO and Edge. Switching to scale-only adaptation leaves CLIP-T close to the full setting. For Style-9, we recommend J = 1 with AdaIN enabled and DCRCI disabled. Among the configurations shown, it has the highest CLIP-T. Compared with FIA-Edit, it gives higher Edge (0.96152 vs. 0.95399) and CLIP-T (0.32506 vs. 0.32392), with nearly equal DINO distance (0.21582 vs. 0.21536). Users can select this setting through the existing configuration file while retaining SPECTRA’s scale–direction weighting and SSMO-guided fusion.

Section 4.9 (“Preservation–Editing Trade-off”) presents the comparisons and recommended configurations for both Color40 (Task 1) and Style-9 (Task 2) in Table 5. It also summarizes the blinded Style-9 human evaluation. Appendix M and Table 19 give the protocol and joint edit-and-preservation success rates.

  1. Closest-prior and component evidence: What benefit remains over W-Edit under matched task, backbone, and editing strength? Which existing controls isolate scale/SSMO contributions in the complete system, and what is the paired uncertainty of Table 13's directional effect? A focused matched comparison would be assessment-relevant, rather than a request for broad additional benchmarking.

Comparison with W-Edit. Thank you for pointing us to W-Edit (Sun et al., ICLR 2026). Both methods use source subband energy to guide preservation. In W-Edit (§3.3, Eq. 10), a cumulative-energy threshold selects which wavelet bands to replace with source bands in intermediate attention features. SPECTRA uses source energy to assign continuous weights across scales and directions at each denoising step. Target attention then modulates the source contribution spatially, and the fused clean prediction is passed to the backbone’s native scheduler. This gives SPECTRA control over how much source information to retain in each directional band and where to retain it.

W-Edit’s main evaluation uses FLUX.1-dev with a wavelet-energy threshold of η = 0.6. We could not locate a public author implementation, so a head-to-head comparison at comparable editing strength remains open. Table 3 in Section 4.4 reports our comparison with native FLUX.1-dev img2img at a shared strength of 0.4. SPECTRA reduces DINO distance (×100) from 0.21492 to 0.18700 on Color40 and from 0.22416 to 0.15860 on Style-9. Edge improves on both datasets, with a slight CLIP-T decrease on Color40 and a slight increase on Style-9. Section 2 (“Related Work”) describes the methodological distinction from W-Edit.

Contribution of the full controller on Color40. We ran a matched 2×2 ablation on SD2.1 using the same 40 Color40 image–prompt pairs. I = 0/1 denotes joint source self-attention and convolutional-feature injection off/on. C = 0/1 denotes the complete spectral controller off/on. The controller applies low-frequency fusion, adaptive scale and direction weighting, SSMO gating, and spectral reconstruction. All four conditions share the same source construction, sampling settings, and SD2.1 implementation, with AdaIN and DCRCI disabled throughout.

Table 17(a). Matched 2×2 ablation on PIE-Bench Color40 (n = 40). Entries are means across the same image–prompt pairs.

ConditionDINO ×100 ↓Edge ↑CLIP-T ↑
Neither (I = 0, C = 0)0.734960.886200.33317
Injection only (I = 1, C = 0)0.199380.963190.31689
Controller only (I = 0, C = 1)0.580960.903620.32918
Full SPECTRA (I = 1, C = 1)0.148320.970600.31081

With injection enabled in both conditions, adding the controller reduces mean DINO distance by 25.6% and increases Edge by 0.00741, with a 0.00608 decrease in CLIP-T. Section 4.8 (“Contribution of the Full Controller”) summarizes the incremental effect, and Appendix L.1 (“Full-controller factorial ablation”) reports the complete results in Table 17.

Scale and SSMO controls. In the full Color40 system, the original SSMO gate yields slightly better mean DINO and Edge than the mean-matched constant and spatially shuffled gates, with similar CLIP-T (Appendix L.2, Table 18(a)). These controls test the contribution of spatial placement. Table 4 in Section 4.5 tests scale depth in the full Style-9 configuration. Increasing depth from one to three scales reduces mean DINO from 0.2029 to 0.1535. The effect of adaptive scale weighting at fixed depth remains open.

Paired uncertainty for Table 13. Table 13 tests whether assigning the same weights to their corresponding directional bands improves preservation. We used 10,000 paired bootstrap resamples of the 40 Color40 cases on SD2.1 at the original generation seed. Each cell reports the mean difference (Correct minus control) and its 95% confidence interval.

ComparisonΔ DINO ×100 ↓Δ Edge ↑Δ CLIP-T ↑
Correct − Shuffled−0.001324 [−0.002495, −0.000193]+0.000656 [+0.000042, +0.001316]−0.000576 [−0.000965, −0.000232]
Correct − Scale-only−0.000211 [−0.001868, +0.001629]+0.000503 [−0.000032, +0.001091]−0.000178 [−0.000546, +0.000200]

Correct assignment improves DINO and Edge relative to shuffled assignment, with both intervals excluding zero, alongside a small decrease in CLIP-T. Relative to Scale-only, the mean DINO and Edge differences also favor correct assignment, but all intervals include zero, so the incremental benefit remains uncertain. Section 4.6 summarizes these comparisons, and Table H.1 reports the intervals directly below Table 13 in Appendix H.

Comments on the paper

1 comment
AuthorsOpenReview ↗

**General Note to All Reviewers**

Note on the Revised Manuscript

We thank all reviewers for their constructive feedback. We have incorporated the additional experiments, analyses, and clarifications discussed in our responses into the revised manuscript. Throughout our rebuttal, we refer to the corresponding sections, tables, and appendices to help reviewers locate these updates. For clarity and completeness, we have addressed each reviewer's comments individually, even when similar concerns were raised by multiple reviewers. This approach ensures that each reviewer can directly follow our responses to their specific questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.