acceptodds
Under review as a conference paper at ICLR 2027

S2AR: Dual-Axis Routing for Continuous Attribute Control in Visual Autoregressive Models

Abstract

Visual autoregressive models have emerged as a powerful and scalable paradigm for image synthesis. However, continuous attribute control within this framework remains underexplored. To address this gap, we present S2AR, a training-free framework for continuous attribute control. Specifically, we formulate attribute control as a dual-axis routing problem across spatial and scale dimensions. S2AR first identifies target visual tokens to construct a spatial localization map, and subsequently determines scale-wise routing weights based on structural sensitivity. Guided by these routing weights, a single control scalar jointly modulates cross-attention and global conditioning, thereby continuously steering attribute intensity. On our constructed Continuous Attribute Control Benchmark, S2AR consistently outperforms existing training-free and LoRA-based methods while faithfully preserving identity and background details.

Primary area:
generative models
Keywords:
Visual Autoregressive Models, Continuous Attribute Control, Text-to-Image

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_f6sXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes S2AR, a training-free method for continuous attribute control in visual autoregressive models. Experimental results show effectiveness in content preservation and monotonic attribute control.

Critical strengths
  1. This paper addresses continuous attribute control in visual autoregressive models. The research problem is clear.

  2. This method adjusts the editing regions and generation scales according to different attributes.

Critical weaknesses
  1. The results of the Ours differ between Table 1 and Table 2.

  2. The scale routing weights are obtained from the reconstruction error of spatial response maps. They do not directly measure how much each generation scale affects the target attribute.

Questions for authors
  1. Please explain why the full-method results differ between Table 1 and Table 2.

  2. Please provide sensitivity analysis of scale. It would be helpful to compare the proposed scale weights with fixed-scale intervention and simpler scale weighting methods.

2 comments

(Part 1/2)

We sincerely appreciate your recognition of our work and your constructive suggestions. We have carefully considered your feedback and address each of your comments below.

Q1. Discrepancy Between Table 1 and Table 2

We thank the reviewer for identifying this inconsistency and acknowledge that the original manuscript did not clearly specify the differing evaluation settings. Table 1 used all 168 CACBench prompts (8 per attribute), whereas Table 2 used 42 randomly sampled prompts (2 per attribute) to reduce the computational cost of extensive analysis experiments. Thus, the results were not directly comparable. To resolve this, we reran the ablations on all 168 prompts, updated Table 2 with task-specific results, and revised the corresponding analysis. The updated ablations show that removing spatial routing reduces controllability and structural preservation, as editing signals are no longer confined to attribute-relevant regions. Removing adaptive scale routing substantially weakens controllability, as a fixed injection schedule cannot accommodate attribute-specific scale requirements. For ease of reference, we provide the revised Table below (refer to Table 2 in the manuscript):

Table 1: Ablation study on all CACBench prompts across three tasks. (Best in bold, second best underlined.)

ConfigurationExp ΔCLIP ↑Exp ρ ↑Exp Rev. ↓Exp DINO ↑Exp LPIPS ↓Obj ΔCLIP ↑Obj ρ ↑Obj Rev. ↓Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn ρ ↑Scn Rev. ↓Scn DINO ↑Scn LPIPS ↓
Full method0.0480.9080.1420.9410.2380.0370.8880.1580.8240.2180.0680.8900.1390.7650.286
w/o Spatial0.0350.8120.2180.9080.2990.0200.6120.3000.7750.3880.0550.7840.1850.6980.480
w/o Scale0.0180.7000.2630.9740.0880.0020.2450.4580.9580.1030.0060.3790.4060.9210.152
w/o 1×1 Lock0.0330.7670.2410.9280.2660.0130.5110.3540.8040.3700.0470.7180.2390.7480.443
w/o Attr Context0.0190.5470.3290.9650.2030.0040.1570.4480.8780.2510.0120.4520.3610.8510.317

(Part 2/2)

Q2. Sensitivity Analysis of Scale

To validate the effectiveness of our adaptive scale weighting strategy, we conducted two additional sets of comparative experiments: (1) Fixed-scale Intervention, where editing interventions are applied to different fixed scale ranges; and (2) Scale Weighting, including uniform weights and linearly increasing weights. We evaluate these settings on three editing tasks using ΔCLIP, DINO, and LPIPS as evaluation metrics.

First, regarding fixed-scale interventions, as the intervention shifts from lower-indexed to higher-indexed scales, ΔCLIP generally decreases across all three tasks, while DINO consistently increases and LPIPS decreases. This suggests that higher-indexed scales tend to preserve the original image content more effectively but struggle to achieve sufficient semantic editing. The optimal fixed scale range varies across tasks, indicating that the effectiveness of fixed-scale interventions is task-dependent. In contrast, our method achieves the highest ΔCLIP across all three tasks, demonstrating that adaptive scale allocation enables more effective semantic editing without requiring manual selection of task-specific scale ranges.

Table 2: Sensitivity analysis of fixed-step interventions across three tasks.

SettingExpr. ΔCLIP ↑Expr. DINO ↑Expr. LPIPS ↓Objects ΔCLIP ↑Objects DINO ↑Objects LPIPS ↓Scenes ΔCLIP ↑Scenes DINO ↑Scenes LPIPS ↓
Ours0.0480.9410.2380.0370.8240.2180.0680.7650.286
Steps 2–40.0010.9580.2620.0180.8350.3560.0400.7670.435
Steps 5–70.0180.9740.0880.0020.9580.1030.0060.9210.152
Steps 8–100.0070.9960.0200.0010.9910.0310.0010.9750.055
Steps 11–130.0010.9990.0040.0010.9980.0080.0000.9930.016

Second, regarding scale weighting strategies, our adaptive weighting strategy achieves better content preservation while improving semantic editing performance. Although linearly increasing weights perform better on certain content preservation metrics, their semantic editing capability is notably limited. These results indicate that simple predefined weighting schemes cannot adequately accommodate the editing requirements of different inputs.

Table 3: Comparison of different scale weighting strategies across three tasks.

SettingExp. ΔCLIP ↑Exp. DINO ↑Exp. LPIPS ↓Obj ΔCLIP ↑Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn DINO ↑Scn LPIPS ↓
Ours0.0480.9410.2380.0370.8240.2180.0680.7650.286
Uniform weights0.0340.9200.2690.0170.8090.3600.0480.7500.440
Linear weights0.0160.9720.1900.0030.9070.2250.0060.8810.254

We have added these comparative experiments and the corresponding sensitivity analysis results to the appendix. In addition, we have included a scale sensitivity analysis for the localization stage to further evaluate the impact of different scale selections on localization performance.

Reviewer_j5VQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes S2AR (Spatial-Scale Attribute Routing), a training-free continuous attribute control framework for visual autoregressive models (VARs), designed to adjust the intensity of specific semantic attributes in generated images while preserving the identity, structure, and background information of the original images as much as possible. The authors point out that different semantic attributes exhibit distinct spatial distributions and scale dependencies during the autoregressive generation process; directly modifying the text prompt may cause the generation trajectory to deviate, thereby affecting non-target regions. To address this issue, S2AR proposes an attribute routing mechanism across two dimensions: spatial and scale. Specifically, spatial routing identifies editing regions related to the target attribute by comparing the differences in model predictions between the original and target text conditions; scale routing adaptively determines the intensity of attribute intervention based on the reconstruction fidelity of the spatial response map at different generation scales. Ultimately, S2AR achieves continuous attribute editing driven by a single control parameter through local cross-attention modulation and global conditional interpolation.

Critical strengths
  1. The research question has practical value, and the methodological design is clearly motivated. This paper focuses on attribute control in visual autoregressive models, investigating how to preserve non-target content while adjusting local or global semantic attributes. Building on the characteristics of multiscale autoregressive generation, the authors model attribute control as two interrelated problems and propose a relatively comprehensive training-free control framework.
  2. The experimental results demonstrate good control over properties and the ability to preserve content.
Critical weaknesses
  1. Presentation: "Scale-Aware Response Routing" in Figure 2 is broken.

  2. Limited methodological novelty and unverified necessity of adaptive scale routing. S2AR's core components closely mirror existing training-free visual autoregressive (VAR) editing techniques: AREdit [1] already leverages source-target prediction distribution differences to localize edits, BitResEdit [2] combines these differences with localized multi-scale residual editing in bitwise AR models, and SSG [3] explores training-free scale-aware guidance. S2AR's primary distinction, deriving scale-wise routing weights from the reconstruction fidelity of spatial response maps, is essentially a resolution-based heuristic whose advantage over simpler scheduling strategies remains unproven. Because Table 2 only ablates scale routing entirely without comparing against fixed weights or linear schedules, the experiments show that scale-dependent control helps, but fail to justify the proposed adaptive formulation.

  3. Missing essential baselines for validating the effectiveness of dual-axis routing. The compared baselines, Concept Sliders [4], AREdit [1], and BitResEdit [2], were designed for either general text-guided editing [1, 2] or diffusion models [4]. Although implemented on a shared Infinity-2B backbone, how these methods were adapted for continuous attribute control is under-explained, obscuring whether the comparison is fair. More importantly, the paper omits simple, direct baselines for inference-time attribute control, such as source-target text embedding interpolation, logit-difference scaling, or unrouted cross-attention difference injection. Without comparing against these straightforward alternatives, the necessity of the proposed dual-axis routing remains unsubstantiated.

  4. Several core design choices rely on unverified heuristics:

    1. Fixed localization scales: Spatial localization aggregates responses from a fixed scale set {5, 6, 7} without sensitivity analysis across varying attribute types, object sizes, or model capacities.
    2. Representability vs. controllability: Scale routing assumes that spatial reconstruction fidelity indicates suitability for semantic intervention. However, spatial representability does not imply semantic controllability, and no controlled experiments validate this link.
    3. Spatial coverage as globalness: Deriving the globalness score from response coverage conflates spatial extent with semantic globalness. Moreover, the local cross-attention modulation and global conditioning pathways are not ablated separately to isolate their individual contributions.
  5. There exists some inconsistency in experiment data. In Table 1, S2AR achieved Spearman’s correlation coefficients of 0.908, 0.888, and 0.890 for the “Faces,” “Objects,” and “Scenes” tasks, respectively, while the corresponding results for the “Full Method” in Table 2 were 0.735. Furthermore, the DINO similarity of the “Full Method” reported in Table 2 is 0.945, which is also higher than the values of 0.941, 0.824, and 0.765 reported for the three tasks in Table 1, respectively. These differences may stem from different evaluation subsets, sample selection, or statistical methods; however, the paper does not currently provide a clear explanation of the composition of the data in Table 2 or its relationship to Table 1.

Questions for authors
  1. Regarding the necessity of the scale-routing mechanism. This paper uses the reconstruction error of the spatial response map to determine the scale-routing weights. Could the authors further explain the advantages of this design over simple strategies such as fixed-scale weights and linearly increasing scheduling, and provide corresponding comparative experiments?
  2. Regarding the Adequacy of the Baseline. Could the authors supplement the paper with a simple, training-free baseline based on the same Infinity backbone—such as text-conditional interpolation, differential scaling of predicted logits, or differential injection of cross-attention—to validate the actual benefits of the dual-axis routing mechanism relative to basic attribute control strategies?
  3. Regarding the rationality of key design choices. Spatial routing aggregates the predicted responses at the 5th, 6th, and 7th scales, while the suitability of a scale is measured using the reconstruction fidelity of spatial responses. Could the authors provide sensitivity experiments on scale selection and further validate the relationship between spatial reconstruction fidelity and actual semantic control performance?
  4. Regarding the consistency of experimental results. In Table 1, the Spearman correlation coefficients for S2AR across the three tasks are all close to 0.9, whereas the result for the Full Method in Table 2 is 0.735. Could the authors clarify whether there are differences in the evaluation data, statistical methods, or experimental settings between the two tables, and explain this apparent discrepancy in the results?
8 comments

(Part 1/5)

We sincerely appreciate your careful and thoughtful review of our work. Your valuable feedback has helped us further strengthen our methodological justification and experimental analysis. We provide detailed responses to each of your comments below.

Q1. Necessity of Scale-Aware Routing

To validate the advantages of our scale-routing mechanism over simpler strategies, we conducted two sets of comparative experiments: editing interventions at different fixed scale ranges, and scale weighting strategies using uniform and linearly increasing weights.

Table 1: Comparison of different scale weighting strategies across three tasks.

SettingExp. ΔCLIP ↑Exp. DINO ↑Exp. LPIPS ↓Obj ΔCLIP ↑Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn DINO ↑Scn LPIPS ↓
Ours0.0480.9410.2380.0370.8240.2180.0680.7650.286
Uniform weights0.0340.9200.2690.0170.8090.3600.0480.7500.440
Linear weights0.0160.9720.1900.0030.9070.2250.0060.8810.254

Scale Weighting Strategies. Compared with uniform and linearly increasing weights, our adaptive weighting strategy achieves higher ΔCLIP across all three tasks, while also outperforming uniform weighting in both DINO and LPIPS. Although linear weighting performs better on some content preservation metrics, its semantic editing capability is substantially weaker. These results demonstrate that adaptive weight allocation based on the reconstruction fidelity of spatial response maps enables more effective attribute control than predefined weighting strategies while maintaining good content fidelity.

Table 2: Sensitivity analysis of fixed-step interventions across three tasks.

SettingExpr. ΔCLIP ↑Expr. DINO ↑Expr. LPIPS ↓Objects ΔCLIP ↑Objects DINO ↑Objects LPIPS ↓Scenes ΔCLIP ↑Scenes DINO ↑Scenes LPIPS ↓
Ours0.0480.9410.2380.0370.8240.2180.0680.7650.286
Steps 2–40.0010.9580.2620.0180.8350.3560.0400.7670.435
Steps 5–70.0180.9740.0880.0020.9580.1030.0060.9210.152
Steps 8–100.0070.9960.0200.0010.9910.0310.0010.9750.055
Steps 11–130.0010.9990.0040.0010.9980.0080.0000.9930.016

Fixed-Scale Interventions. As the intervention shifts from earlier to later scales, DINO progressively increases and LPIPS decreases, while semantic editing strength (ΔCLIP) generally declines. This indicates that later scales better preserve the original image content but struggle to produce sufficient attribute changes. Moreover, the optimal fixed intervention range is Steps 5–7 for Expressions and Steps 2–4 for Objects and Scenes, indicating that different tasks have different scale requirements. In contrast, S2AR achieves higher ΔCLIP than all fixed-scale configurations across all three tasks, further demonstrating the effectiveness of adaptive scale allocation for attribute editing.

(Part 2/5)

Q2. Adequacy of Baselines

To validate the advantages of dual-axis routing over simple attribute control strategies, we introduced three additional training-free baselines using the same frozen Infinity-2B backbone: (1) Pooled text-conditioning interpolation, which interpolates between the original and edited text conditions; (2) Logit-difference scaling, which directly scales the predicted logit differences; and (3) Unrouted cross-attention difference injection, which directly injects cross-attention output differences without spatial or scale routing. All methods were evaluated on the full CACBench under identical experimental settings.

Table 3: Comparison with simple training-free baselines on CACBench.

MethodExp ΔCLIP ↑Exp ρ ↑Exp Rev. ↓Exp DINO ↑Exp LPIPS ↓Obj ΔCLIP ↑Obj ρ ↑Obj Rev. ↓Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn ρ ↑Scn Rev. ↓Scn DINO ↑Scn LPIPS ↓
Pooled text-conditioning interpolation-0.001-0.0380.5040.9810.1820.000-0.0250.4980.9180.2030.0000.0400.5060.8980.223
Logit-difference scaling0.0370.8240.2170.8610.3570.0250.6470.2970.7370.4200.0580.7890.1880.6730.502
Unrouted cross-attention difference injection0.0370.7950.2170.8660.3570.0250.6680.2780.7420.4210.0580.7840.1930.6740.502
S2AR0.0480.9080.1420.9410.2380.0370.8880.1580.8240.2180.0680.8900.1390.7650.286

Experimental Results. Pooled text-conditioning interpolation produces almost no effective attribute changes, and its high content preservation scores largely reflect the limited editing effect. In contrast, logit-difference scaling and unrouted cross-attention difference injection produce noticeable attribute changes but underperform S2AR in both editing effectiveness and content preservation. Specifically, S2AR achieves the highest ΔCLIP and ρ and the lowest Rev. across all three tasks, while also consistently outperforming these two direct intervention strategies in DINO and LPIPS. These results demonstrate that dual-axis routing enables more effective and monotonic attribute control while better preserving the original image content.

Baseline Adaptations. We have also added detailed implementation descriptions to the Appendix explaining how existing baselines are adapted for continuous attribute control. Specifically, for AREdit, BitResEdit, and Concept Sliders-style LoRA, we enable continuous control by adjusting the editing threshold, scaling the spatial gate, and scaling the LoRA weight updates, respectively. All methods are evaluated using the same CACBench prompts, random seeds, five control levels, and evaluation metrics to ensure consistent evaluation and fair comparison. We have also included ATA in the revised Table 1. Since the GitHub repository linked in the ATA paper remains inaccessible (404 error) and the paper provides insufficient architectural details for exact reproduction, we independently implemented a best-effort reproduction based on its methodological descriptions. At the request of another reviewer, we have also added precomputed results for the baseline methods to our interactive demo, allowing reviewers to directly compare these baselines with S2AR under matched settings.

(Part 3/5)

Q3. Justification of Key Design Choices

To validate the scale selection for spatial response map construction, we conducted additional sensitivity experiments using alternative scale ranges, including {2–4}, {8–10}, {11–13}, and {1–13}, compared with the default {5–7}. We also evaluated 21 individual attributes to further examine the performance and limitations of our method across different attributes.

Scale Selection Sensitivity Analysis. The results show that different scale ranges exhibit varying trade-offs between attribute control and content preservation. Although some scale configurations achieve higher ΔCLIP or better control monotonicity, they generally introduce greater changes to the original image content. In contrast, the default scales {5–7} achieve the lowest LPIPS across all three tasks while maintaining good attribute control and monotonicity. These results indicate that {5–7} provides a reasonable balance between attribute editing and content preservation, supporting its selection as the default scale range for spatial response map construction.

Table 4: Comparison of different scale ranges for region identification. (Best in bold, second best underlined‾\underline{\text{underlined}})

ScalesExp ΔCLIP ↑Exp ρ ↑Exp Rev. ↓Exp DINO ↑Exp LPIPS ↓Obj ΔCLIP ↑Obj ρ ↑Obj Rev. ↓Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn ρ ↑Scn Rev. ↓Scn DINO ↑Scn LPIPS ↓
Scales {2–4}0.0480.8790.1620.9070.3080.0430.9940.1170.7980.2500.0750.9100.1260.7430.322
Scales {5–7} (Default)0.0480.9080.1420.9410.2380.0370.8880.1580.8240.2180.0680.8900.1390.7650.286
Scales {8–10}0.0470.8790.1360.9400.2470.0350.8650.1690.8270.2230.0700.8760.1450.7610.294
Scales {11–13}0.0460.8680.1530.9310.2740.0340.8580.1760.8290.2260.0670.8790.1570.7700.290
Scales {1–13}0.0500.9040.1220.9160.2820.0410.9690.1420.8000.2500.0760.9210.1170.7330.324

Spatial Reconstruction Fidelity and Semantic Controllability. Our motivation for using Response Map Fidelity is that different scales vary in their ability to represent attribute-related spatial response structures. Due to the limited spatial resolution of coarser scales, the response structures of localized attributes may not be adequately preserved, making it difficult to confine editing interventions to the target regions. Therefore, we use the reconstruction fidelity of spatial response maps to assess how well each scale preserves attribute-related spatial structures and accordingly reduce intervention weights at scales with insufficient spatial representational capacity, thereby limiting potential effects on non-target regions. Importantly, we do not assume that higher reconstruction fidelity necessarily implies stronger semantic editing capability; we use it only as a basis for assigning scale-wise intervention weights. The results in Table 1 further show that adaptive weighting based on Response Map Fidelity achieves higher ΔCLIP than both uniform and linearly increasing weights across all three tasks, providing empirical support for the effectiveness of this design.

Reviewer_m9CcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes S2AR, a training-free method for continuous attribute control in scale-wise visual autoregressive text-to-image models (Infinity-2B/8B). Given a base prompt, an attribute word, and a strength α∈[0,1], the method generates an anchor image, then compares model logits under the base and attribute-augmented prompts along the anchor trajectory to build a spatial editing mask from the bit-flipping component of the difference. A per-scale weight is derived from how well this mask is preserved under downsampling to each scale’s resolution, and a globalness score from its spatial spread. At generation time, the attribute signal is injected into cross-attention outputs within the mask and, for global attributes, through interpolation of prompt embeddings, both scaled by α and the per-scale weights. The authors introduce CACBench (21 attributes, three categories) and compare against AREdit, BitResEdit, and a LoRA slider on the same backbone using ΔCLIP, Spearman ρ, reversal rate, DINO, and LPIPS, reporting the best monotonicity and preservation scores across categories, along with ablations, a transfer to Infinity-8B, and efficiency measurement.

Critical strengths

1: The problem is well motivated: continuous attribute control is underexplored in VAR models, and the paper gives a clear reason why coarse-to-fine generation requires scale-aware intervention.

2: The routing design is simple and interpretable. The bit-flip-directed response (Eq. 4), the downsampling-based scale weight (Eq. 5), and the globalness gate (Eq. 6) are parameter-free and handle both local and global attributes with one mechanism.

3: The method is training-free, needs no reference image, is cheaper than all baselines (Table 3), and transfers to Infinity-8B without retuning.

4: Fig. 4 and the ablations in Table 2 support the central claim that spatial routing drives preservation and scale routing drives attribute strength and monotonicity.

Critical weaknesses

1: Presentation. I have to be honest: the writing and figures gave me the impression that the paper was put together in a hurry. In Fig. 2 the boxes and arrows are visibly misaligned, and the labels don’t match the equations in Sec. 4. Fig. 1 doesn’t say which scale each column is. Fig. 4(b) is too small to read and uses terms (“semantic pressure”, “grid fidelity”) that never appear in the text. The intro and conclusion say you inject “prediction differences”, but Sec. 4.3 actually injects cross-attention shifts. None of these is fatal on its own, but together they made it hard to trust the rest.

2: Amount of experiments. The core experiments are fine as far as they go, but there isn’t much there: one main table with no standard deviations, one ablation table whose numbers don’t match Table 1 (on what subset?), one qualitative figure, one efficiency table. I’d want to see at least a sensitivity check on the hard-coded scales {5,6,7} and the thresholding, per-attribute results, some failure cases, and a comparison with ATA, which is the only prior slider method for VAR models and is cited but never compared.

3: The benchmark is private. Every number in the paper comes from CACBench, but I can’t find the attribute list, the prompts, the seeds, or how α is fed to AREdit, BitResEdit, and the LoRA baseline, and there’s no promise to release any of it. Without that, I genuinely cannot tell how much the margins over the baselines are worth. Please release the benchmark.

4: I’d like to try it. The demos in Fig. 3 and Fig. 5 look good, honestly better than I expected. But with one example per task, no code, and no hyperparameters, I can’t rule out that they’re cherry-picked or hand-tuned. The nice thing about a training-free method on a public checkpoint is that this is easy to settle: an anonymous code release or a small interactive demo (a Hugging Face Space would do) would let reviewers drag the slider themselves. With the help of agents (codex or claude code), that can be provided easily and if it works as shown, I’d be happy to raise my score.

Questions for authors

Please see the weaknesses.

To sum up: I think the method is practical and the idea has real value, but the paper as written is not in good shape. I would be willing to support acceptance if three things happen:

1: The authors commit to a substantial rewrite and redraw the figures. I’ll be candid that some reviewers and ACs will still hold the current presentation against the paper, and a cleaned-up resubmission to the next venue might serve the work better than a rushed camera-ready.

2: The other reviewers do not find further inconsistencies or problems beyond the ones I listed.

3: During the rebuttal the authors release an anonymous demo, and the reviewers who try it agree that the control effect is as impressive as the figures suggest. If that happens, I will gladly raise my score.

6 comments

(Part 1/4)

We sincerely thank you for your careful review of our work. Following your suggestions, we have systematically improved the manuscript's presentation, experimental validation, and reproducibility materials. In particular, we have released an anonymous interactive demo that allows reviewers to directly compare our method with baselines under matched settings.

Q1. Presentation Problems

To improve clarity, we made the following revisions to the manuscript:

  1. Figure 1. We added explicit scale indices to each column to provide a clearer reference.
  2. Figure 2. We redrew Figure 2 to present a more intuitive and structured overview of our framework pipeline.
  3. Figure 4(b). We have enlarged Fig. 4(b) for better readability and replaced the ambiguous terms with “Mean Directional Response” and “Response Map Fidelity,” which are now clearly explained in the revised manuscript.
  4. Injection Mechanisms. We clarified the injection mechanism in the revised manuscript: we inject cross-attention output shifts during generation, and prediction differences are used to construct the spatial localization map. We have revised the relevant statements in the Introduction and Conclusion to make them consistent with Sec. 4.3.
  5. Standard Deviations. We added standard deviations across five random seeds to the revised Table 1.

Q2. Extensive Analyses

2.1 Discrepancy Between Table 1 and Table 2

We thank the reviewer for identifying this inconsistency and acknowledge that the original manuscript did not clearly specify the differing evaluation settings. Table 1 used all 168 CACBench prompts (8 per attribute), whereas Table 2 used 42 randomly sampled prompts (2 per attribute) to reduce the computational cost of extensive analysis experiments. Thus, the results were not directly comparable. To resolve this, we reran the ablations on all 168 prompts, updated Table 2 with task-specific results, and revised the corresponding analysis. The updated ablations show that removing spatial routing reduces controllability and structural preservation, as editing signals are no longer confined to attribute-relevant regions. Removing adaptive scale routing substantially weakens controllability, as a fixed injection schedule cannot accommodate attribute-specific scale requirements. For ease of reference, we provide the revised Table below (refer to Table 2 in the manuscript):

Table 1: Ablation study on all CACBench prompts across three tasks. (Best in bold, second best underlined.)

ConfigurationExp ΔCLIP ↑Exp ρ ↑Exp Rev. ↓Exp DINO ↑Exp LPIPS ↓Obj ΔCLIP ↑Obj ρ ↑Obj Rev. ↓Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn ρ ↑Scn Rev. ↓Scn DINO ↑Scn LPIPS ↓
Full method0.0480.9080.1420.9410.2380.0370.8880.1580.8240.2180.0680.8900.1390.7650.286
w/o Spatial0.0350.8120.2180.9080.2990.0200.6120.3000.7750.3880.0550.7840.1850.6980.480
w/o Scale0.0180.7000.2630.9740.0880.0020.2450.4580.9580.1030.0060.3790.4060.9210.152
w/o 1×1 Lock0.0330.7670.2410.9280.2660.0130.5110.3540.8040.3700.0470.7180.2390.7480.443
w/o Attr Context0.0190.5470.3290.9650.2030.0040.1570.4480.8780.2510.0120.4520.3610.8510.317

(Part 2/4)

2.2 Sensitivity Analysis of Scale Selection for Region Identification

To justify our choice of scales {5–7} for region identification, we conducted additional sensitivity experiments with alternative scale configurations, including {2–4}, {8–10}, {11–13}, and the full scale range {1–13}, across three tasks: Expressions, Objects, and Scenes. The results show that as the selected scales shift from finer to coarser scales, ΔCLIP increases across all three tasks, ρ generally improves, while DINO decreases and LPIPS tends to increase. This indicates that coarser scales enhance attribute controllability at the cost of content preservation. Using the full scale range {1–13} also achieves strong attribute control and high monotonicity, but exhibits relatively weaker content preservation. In contrast, our default intermediate scales {5–7} achieve the lowest LPIPS across all three tasks while maintaining high control monotonicity, demonstrating a better balance between attribute controllability and content preservation.

Table 2: Comparison of different scale ranges for region identification. (Best in bold, second best underlined‾\underline{\text{underlined}})

ScalesExp ΔCLIP ↑Exp ρ ↑Exp Rev. ↓Exp DINO ↑Exp LPIPS ↓Obj ΔCLIP ↑Obj ρ ↑Obj Rev. ↓Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn ρ ↑Scn Rev. ↓Scn DINO ↑Scn LPIPS ↓
Scales {2–4}0.0480.8790.1620.9070.3080.0430.9940.1170.7980.2500.0750.9100.1260.7430.322
Scales {5–7} (Default)0.0480.9080.1420.9410.2380.0370.8880.1580.8240.2180.0680.8900.1390.7650.286
Scales {8–10}0.0470.8790.1360.9400.2470.0350.8650.1690.8270.2230.0700.8760.1450.7610.294
Scales {11–13}0.0460.8680.1530.9310.2740.0340.8580.1760.8290.2260.0670.8790.1570.7700.290
Scales {1–13}0.0500.9040.1220.9160.2820.0410.9690.1420.8000.2500.0760.9210.1170.7330.324

2.3 Sensitivity Analysis of Thresholding qq

We evaluate the impact of the spatial response normalization quantile qq across all three tasks. As shown in the table, as qq increases from 0.80 to 0.99, DINO consistently increases and LPIPS decreases across all three tasks, while ΔCLIP\Delta\mathrm{CLIP} exhibits an overall decreasing trend. This indicates that a higher qq more strictly restricts the editing regions, improving content preservation at the cost of attribute editing effectiveness. Furthermore, increasing qq from 0.80 to 0.95 substantially reduces LPIPS with only minor changes in ΔCLIP\Delta\mathrm{CLIP}, suggesting that content preservation can be significantly improved with minimal loss in editing strength. In contrast, further increasing qq to 0.99 improves content preservation metrics but also reduces semantic editing strength. Therefore, we select q=0.95q=0.95 as the default setting to achieve a favorable balance between attribute editing effectiveness and content preservation.

Table 3: Effect of spatial response normalization quantile qq across three tasks. (Best in bold, second best underlined‾\underline{\text{underlined}})

Quantile qExp ΔCLIP ↑Exp DINO ↑Exp LPIPS ↓Obj ΔCLIP ↑Obj DINO ↑Obj LPIPS ↓Scn ΔCLIP ↑Scn DINO ↑Scn LPIPS ↓
q=0.800.0480.8960.2780.0410.8100.2400.0740.7220.318
q=0.900.0480.9100.2630.0390.8130.2330.0720.7440.303
q=0.95 (Default)0.0480.9410.2380.0370.8240.2180.0680.7650.286
q=0.990.0420.9790.2140.0350.8600.2060.0620.8030.260

(Part 3/4)

2.4 Per-attribute results

We report the results for individual attributes to examine the performance and limitations of S2AR across different attributes. Specifically, for expression editing, surprise and anger exhibit stronger semantic editing effects than sadness and fear. For object editing, surface attributes such as rust and wetness can be effectively edited while preserving the original object content, whereas attributes involving structural changes, such as cracks, tend to cause greater deviations from the original appearance. For scene editing, visually distinctive attributes such as snow, winter, and fog exhibit stronger editing effects, while less visually distinctive attributes such as wind, summer, and spring show relatively weaker effects. Overall, these results indicate that attributes with clear visual cues are generally easier to edit effectively, whereas subtle facial expressions and attributes involving structural changes remain more challenging.

Table 4: Per-attribute performance analysis across three tasks.

ExpressionsΔCLIP ↑DINO ↑LPIPS ↓ObjectΔCLIP ↑DINO ↑LPIPS ↓SceneΔCLIP ↑DINO ↑LPIPS ↓
surprised0.0680.9030.222cracked0.0470.6960.233snowy0.1180.6610.357
angry0.0660.9340.243rusty0.0420.8680.205winter0.1090.6770.345
disgusted0.0510.9390.238mossy0.0390.7850.273foggy0.0870.7130.345
smiling0.0430.9610.254wet0.0360.8710.211rainy0.0730.7690.294
fearful0.0350.9520.244dusty0.0300.8740.187nighttime0.0710.7310.332
sad0.0270.9570.228rough0.0300.8520.197autumn0.0690.7960.311
spring0.0330.8370.209
summer0.0290.8460.191
windy0.0270.8540.187
Mean0.0480.9410.238Mean0.0370.8240.218Mean0.0680.7650.286

2.5 About Failure Cases

We have added qualitative analyses of representative failure cases to the Appendix. Specifically, when editing the "sad" attribute of a stage musician, the intended sad expression remains barely perceptible even as the control strength α increases from 0 to 1. This may be because the face occupies a relatively small region in the waist-up portrait, making fine-grained facial localization challenging. Another representative failure case involves "snowy" editing of a coastal village. Although increasing α produces more snow accumulation, it also noticeably alters the outlines and roof styles of the original buildings. This may be attributed to the generative model's learned semantic priors, which associate snowy environments with particular architectural appearances. These examples highlight two remaining limitations of our method: precise localization of small target regions and disentanglement of desired editing attributes from strongly correlated semantic priors.

2.6 Comparison with ATA

We became aware of ATA in August and subsequently included it in our discussion. However, as of the time of this rebuttal, the GitHub link provided in the ATA paper remains inaccessible (returning a 404 error), and the paper lacks sufficient details about the specific model architecture, making exact reproduction difficult. Nevertheless, we made our best effort to reproduce ATA based on the methodological descriptions provided in the paper and have included it in the quantitative comparison in the revised Table 1. In addition, we have added precomputed results from our ATA reproduction to our Hugging Face Space, allowing reviewers to directly compare ATA with our method under matched settings.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.