acceptodds
Under review as a conference paper at ICLR 2027

Kingfisher: Fast Personalized Visual Autoregressive Generation via Scale-Aware Prompt-Conditioned Scale Decomposition

Abstract

Personalized image generation has been extensively studied in diffusion models, often relying on subject-specific optimization of model weights, textual embeddings, LoRA modules, or adapters. In this work, we explore a different direction for fast personalization in visual autoregressive models. Our key observation is that a short prefix of early autoregressive scale codes is sufficient to strongly determine both object structure and global appearance, while later scales primarily refine the established generation trajectory. Based on this insight, we introduce Kingfisher, a model-tuning-free framework that decomposes early residual scale codes into content- and style-related components through prompt-conditioned scale-code subspaces. Given two real images, Kingfisher first decomposes their early scale codes, then recomposes the content components of one image with the style components of the other into a compact semantic prefix. The frozen generator subsequently predicts the remaining scales through bitwise self-correction for coherent refinement. Experiments on CSD-100 show that Kingfisher achieves strong content preservation and the highest style and text alignment among the evaluated methods, while reducing pair-specific preparation time by approximately 54x compared with diffusion-based methods and 7.5x compared with VAR-based alternatives. These results demonstrate that effective personalization can be achieved by directly manipulating the native scale-space representation of a frozen visual autoregressive generator.

Primary area:
generative models
Keywords:
Personalized image generation, Tuning-free framework, Visual autoregressive model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZypjOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Kingfisher performs image-guided content–style composition using a frozen Infinity-2B generator. Prefix reconstruction and prompt-routing experiments motivate intervention in the early residual scales, which jointly influence structure and appearance. Prompt-conditioned predictions define neutral-appearance and object-content subspaces through SVD; per-image optimization then estimates content and appearance components. The method combines these components into a four-scale prefix and lets Infinity generate the remaining scales under a target text prompt. On its CSD-100 evaluation, Kingfisher achieves strong style and text alignment and low reported preparation time, with a measurable trade-off between stylization and content similarity.

Critical strengths
  1. The controlled prefix-reconstruction and prompt-routing analyses provide evidence that a compact early prefix is an effective intervention point for both structure and appearance. Kingfisher turns this observation into a distinct cross-image composition method that operates on the native residual-code trajectory of a frozen generator, rather than relying on model-weight adaptation.

  2. The evaluation provides substantial evidence within its chosen setting: 500 ordered pairs constructed from CSD-100, multiple baselines, prefix and objective ablations, and a human preference study. Kingfisher leads four of five automatic metrics in Table 1 and receives 42.3% of the aggregate human preference votes. The reported preparation time is also considerably lower than that of the evaluated personalization baselines.

Critical weaknesses
  1. Comparative attribution and generality remain limited. All Kingfisher experiments use Infinity-2B, whereas the CSD-VAR baseline uses Switti despite the original CSD-VAR paper reporting an Infinity-based variant. A comparison under the same backbone and cross-image protocol would better establish the method-specific gain.

  2. The evidence does not isolate the marginal benefit or complete cost of per-image optimization. Table 7 ablates individual losses but does not compare the full method with its zero-step SVD initialization. Table 2 does not clearly specify whether its preparation time includes encoding, prompt-conditioned predictions, and SVD. These omissions limit how precisely the claimed quality–efficiency trade-off can be assessed.

Questions for authors
  1. Backbone-controlled comparison. Since the original CSD-VAR paper reports an Infinity-based variant, could the authors evaluate it on Kingfisher’s 500 ordered pairs under matched prompts, resolution, metrics, and timing conditions? If this is infeasible, please clarify the constraint.

  2. Optimization benefit and timing scope. Could the authors compare the full 250-step method against a zero-step variant using the SVD-initialized CiC_i and SiS_i, reporting both quality and end-to-end preparation time? Please specify whether the reported preparation time includes image encoding, prompt-conditioned scale predictions, and SVD.

  3. Minor presentation points: Equation (11) defines R3∗=A3R_3^*=A_3, but Table 5 treats the two as distinct; the operation defining R3∗R_3^* in that ablation should be stated. Also, Table 2 gives a CSD-VAR-to-Kingfisher preparation-time ratio of approximately 7.3×7.3\times, rather than the “more than 10×” stated in the main text.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Zypj (Part 1/1)

We thank the reviewer for the constructive comments, for recognizing the prefix analysis and the substantial evaluation, and for the careful reading of our notation and timing.

W1/Q1: Backbone-controlled comparison.

This comparison is feasible, and we have carried it out under fully matched conditions: the same frozen Infinity-2B backbone, the same 500 ordered pairs, target prompts, metrics, resolution, and seeds, and the same single-A100 timing setup. Under these conditions, Kingfisher improves over CSD-VAR on all five metrics: CSD-C 0.5516 vs. 0.4816, CLIP-I 0.7888 vs. 0.7178, CSD-S 0.5265 vs. 0.3232, DINO 0.4366 vs. 0.2594, and CLIP-T 0.3113 vs. 0.2877. To further confirm the method-specific gain, we also applied Kingfisher unchanged to Switti, CSD-VAR's original backbone, where it again outperforms CSD-VAR on all five metrics. Both comparisons are included in revised Table 1. Appendix I specifies that the Switti run keeps the same prompt sets, subspace ranks, decomposition objective, 250 optimization steps, prefix rule, and β=0.5\beta = 0.5, with no Switti-specific tuning.

W2/Q2: Zero-step baseline and optimization benefit.

We added a zero-step variant that directly uses the projection-based initialization Ci(0),Si(0)C_i^{(0)}, S_i^{(0)}. The 250-step optimization yields substantial gains: CSD-C improves from 0.4929 to 0.5504, CLIP-I from 0.6897 to 0.7922, CSD-S from 0.4679 to 0.5210, and CLIP-T from 0.2396 to 0.3127. The zero-step variant shows a slightly higher DINO (0.4782 vs. 0.4367), but together with its drop in content and text alignment, this indicates that the unrefined components still carry object information from the style image. The optimization is what separates the two factors. This gain comes at a modest cost. The zero-step variant shares image encoding, the prompt-conditioned scale predictions, and SVD with the full method, so the 250 optimization steps account for only part of the 11.77 s per-reference extraction time. Even with the optimization included, Kingfisher remains 10.5×10.5\times faster than CSD-VAR on the same backbone, so the quality gain does not compromise the efficiency advantage. The comparison is reported in Appendix G, Table 8, and illustrated in Fig. 5.

Regarding timing scope, the reported extraction time is end-to-end. It includes image encoding, the prompt-conditioned scale predictions, SVD, and the 250 optimization steps; we have made this explicit in the revised Table 2, which now separates reusable per-reference extraction, pair preparation, and inference. Even with all these steps included, Kingfisher requires 11.77 s/reference and 23.54 s/pair on Infinity, compared with 123.98 s/reference and 247.96 s/pair for CSD-VAR on the same backbone. Since the per-reference representation is reusable, this cost is paid once per image, however many pairs that image appears in.

Q3: Presentation corrections.

We thank the reviewer for spotting these. In Table 5, the last row is an ablation in which the third scale is also appearance-rectified, R3∗=A3−S3A+S3BR_3^* = A_3 - S_3^A + S_3^B, whereas the default prefix in Eq. (11) keeps R3∗=A3R_3^* = A_3. The Table 5 caption now states this operation explicitly, consistent with Sec. 4.3. The efficiency statement has also been corrected to the matched-backbone ratios in Table 2 and the associated text.

Dear Reviewer Zypj,

As the discussion period closes on November 18 (AoE), we would like to kindly check whether our response has addressed your main concerns: the fully matched CSD-VAR comparison on Infinity-2B and the additional Switti evaluation (Table 1, Appendix I); the zero-step SVD-initialization baseline isolating the benefit of the 250-step optimization (Fig. 5 and Appendix G, Table 8); the clarified end-to-end timing, which includes image encoding, prompt-conditioned scale predictions, SVD, and optimization, together with reusable per-reference extraction costs (Table 2); and the corrected third-scale notation in Table 5.

If these revisions address your concerns, we would be grateful if you would consider updating your score. We are happy to clarify anything further.

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

This paper presents Kingfisher, a fast personalization framework for visual autoregressive image generation. The method is motivated by the observation that the early residual scales of a visual autoregressive model strongly determine object structure and global appearance, while later scales mainly refine details. Given a content image and a style image, Kingfisher uses predictions from a frozen Infinity-2B model to construct prompt-conditioned content and neutral-appearance subspaces. It decomposes selected early scale codes into content- and appearance-related components, recomposes these components into a short semantic prefix, and lets the frozen autoregressive generator predict the remaining scales. The method therefore requires no model fine-tuning, LoRA modules, adapters, or learned subject embeddings. Experiments on 500 content–style pairs constructed from CSD-100 show competitive content preservation, strong style and text alignment, and substantially lower pair-specific preparation time than the evaluated diffusion and VAR-based baselines. The paper also includes ablations of its prefix composition, decomposition objective, and structural correction coefficient.

Critical strengths
  1. The paper introduces a useful way to personalize visual autoregressive generators through their native scale-code representation. The prefix-reconstruction and prompt-routing analyses provide evidence that early consecutive scales form a compact semantic bottleneck controlling both structure and appearance. This observation leads to a technically interesting alternative to modifying model parameters, adapters, or learned embeddings.

  2. Kingfisher provides a strong quality–efficiency trade-off. On the reported CSD-100 evaluation, it achieves the best CLIP-I, CSD-S, DINO, and CLIP-T scores among the evaluated methods, as well as the highest human preference rate. Its pair-specific preparation time is also substantially lower than those of the compared diffusion and VAR-based methods. The ablations generally support the proposed decomposition objective and scale-aware prefix design.

Critical weaknesses
  1. The evaluation coverage and statistical validation remain limited. The main results are obtained from 500 pairs constructed from CSD-100, with only one output generated per pair using a fixed random seed. No confidence intervals or significance tests are reported. It is therefore unclear whether the improvements remain stable across random seeds, prompt variations, and broader data distributions.

  2. The paper lacks a controlled comparison using the same backbone. Kingfisher, the diffusion baselines, and CSD-VAR use different generative backbones. Differences in generation quality and efficiency may therefore be partially attributable to the underlying models rather than solely to the proposed personalization method. Simple Infinity-2B baselines, such as direct early-scale swapping or interpolation, would help isolate the contribution of the proposed decomposition objective.

Questions for authors
  1. Could the authors provide controlled baselines using the same frozen Infinity-2B backbone, such as direct early-scale swapping or interpolation, or projection-only decomposition without joint optimization? These comparisons would help determine whether the improvements come from early-scale manipulation in general or from the proposed decomposition and optimization objective.

  2. Please clarify the efficiency claim in Section 5.2. According to Table 2, the preparation-time speedup over CSD-VAR is 171.65/23.54≈7.3×171.65/23.54 \approx 7.3\times, which is consistent with the approximately 7.5×7.5\times claim in the abstract. However, Section 5.2 describes this as “more than a 10×10\times reduction.”

  3. Could the authors report the main results over multiple random seeds, together with confidence intervals or paired significance tests? If a full additional evaluation is not feasible, please discuss the observed sensitivity to random seeds and prompt variations.

2 comments
AuthorsOpenReview ↗

Response to Reviewer QLJn (Part 1/1)

We are grateful to the reviewer for the helpful suggestions and for recognizing the native scale-code personalization and the strong quality–efficiency trade-off of Kingfisher.

W1/Q3: Multi-seed evaluation and stability.

We reran all methods with two fixed seeds (2026 and 2027) over all 500 ordered pairs and now report mean ±\pm standard deviation in revised Table 1. Kingfisher on Infinity-2B obtains 0.5516±0.00040.5516 \pm 0.0004 CSD-C, 0.7888±0.00170.7888 \pm 0.0017 CLIP-I, 0.5265±0.00610.5265 \pm 0.0061 CSD-S, 0.4366±0.00220.4366 \pm 0.0022 DINO, and 0.3113±0.00110.3113 \pm 0.0011 CLIP-T. These results highlight an additional strength of Kingfisher: it is more stable than optimization-based personalization methods (e.g., UnZipLoRA CSD-S 0.4849±0.16630.4849 \pm 0.1663), because it operates on a well-constrained scale-code representation and learns no new parameters. Moreover, on CLIP-I and CLIP-T, Kingfisher's margins over the next-best method (0.017 and 0.023) are an order of magnitude larger than the observed seed variation, indicating that the improvements are not an artifact of seed choice. Regarding prompt variation, the new-subject experiment (Appendix H.1), which uses different subject prompts, shows consistent gains in appearance transfer. With two seeds, formal significance tests have limited power, so we report mean and standard deviation; evaluating broader data distributions beyond CSD-100 is an important direction for future work. We also state explicitly in Sec. 5.1 that no best-of-seed selection is used.

W2/Q1: Controlled Infinity-2B baselines.

On the same frozen Infinity-2B backbone, we added direct early-scale swapping, interpolation, direct prefix mixing, random rank-matched subspaces, and projection-only (zero-step) decomposition without joint optimization. The results consistently favor Kingfisher. Swapping preserves content but transfers little style (CSD-S 0.3617 vs. 0.5210). Direct mixing and interpolation blend both references into ambiguous shapes and mixed appearances. Random subspaces and zero-step initialization lose content and text alignment (CLIP-T 0.2354 and 0.2396 vs. 0.3127). This indicates that the improvements stem from the proposed prompt-conditioned decomposition and its optimization objective rather than from generic early-scale manipulation. The controls are shown in Fig. 5 and Appendix G, Table 8.

To further rule out backbone effects, we evaluated CSD-VAR on the same Infinity-2B backbone, where Kingfisher improves CSD-C/CLIP-I/CSD-S/DINO/CLIP-T from 0.4816/0.7178/0.3232/0.2594/0.2877 to 0.5516/0.7888/0.5265/0.4366/0.3113 (revised Table 1).

Q2: Efficiency claim.

We thank the reviewer for the careful check. The "more than 10×10\times" statement mistakenly compared CSD-VAR on Switti with Kingfisher on Infinity. With the corrected matched-backbone comparison, the speedup is larger than the cross-backbone ratio of 7.3×7.3\times. On Infinity-2B, Kingfisher requires 11.77 s/reference versus 123.98 s/reference for CSD-VAR (10.5×10.5\times), and on Switti 6.95 s versus 85.83 s (12.3×12.3\times). Revised Table 2 now separates reusable per-reference extraction from pair-level preparation, and the Abstract and Sec. 5.2 consistently use these matched-backbone figures.

Dear Reviewer QLJn,

As the discussion period closes on November 18 (AoE), we would like to kindly check whether our response has addressed your main concerns: the two-seed evaluation with mean and standard deviation and no best-of-seed selection (Table 1, Sec. 5.1); the same-backbone CSD-VAR comparison on Infinity-2B (Table 1); the controlled Infinity-2B baselines, including swapping, interpolation, direct mixing, random subspaces, and zero-step initialization (Fig. 5 and Appendix G, Table 8); and the corrected matched-backbone efficiency accounting (Table 2, Sec. 5.2).

If these revisions address your concerns, we would be grateful if you would consider updating your score. We are happy to clarify anything further.

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

This paper proposes Kingfisher, a framework for combining the content of one image with the appearance of another using a frozen visual autoregressive model. Motivated by the influence of early scales, it constructs prompt-conditioned subspaces to estimate content and style components, recomposes them into a four-scale prefix, and autoregressively generates the remaining scales. On CSD-100, it reports strong style/text alignment, competitive content preservation, and reduced preparation time compared with the evaluated baselines.

Critical strengths

S1. A concrete methodological contribution. Constructing prompt-conditioned subspaces directly in visual scale-code space is an interesting alternative to learning textual representations or auxiliary memories. The contribution is specifically the subspace-based decomposition and prefix composition, rather than scale-aware personalization itself.

S2. Promising results and clear organization. Kingfisher reports the strongest style alignment and human-preference share among the evaluated methods. The paper is generally well organized, and the loss and prefix-composition ablations help explain its design choices.

Critical weaknesses

W1. Novelty and problem scope need more precise positioning. As discussed in the cited prior work [2–5], early-scale control and test-time manipulation of VAR representations are not new directions. The novelty therefore lies primarily in prompt-conditioned subspace decomposition. Moreover, the evaluation demonstrates cross-image content–style composition under a fixed prompt template, rather than broader personalization through independently reusable concepts and recontextualization. This distinction matters when selecting baselines and interpreting the claims.

W2. The comparisons do not isolate the methodological advantage. Kingfisher uses Infinity-2B, whereas its CSD-VAR baseline uses Switti, although the original CSD-VAR supports both backbones. A same-Infinity comparison is needed to separate method improvements from backbone differences. Training-free style-transfer methods, particularly [2,3], are also important missing comparisons given the task and efficiency claims.

The distinction should not be framed simply as frozen versus fine-tuned generators: CSD-VAR also freezes pretrained generator weights. Kingfisher instead optimizes visual-code components, including 250 test-time optimization steps, rather than textual embeddings and K–V memories.

W3. The subspace interpretation is insufficiently validated. The scale analysis motivates editing early scales, but the specific recipe of modifying scales 1–2, preserving scale 3, and correcting scale 4 remains empirically selected. Although the ablations support this configuration, they do not establish whether prompt-conditioned subspaces are necessary compared with random rank-matched subspaces, direct prefix mixing, or projection-only initialization. Furthermore, “neutral” photographic prompts define a particular appearance reference, not a style-free space. Stronger controls are needed to substantiate the claimed content–style separation.

W4. Efficiency accounting needs clarification. CSD-VAR decomposes each reference independently, so preparation can potentially be reused across pairs. Reporting only per-pair cost may obscure this distinction. The paper should separate per-reference extraction, pair composition, and generation costs. Additionally, the stated “more than 10×” preparation reduction is inconsistent with Table 2’s 171.65 versus 23.54 seconds.

[1] Voronov, Anton, et al. "Switti: Designing scale-wise transformers for text-to-image synthesis." arXiv preprint arXiv:2412.01819 (2024) [2] Lee, Kyoungmin, et al. “A Training-Free Style-Personalization via SVD-Based Feature Decomposition.” CVPR 2026 [3] “A Training-Free Style-aligned Image Generation with Scale-wise Autoregressive Model.” arXiv:2504.06144. [4] “EditInfinity: Image Editing with Binary-Quantized Generative Models.” arXiv:2510.20217. [5] “Edit the Bits, Diff the Codes: Bitwise Residual Editing for Visual Autoregressive Models.” arXiv:2606.13558.

Questions for authors

Q1. Can the authors compare against CSD-VAR on the same Infinity backbone, and against training-free VAR style-transfer methods [2,3]? Q2. Do prompt-conditioned subspaces outperform random rank-matched subspaces, direct prefix mixing, and initialization without optimization? Q3. Can extracted components support recontextualization and style transfer to new text-specified subjects beyond the fixed composition template? Q4. How are reusable per-reference preparation and pair-specific costs accounted for? Does the approach also work on Switti?

5 comments
AuthorsOpenReview ↗

Response to Reviewer PdSA (Part 1/2)

We thank the reviewer for the detailed and constructive feedback, and for recognizing prompt-conditioned subspace decomposition in scale-code space as a concrete methodological contribution. We address each point below.

W1/Q3: Novelty and problem scope.

We agree that the importance of early scales in VAR models has been observed before (SVD-PFB, Lee et al., 2026; Style-aligned, Park et al., 2025; EditInfinity, Wang et al., 2026; Edit the Bits, Zhang et al., 2026). However, these works use early or intermediate scales as control knobs: they swap, inject, align, or edit scale representations or bits to steer generation. To the best of our knowledge, none of them decomposes the scale codes of a real image into explicit content and appearance components. Kingfisher does this by constructing prompt-conditioned subspaces from the frozen generator's own predictions, so that the factors are defined by the generator itself, without learned tokens, adapters, or hand-picked scales. This is what allows content from one real image and appearance from another to be recombined within a single compact prefix. Our scale-wise analysis serves a complementary role: beyond confirming the importance of early scales in Infinity, it identifies that a four-scale prefix is sufficient to determine the generation trajectory, which defines where the decomposition should operate. Following the reviewer's suggestion, we revised the Abstract, Introduction (Sec. 1), contribution statements, and Related Work (Sec. 2) to explicitly acknowledge prior works on early/intermediate-scale control and to position our main methodological contribution as prompt-conditioned scale-code decomposition.

Regarding scope, we agree that the main benchmark evaluates cross-image content–style composition under a fixed template. We note that this follows the established CSD-100 protocol of CSD-VAR and is a challenging setting: each concept is given by a single real image, and content and style must be disentangled from different references at the same time. To test generality beyond this template, we added two experiments. For style transfer to new text-specified subjects (Appendix H.1, Table 9, Fig. 10; 150 samples), Kingfisher achieves the strongest appearance transfer (CSD-S 0.4545, DINO 0.3256) compared with SVD-PFB (0.3325/0.2405) and Style-aligned (0.4070/0.2744), while maintaining comparable subject alignment, even though both baselines were designed specifically for this text-driven setting. In a subject-recontextualization experiment (Appendix H.2, Fig. 11), content components extracted from five images of the same subject are placed in new text-specified contexts without any training. Together, these results suggest that the decomposed components are reusable beyond the paired setting.

W2/Q1: Same-backbone and training-free comparisons.

We ran CSD-VAR on the same frozen Infinity-2B backbone and, conversely, applied Kingfisher unchanged to Switti. Averaged over two seeds on Infinity-2B, Kingfisher improves over CSD-VAR on all five metrics: CSD-C 0.4816→0.55160.4816 \rightarrow 0.5516, CLIP-I 0.7178→0.78880.7178 \rightarrow 0.7888, CSD-S 0.3232→0.52650.3232 \rightarrow 0.5265, DINO 0.2594→0.43660.2594 \rightarrow 0.4366, and CLIP-T 0.2877→0.31130.2877 \rightarrow 0.3113. On Switti, CSD-VAR's own backbone, Kingfisher again outperforms CSD-VAR on all five metrics, using exactly the same settings without any backbone-specific tuning. The advantage holds in both directions, which indicates that it stems from the proposed decomposition rather than from the backbone.

We also re-implemented the training-free VAR methods SVD-PFB (Lee et al., 2026) and Style-aligned (Park et al., 2025) on Infinity-2B and adapted them to the two-real-image setting. Kingfisher outperforms SVD-PFB on all five metrics and outperforms Style-aligned on four of five (CLIP-I, CSD-S, DINO, CLIP-T). Style-aligned obtains a higher CSD-C (0.5709), which reflects its more conservative content preservation at the expense of style and text alignment. Kingfisher thus achieves the most balanced performance among the evaluated VAR methods.

We agree that CSD-VAR also keeps the generator weights frozen, and that Kingfisher involves test-time optimization, albeit of scale-code components rather than model parameters. This clarifies our contribution: with the same frozen backbone, the main difference is what is optimized and where. CSD-VAR optimizes textual embeddings and K–V memories, which requires repeated passes through the generator. Kingfisher operates directly in the generator's native visual scale-code space, with subspaces precomputed once from prompt-conditioned predictions, so its 250 optimization steps act on compact code tensors. This representational choice largely accounts for the 10.5×10.5\times lower extraction cost under the same backbone. We now state this distinction explicitly in Related Work (Sec. 2), the efficiency analysis (Sec. 5.2), and Appendix I. All new results are reported in revised Table 1 and Fig. 4.

AuthorsOpenReview ↗

Response to Reviewer PdSA (Part 2/2)

W3/Q2: Validation of the subspace interpretation.

Are prompt-conditioned subspaces necessary? On the same frozen Infinity-2B backbone, 500 pairs, and seed 2026, we compared against zero-step (projection-only) initialization, random rank-matched subspaces, direct prefix mixing, early-scale swapping, and interpolation. Kingfisher obtains CSD-S/CLIP-T of 0.5210/0.3127, compared with 0.4679/0.2396 (zero-step), 0.4969/0.2354 (random subspaces), 0.4408/0.2497 (direct mixing), 0.3617/0.2617 (swapping), and 0.4584/0.2471 (interpolation). Swapping essentially copies the content image and transfers little style. Mixing and interpolation blend the two references and yield ambiguous shapes. Random subspaces and zero-step initialization reach higher DINO similarity to the style image, but their sharp drop in content and text alignment shows that they transfer the style image's object information together with its appearance. Only the full method balances all three criteria, indicating that the gains stem from the prompt-conditioned decomposition rather than from early-scale intervention alone.

Prefix configuration. The configuration is guided by our scale-wise analysis and validated systematically in Table 5, which reveals a consistent trend: rectifying more early scales steadily increases style alignment and decreases content preservation, and rectifying the third scale in particular sharply damages object identity. Preserving scale 3 as a semantic anchor and applying a mild content correction at scale 4 is therefore well motivated. In addition, the same configuration transfers to Switti without any retuning and still outperforms CSD-VAR there, indicating that it is not specific to Infinity.

"Neutral" prompts. We agree with the reviewer that neutral photographic prompts define an appearance reference rather than a strictly style-free space, and we have clarified this wording. This choice is intentional: measuring SiS_i relative to a natural-photograph reference gives a practical definition of appearance for transfer, and it requires no style labels, style datasets, or learned style encoders. We now describe this as an operational photographic reference (Sec. 4.1) and list the exact prompts in Appendix B, Table 4.

The controls are shown qualitatively in Fig. 5 (Sec. 5.3) and quantitatively in Appendix G, Table 8.

W4/Q4: Efficiency accounting and reuse.

Since both CSD-VAR and Kingfisher process each reference independently, we now report reusable per-reference extraction, pair-level preparation, and per-image inference separately in revised Table 2. Pair composition itself only recombines the precomputed components via Eq. (11) and adds negligible cost, so pair-level preparation is dominated by extracting the two references. We also thank the reviewer for spotting the inconsistent "more than 10×10\times" statement, which compared methods across different backbones. With matched backbones, the speedup is larger than the ratio the reviewer derived: Kingfisher requires 11.77 s versus 123.98 s per reference on Infinity-2B (10.5×10.5\times), and 6.95 s versus 85.83 s on Switti (12.3×12.3\times). Compared with diffusion-based methods, extraction is at least 55×55\times faster, and inference costs only 1.61 s/image on Infinity and 0.72 s/image on Switti. Kingfisher therefore works well on Switti in terms of both quality (Table 1) and efficiency, and the earlier inconsistent statement has been removed.

Dear Reviewer PdSA,

As the discussion period closes on November 18 (AoE), we would like to kindly check whether our response has addressed your main concerns: the novelty and scope, now centered on prompt-conditioned scale-code decomposition and extended beyond the two-image setting (Appendix H); the same-backbone CSD-VAR comparison on Infinity-2B and cross-backbone validation on Switti (Table 1); the additional training-free VAR baselines (Table 1, Fig. 4); the controlled validation of the decomposition, including zero-step, random subspaces, direct mixing, swapping, and interpolation, together with the clarified prefix design and neutral-appearance reference (Fig. 5, Appendix G/Table 8, Secs. 4.1 and 4.3); and the revised distinction from CSD-VAR and reusable efficiency accounting (Sec. 5.2, Table 2, Appendix I).

If these revisions address your concerns, we would be grateful if you would consider updating your score. We are happy to clarify anything further.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revision Cover Letter (Part 1/2)

We thank all reviewers for their constructive feedback. We are encouraged that the reviewers found the use of prompt-conditioned subspaces in the native scale-code space a concrete and interesting alternative to weight, embedding, or adapter adaptation (PdSA, QLJn, Zypj), and recognized the strong quality–efficiency trade-off, the best style/text alignment, and the highest human preference among the evaluated methods (PdSA, QLJn, Zypj).

The reviewers' main concerns and our responses are summarized below.

  • Novelty and scope (PdSA). We agree that the importance of early scales in VAR models has been observed before. Prior works use these scales as control knobs, whereas Kingfisher decomposes the scale codes of real images into explicit content and appearance components via prompt-conditioned subspaces. We now credit prior work explicitly and position our main contribution as prompt-conditioned scale-code decomposition. New experiments show that the extracted components also support style transfer to new text-specified subjects and, in a preliminary setting, subject recontextualization.

  • Backbone-controlled comparisons (PdSA, QLJn, Zypj). On the same frozen Infinity-2B backbone, Kingfisher outperforms CSD-VAR on all five metrics (e.g., CSD-S 0.3232→0.52650.3232 \rightarrow 0.5265, DINO 0.2594→0.43660.2594 \rightarrow 0.4366), and it does so again on Switti without retuning. The gain therefore comes from the method, not from the backbone. Kingfisher also achieves the most balanced performance against SVD-PFB and Style-aligned.

  • Distinction from CSD-VAR (PdSA). Both methods keep the generator frozen; the difference lies in what is optimized. CSD-VAR optimizes textual embeddings and K–V memories, whereas Kingfisher optimizes compact scale-code components with respect to subspaces precomputed once per reference, which largely accounts for its lower cost.

  • Necessity of the decomposition and design choices (PdSA, QLJn, Zypj). Same-backbone controls (zero-step initialization, random rank-matched subspaces, direct prefix mixing, early-scale swapping, and interpolation) all fail to match Kingfisher's content–style balance, showing that both the prompt-conditioned subspaces and the 250-step optimization contribute. The prefix configuration follows a consistent trend in the prefix ablation and transfers to Switti unchanged, and the neutral prompts are clarified as an operational photographic reference.

  • Statistical stability (QLJn). With two independent seeds, Kingfisher shows small standard deviations (e.g., CSD-S 0.5265±0.00610.5265 \pm 0.0061) and is more stable than optimization-based baselines. Its margins on CLIP-I and CLIP-T exceed the seed variation by an order of magnitude. No best-of-seed selection is used.

  • Efficiency accounting (PdSA, QLJn, Zypj). We removed the inconsistent "more than 10×10\times" claim and now report reusable per-reference extraction (end-to-end), pair preparation, and inference separately. On matched backbones, Kingfisher reduces extraction cost by 10.5×10.5\times (Infinity-2B) and 12.3×12.3\times (Switti) compared with CSD-VAR, and by at least 55×55\times compared with diffusion-based methods.

  • Presentation (Zypj). We clarified the third-scale notation in the prefix ablation and corrected the associated efficiency text.

AuthorsOpenReview ↗

Revision Cover Letter (Part 2/2)

Accordingly, the paper revisions are as follows:

  • Novelty positioning. The revised Abstract, Introduction, and contribution statements acknowledge prior works on early-scale control and make the distinct motivation explicit: prior VAR methods use early scales as control signals, whereas our scale-wise analysis (Sec. 3) shows that content and appearance are entangled within these codes, which motivates separating them in the native scale-code space of each real image. The main contribution is accordingly positioned as prompt-conditioned scale-code decomposition. A new Related Work paragraph (Sec. 2) contrasts Kingfisher with training-free VAR control methods, which steer generation through early scales, and with CSD-VAR, which represents content and style as textual embeddings and K–V memories rather than as components of the visual scale codes.

  • Backbone-controlled comparisons. We added CSD-VAR on the same frozen Infinity-2B backbone and evaluated Kingfisher on Switti (revised Table 1, Fig. 4), with implementation details and the CSD-VAR distinction in Appendix I.

  • Additional VAR baselines and controls. We added SVD-PFB and Style-aligned to revised Table 1 and Fig. 4. Controlled Infinity-2B variants are discussed in Sec. 5.3 and Fig. 5, with quantitative results in Appendix G, Table 8.

  • Method clarifications. Sec. 4.1 clarifies that the neutral prompts provide an operational photographic reference rather than a style-free space, and Sec. 4.3 grounds the prefix configuration in the scale-wise analysis (Sec. 3) and the prefix ablation (Table 5).

  • Multi-seed evaluation. We reran the main benchmark with two seeds (2026 and 2027) and report mean and standard deviation for all methods in revised Table 1; Sec. 5.1 states that no best-of-seed selection is used.

  • Efficiency accounting. Revised Table 2 separates reusable per-reference extraction, pair-level preparation, and per-image inference, and its caption specifies that extraction time is end-to-end. Sec. 5.2 explains the efficiency gap over CSD-VAR, and the Abstract and contributions now use matched-backbone speedups only.

  • Beyond two-image composition. Appendix H adds style transfer to new text-specified subjects (Appendix H.1, Table 9, Fig. 10) and a preliminary subject-recontextualization extension (Appendix H.2, Fig. 11).

  • Presentation corrections. The Table 5 caption now states the third-scale operation explicitly, consistent with Eq. (11) and Sec. 4.3.

We thank the reviewers again for the suggestions, which substantially improved the positioning, controls, and evaluation of the paper. Detailed responses to each reviewer are provided below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.