acceptodds
Under review as a conference paper at ICLR 2027

Bind-and-Route: Explicit Reference Binding and Attention Routing for Multi-Subject Personalized Generation

Abstract

Multi-subject personalized generation aims to synthesize multiple customized subjects in a single image while preserving both textual controllability and reference appearance fidelity. However, we observe that existing personalization methods degrade substantially as the number of reference subjects increases, leading to reduced text alignment and weakened appearance consistency. By analyzing multi-modal attention patterns, we find that such degradation is closely related to Text-Reference Misalignment and Subject Confusion: the text of a subject and its corresponding reference condition may attend to different generation regions, while the target generation regions of a subject may attend to mismatched reference conditions. To address these issues, we propose Bind-and-Route, a framework for robust multi-subject personalization. First, we establish explicit text-reference associations through placeholder-reference contrastive learning and bidirectional binding resamplers, which mutually inject the associated textual and reference information to facilitate binding in multi-modal attention. Second, we introduce a lightweight attention router, which uses image-text attention as a localization prior to route image tokens toward matched reference conditions. Experiments demonstrate that our method can achieve strong text alignment and reference appearance fidelity, offering an effective solution under unified multi-modal conditioning.

Primary area:
generative models
Keywords:
Personalized image generation, Image customization

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_MyB6OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Bind-and-Route targets multi-subject personalized image generation, where increasing reference subjects causes text-reference misalignment and subject confusion. It introduces bidirectional cross-attention resamplers and contrastive binding losses to associate each placeholder with its matched reference. It further proposes a spatial router that uses noise-to-text attention priors to redistribute image-to-reference attention without inference-time masks. Experiments on PSRBench, XVerseBench, qualitative comparisons, ablations, and a small user study show competitive overall performance.

Critical strengths
  1. The paper offers a mechanistic decomposition of multi-subject failures. Explicit Reference Binding addresses incorrect text-reference correspondence, while Attention Routing addresses incorrect spatial assignment, making the proposed interventions interpretable and modular.

  2. The method targets the four-reference setting, where existing approaches degrade substantially. It improves the PSRBench average and aesthetic scores, while remaining mask-free at inference and outperforming the same-data OminiControl baseline.

Critical weaknesses
  1. The Bind contrastive loss does not define the source of AP2RA^{P2R} and AR2PA^{R2P} precisely. The paper omits the layer, heads, timestep, and aggregation procedure, leaving a key training signal under-specified.

  2. The Bind resampler is not described sufficiently for reproduction. The paper does not clarify its timestep execution or caching, the training timestep schedule, or basic architecture details such as dimensions, head count, FFN configuration, depth, parameter count, and weight sharing.

  3. Route is evaluated only through a coarse cumulative ablation. The paper does not isolate the contributions of the text and background priors, offset correction, three loss terms, loss weights, mask thresholds, block and timestep ranges, or text-prior detachment. Therefore, the claimed gains cannot be attributed to specific design choices.

  4. The experimental comparison is not fully controlled or fair. The baselines use different backbones and inference step counts, while Bind-and-Route introduces LoRA, bidirectional resamplers, a spatial router, and placeholders with 285M trainable parameters. Without matched parameter budgets, training compute, and adaptation procedures, the reported gains cannot be cleanly attributed to the proposed binding and routing mechanisms.

Questions for authors

See Weakness.

6 comments
AuthorsOpenReview ↗

Response to reviewer MyB6 (part 1 of 2)

Please refer to our revised manuscript and the Summary of Changes in the Revised Manuscript comment. Below is our response to the weaknesses:

1. Attention source and aggregation for the binding contrastive loss. For the token-level attention aggregation AP2RA^{\text{P2R}} and AR2PA^{\text{R2P}} defined after Eq. (1) and Eq. (3), we average the attention probabilities over all heads within each block. The bidirectional contrastive loss is computed and averaged across all 19 double-stream and 38 single-stream blocks. There is no block-range or timestep-range restriction for this objective.

2. Binding resampler architecture and execution details. Appendix A specifies that the binding resamplers are applied before the first transformer block at each timestep, while Section 3.1 and Figure 3 describe their bidirectional residual feature injection. Both resamplers use input, internal, and output dimensions of 3072 and consist of one pre-LayerNorm cross-attention layer with eight heads, followed by a pre-LayerNorm FFN, both with residual connections. The FFN has dimensions 3072–12288–3072 and uses GELU activation; all dropout rates are zero. The detailed architecture is shown in Figure 8. The two directions are instantiated independently, while subjects share the mapper within each direction. Each mapper contains 113,298,432 parameters, giving 226,596,864 parameters for the two resamplers together. All trainable modules, including the semantic resampler, are optimized using FLUX. 1-dev’s default timestep sampling schedule and rectified-flow training objective. Resampler outputs are recomputed without caching during both training and inference. We have supplemented Appendix A with numerical configurations and execution details.

3. Ablation study. We have expanded router range selection experiments in Appendix A (Table 4) and ablation studies in Appendix D (Table 7).

Text and background priors. The prior-only variant (Prior only) directly uses the text priors to guide the reference image attention. Compared with the Direct attention mask loss variant, it achieves better aesthetic quality by using text attention to guide reference image attention. However, without the router to further distinguish the regions associated with different subjects, Subject Confusion remains, resulting in lower subject consistency than the full method. This distinction is also illustrated by the mean text prior and mean router alpha visualized in the last two columns of Fig. 9, where the router produces more spatially distinct localized reference-region assignments. The base prior serves as a background branch that absorbs routing probability in background regions. Removing it makes reference image attention more diffuse (Router w/o pbasetextp^{\text{text}}_{\text{base}}). Although this variant yields a marginal improvement in subject consistency in the 4-reference setting, it reduces semantic alignment and aesthetic quality.

Offset correction. During training, we observe that omitting the offset causes the total reference attention mass after routing to decrease substantially, as illustrated in Fig. 10. Router w/o offset slightly improves semantic alignment but reduces subject consistency, consistent with the decrease in reference attention mass.

Lrouter\mathcal{L_{\text{router}}} weights. The ablation of Lrouter\mathcal{L_{\text{router}}} coefficients shows that Lce\mathcal{L_{\text{ce}}} (ce only (0.5/0/0)) improves subject consistency, compared with Router w/o Lrouter\mathcal{L_{\text{router}}}, but a large λce=0.6\lambda_{\text{ce}}=0.6 (ce only (0.6/0/0)) can compromise semantic alignment and aesthetic quality. Adding the negative (ce + neg(0.5/0.1/0)) and margin losses (ce + margin (0.5/0/0.1)) also improves subject consistency.

Router range selection. Following the reviewers’ suggestions, we found that applying the router throughout the entire generation process yielded larger gains. Rather than testing every possible routing range, we used the analysis in Appendix A to select candidate ranges and adopted the best-performing configuration among those tested. An exhaustive search over all hyperparameter settings is infeasible; we therefore select a well-motivated, reasonable configuration rather than seek a globally optimal one.

4. Backbone fairness. Except for MS-Diffusion, which uses SDXL, all compared methods use DiT backbones. Qwen-Image-Edit and OmniGen2 do not use backbones inherently weaker than FLUX.1-dev. All FLUX.1-based methods are evaluated with the same number of denoising steps, while the other methods use their recommended default settings. Thus, the inference settings do not deliberately favor our method.

AuthorsOpenReview ↗

Response to reviewer MyB6 (part 2 of 2)

5. Parameter-count ablation. FLUX.1-dev transformer contains about 12B parameters, and our 285M trainable parameters account for 2.38% of its total parameter count. To provide a more capacity-controlled comparison, we additionally evaluate a baseline with rank-95 LoRA, comprising 344M trainable parameters:

Method3-ref Subject3-ref Semantic3-ref Aesthetic3-ref Average4-ref Subject4-ref Semantic4-ref Aesthetic4-ref Average
Baseline0.51900.86380.94690.77660.40260.85430.94620.7344
Baseline, 344M LoRA0.52360.86650.98780.79260.43940.73190.77640.6492
Ours (full)0.52750.89610.96390.79580.48920.86440.98790.7805

Increasing the baseline’s trainable parameter count improves its performance on the 3-reference subset compared with our current baseline. However, on the four-reference subset, semantic alignment and aesthetic quality deteriorate substantially, resulting in an Average score below that of the original baseline. The 344M LoRA baseline also underperforms our full method on every reported metric except Aesthetic in the 3-reference setting, despite having more trainable parameters. These results suggest that increasing model capacity alone is insufficient to reproduce the gains achieved by our method in the evaluated settings.

Thank you for the detailed responses and the additional results, which have partly addressed my concerns. However, considering the overall quality of the paper, and since I agree with some of the points raised by Reviewer 3Nna, I have decided to raise my final score to 2.

Reviewer_3NnaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies multi-subject personalized generation with diffusion transformers, where text, image and several reference images interact in one attention space. Based on attention maps, the authors argue that quality drops as references are added for two reasons. A subject's text and its reference attend to different regions, and the region of one subject attends to the wrong reference. The proposed Bind-and-Route is built on FLUX.1-dev and OminiControl and has two parts. Explicit Reference Binding adds placeholder embeddings and bidirectional resamplers that inject reference features into placeholders and text features into reference tokens. It also adds contrastive losses on the attention between placeholders and references. Attention Routing uses the attention from image tokens to each subject's text as a location prior. A small convolutional router, trained with foreground masks, turns this prior into logit biases over reference tokens, with an offset that keeps the total attention on references unchanged. Experiments on PSRBench with three and four references and on XVerseBench compare against eight recent methods, together with an ablation and a small user study.

Critical strengths
  1. The diagnosis is clear and each component targets a specific failure. The paper separates two failure modes in multi-reference attention and designs one mechanism for each. Using the model's own text attention as a location prior is a neat choice. It needs no masks at inference, and the offset that preserves the total attention on references is a thoughtful detail that protects text alignment.

  2. The benefit grows with the number of references. With four references on PSRBench, the method clearly improves Subject and Aesthetic scores over the OminiControl baseline trained on the same data, which rules out data as the main source of the gain. The cumulative ablation and the comparison with eight recent methods give a reasonably broad picture.

Critical weaknesses
  1. The motivating analysis is only qualitative. Text-Reference Misalignment and Subject Confusion are illustrated with a few attention maps in Figure 1. There is no quantitative measure across the benchmark, such as the overlap between a subject's text attention and its reference attention, or the share of a subject region's attention that goes to wrong references. There is also no evidence that Bind-and-Route actually reduces these quantities. The claim that existing methods degrade as references are added is not shown as a trend from one to four references in the main paper. Without this, the link between the diagnosed problems and the reported gains is asserted rather than shown.

  2. The gains are small and concentrated in the aesthetic score. With three references, Subject improves by about 0.005 over the baseline and stays below DreamO. Semantic is lower than the baseline and several other methods, and adding the router lowers Semantic in the ablation. The best Average is largely driven by the HPSv3-based Aesthetic score, which measures preference rather than personalization, and averaging three metrics on different scales amplifies this. On XVerseBench in the appendix, the method is behind XVerse and DreamO on average with one and two references, and well behind on identity similarity. The claim of improving both text alignment and appearance consistency should be qualified.

  3. The evaluation is too small to support the differences. PSRBench has 50 cases per setting, and results come from a single run without seeds or confidence intervals. Most ablation differences in the three-reference setting are in the third decimal place. The user study uses 10 questionnaires on five cases with coarse rank bins. Results over several seeds with confidence intervals, and a larger user study, are needed before these numbers can be trusted.

  4. The effect of mask supervision is not separated from the router. The router is trained with GroundingDINO foreground masks, which the baseline and most compared methods do not use. The ablation adds the router and the mask losses together, so it is unclear whether the gain comes from the router design or simply from extra spatial supervision. Two controls would settle this: the router without mask losses, and a plain mask loss on reference attention without the router. The router is also close in spirit to the routing-based personalization methods cited in the related work, and the differences should be discussed. If these controls show that the router itself drives the gain, I would be willing to raise my score.

  5. Practical details are missing. The router needs explicit head-averaged attention maps from many blocks, which usually rules out fused attention kernels. Memory and runtime overhead relative to the baseline should be reported. At inference, each reference must be linked to a text span, but the paper does not explain how these spans are set on the benchmarks or how robust the method is to them. The block range and timestep range of the router are chosen based on training performance, with no sensitivity analysis.

  6. Standard personalization metrics are missing. The main comparison relies on the PSRBench scores, whose Subject and Semantic metrics are not described in the paper, and on an HPSv3-based aesthetic score. The metrics most commonly used in subject-driven generation are not reported on the main benchmark: CLIP-T for text alignment, and CLIP-I and DINO for subject fidelity. DINO appears only in the appendix as the IP score on XVerseBench, where the method is not among the best. In the multi-subject setting, these scores should be computed per subject on cropped or segmented regions, so that a missing or fused subject is penalized. Without these metrics, it is hard to compare the method with the large body of prior work that reports them.

  7. Several closely related works are not cited or discussed. The paper positions itself against zero-shot multi-subject methods, but omits works that address the same binding and confusion problems more directly. MIP-Adapter identifies the same object confusion problem with multiple reference images and resolves it by weighting each reference's features at every latent position, which is close in spirit to Attention Routing. GroundingBooth binds each subject to its own region through a subject-grounded cross-attention layer and supports multiple subjects in a zero-shot setting. The recent MultiCompose targets per-subject attribute binding in multi-subject personalization and proposes a benchmark for attribute misalignment. MultiCompose appeared shortly before the deadline, so I do not expect a full comparison. Still, these works should be discussed so that readers can see how explicit binding and routing differ from these approaches.

some Minor weaknesses. In Table 4, the XVerse row for three references is identical to the Qwen-Image-Edit row for two references in all five numbers, which looks like a copy error. The text's claim that XVerse has the lowest AES with three references relies on this row. Section 3.1 and Figure 3 insert placeholders into the prompt text, while Section 4.4 and the appendix say they are added to the encoded embeddings. The limitations say Echo-4o was generated by Nano-banana, but the cited Echo-4o paper describes GPT-4o synthetic images. The introduction attributes textual inversion to Mokady et al. 2023, which is Null-text Inversion. Textual inversion is Gal et al.

References

  • MIP-Adapter: Qihan Huang, Siming Fu, et al. Resolving Multi-Condition Confusion for Finetuning-Free Personalized Image Generation. AAAI 2025. https://github.com/hqhQAQ/MIP-Adapter
  • GroundingBooth: Zhexiao Xiong, Wei Xiong, Jing Shi, He Zhang, Yizhi Song, Nathan Jacobs. GroundingBooth: Grounding Text-to-Image Customization. TMLR 2026. https://groundingbooth.github.io
  • MultiCompose: Ruirui Zhang, Zhengkai Zhao, Pan Gao. MultiCompose: Multi-Concept Personalized Composition with Per-Subject Attribute Binding. ACM MM 2026. https://arxiv.org/abs/2608.03708
Questions for authors

Requests that could change my assessment

  1. Can you quantify Text-Reference Misalignment and Subject Confusion on the benchmark, for the baseline and for your method, as the number of references grows? This would directly support the paper's main motivation.
  2. Can you report the router without mask losses and a plain mask loss without the router, and give PSRBench results over several seeds with confidence intervals? Please also add CLIP-T, CLIP-I and DINO scores computed per subject. This would show whether the gains are reliable and where they come from.
  3. How are subject text spans and placeholders specified at inference on the benchmarks? Are placeholders inserted into the prompt text or added to the encoded embeddings? What is the memory and runtime overhead compared with OminiControl?
8 comments
AuthorsOpenReview ↗

Response to reviewer 3Nna (part 1 of 2)

Please refer to our revised manuscript and Summary of Changes in the Revised Manuscript. Below is our response to the questions:

1. Quantifying the motivating analysis. We have added a quantitative analysis of text–reference attention alignment to Section 4.3. For every subject, we extract the attention that the generated image tokens pay to the subject's label tokens and to its reference image tokens, and measure their spatial agreement with Pearson correlation and soft IoU. The results are shown in Tab. 2 and Fig. 9. The baseline’s soft IoU decreases from 0.4372 to 0.4348, 0.3878, and 0.3793 across 1- to 4- reference settings, whereas Bind-and-Route achieves 0.4579, 0.4722, 0.4340, and 0.4349, respectively, substantially mitigating the decline observed for the baseline. Our method also achieves higher Pearson correlation in every setting. These results demonstrate that Bind-and-Route consistently improves text–reference alignment relative to the baseline and effectively alleviates the diagnosed misalignment.

The visualizations in Fig. 9 further show that our reference attention follows the corresponding text-grounded regions more closely, with less leakage into other subjects’ regions, providing qualitative evidence of reduced Subject Confusion. Within PSRBench, increasing the reference count from three to four leads to a smaller decrease in subject consistency for our method (0.5275 → 0.4892, -0.0383) than for the baseline (0.5190 → 0.4026, -0.1164). Together, these findings provide evidence consistent with improved attention alignment and reduced cross-subject interference contributing to the gains in multi-subject generation.

2. Router ablations and evaluation reliability. First, we would like to clarify that the router inputs are detached. The Lrouter\mathcal{L_{\text{router}}} directly supervises the router outputs, and their gradients update only the router without propagating back to the DiT, as noted in Appendix A (“The text prior is detached, so these losses only affect the router”). The router is designed to integrate multimodal contextual information and suppress noise that may arise from directly using text attention maps as localization priors during inference. It reduces this noise by allowing text tokens associated with different subjects to compete within the same generated region and encouraging more differentiated attention allocation. The mask supervision objectives impose more direct and stronger constraints on this goal by acting on the router outputs, while the router provides the mechanism for integrating multimodal contextual information. In the “Direct attention mask loss” ablation in Tab. 7, we apply mask supervision directly to reference image attention. We have also added a complementary “Router w/o Lrouter\mathcal{L_{\text{router}}}” ablation. The results are reported in Tab. 7.

Under the 3-/4-reference settings, Direct attention mask loss achieves Average scores of 0.7821/0.7601, and Router w/o Lrouter\mathcal{L_{\text{router}}} achieves 0.7908/0.7593, compared with 0.7958/0.7805 for the full method. The full method also achieves the highest Subject scores in both settings. These results indicate that neither spatial supervision alone nor routing without explicit mask supervision matches their combination in overall performance. We therefore attribute the improvement to their complementary contributions, rather than to either component alone.

Regarding the concerns about statistical reliability, our original PSRBench results were already averaged over four random seeds (42–45). We have now added standard deviations to PSRBench results. We have also added CLIP-I and CLIP-T results to Table 1. Since PSRBench’s current subject-consistency metric is based on DINO similarity, reporting DINO separately would be redundant. Our method also achieves strong scores on the additional CLIP-I and CLIP-T metrics, with clear improvements over the baseline.

3. Placeholder placement. The opening of Section 3.1 describes a common practice rather than our implementation. Our implementation is consistent with the description in Sec. 4.4 and Appendix A. On the benchmarks, subject descriptions are provided directly by the benchmark and are already included in the prompts. We locate each description’s token span at the T5 token level and insert learnable placeholder embeddings immediately after the corresponding span in the encoded embedding sequence. No placeholder tokens are inserted into the raw prompt. This avoids perturbing contextual semantics during text encoding, which we observed can reduce semantic alignment when placeholders are inserted before encoding. DreamO also adopts this pre-encoding insertion scheme and exhibits limitations on semantic alignment metrics. We have clarified this procedure in Sec. 3.1 and revised Figure 3 by removing placeholders from the prompt and retaining them only in the “Text embeddings” representation, consistent with our actual implementation.

AuthorsOpenReview ↗

Response to reviewer 3Nna (part 2 of 2)

4. Inference overhead. On the 3-reference subset, our method takes 11.1 s to generate an image—2.5 s longer than the baseline and 0.3 s longer than with the previous routing range. The router is the main source of the additional runtime, accounting for approximately 2.2 s. Peak GPU memory usage is 35.9 GiB, an increase of 3.7 GiB over the baseline. Expanding the routing range adds only about 4 MiB of peak memory usage, which is negligible in comparison.

5. Performance gains and limitations. Regarding Weakness 2, since the compared methods, except OminiControl, use different training datasets, and we cannot reproduce the training procedures of methods such as DreamO, we cannot precisely attribute the performance differences to specific factors. The most direct evidence of our method’s effectiveness comes from comparison with OminiControl, which uses the same pretrained model and training data as ours. Our implementation is also built on the OminiControl codebase. Compared with OminiControl, our method improves overall performance, particularly in the 4-reference setting. It also substantially improves ID scores on XVerseBench. Our method still falls considerably behind the best-performing method on the ID metric. We hypothesize that this gap may partly stem from insufficient human samples in our training data. Preserving human identity in personalized generation is indeed an important research topic, but it is not the primary focus of our current work.

6. Practical details. Regarding Weakness 5, we have added details on hyperparameter selection and training in Sec. 4.1 and Appendix A. We found that applying the router throughout the entire generation process yielded larger gains. Rather than testing every possible routing range, we used the analysis in Appendix A to select candidate ranges and adopted the best-performing configuration among those tested. An exhaustive search over all hyperparameter settings is infeasible; we therefore select a well-motivated, reasonable configuration rather than seek a globally optimal one.

7. Additional related work. Regarding Weakness 7, we incorporate the additional related work into Sec. 2 and clarify the following differences.

MIP-Adapter uses text-derived spatial relevance to weight each reference’s attention output. In contrast, our router normalizes across reference and background branches to explicitly model their competition. Within each attention computation, our common offset preserves the aggregate attention mass assigned to reference tokens relative to the non-reference branches.

GroundingBooth associates subjects with spatial regions through explicit target bounding boxes. Our method instead derives localization priors from the model’s own text attention and aligns reference injection with the semantic layout emerging from the pretrained model, rather than imposing an externally specified layout.

MultiCompose combines independently fine-tuned concepts through a multi-stage pipeline that uses automatic detection and segmentation to obtain masks for spatially fusing subject-specific predictions. Our method requires neither per-subject adaptation nor inference-time masks: after generic training, it handles new reference subjects by routing reference attention within a unified model. It therefore avoids both per-concept fine-tuning and the separate detection and segmentation stages.

8. Corrections. Regarding the Minor Weakness, the reported 3-reference results for XVerse on XVerseBench were mistakenly copied from the 2-reference results for Qwen-Image-Edit. We have corrected this error in the revised manuscript. The correct 3-reference results for XVerse are as follows:

MethodDPGAESIPIDAvg
XVerse80.9452.3658.8561.0263.29

As the number of references increases to three, XVerse’s DPG and AES scores on XVerseBench decline substantially, consistent with the collage-like artifacts observed in the qualitative results (Fig. 5). We have also corrected the misattribution of Gal et al.’s textual inversion to Mokady et al.’s null-text inversion. Thank you for pointing out these issues.

Thank you for the detailed response, which addresses several of my concerns. My main remaining concern is the learned router. In Table 7, the “Prior only” variant without a learned router is close to the full method (Subject 0.5243 vs. 0.5275 and 0.4686 vs. 0.4892), and the router without mask losses even lowers Subject. So the gain seems to come mainly from using text attention as a prior, not from the learned router itself. Could the authors clarify what the learned router adds beyond the prior?

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

This paper presents Bind-and-Route for multi-subject personalized image generation. Built on FLUX.1-dev and OminiControl, it links each subject’s text description to its reference image and uses text attention to guide where reference information is applied. The method aims to reduce subject confusion and improve prompt following, without requiring masks or per-subject training at inference time.

Critical strengths
  1. The method is well motivated and introduces useful technical ideas. Explicit binding establishes which reference subject a textual description refers to, while spatial routing determines which reference image each generated region should attend to. The two components are complementary and directly address the failure modes discussed in the paper.

  2. The experiments are extensive and include strong baselines such as Qwen-Image-Edit and OmniGen2. The improvements are particularly clear in the four-reference setting. Qualitative results also show improvements in subject interactions, prompt following, and reference appearance preservation.

  3. The paper is clearly written, and the figures make the method easy to follow.

Critical weaknesses
  1. The router is enabled in blocks 11–40 and within the timestep interval [0.3, 0.7]. While the appendix briefly explains how these ranges were chosen, ablation results would help clarify the reasons for these choices and the method’s sensitivity to different ranges.

  2. The method is evaluated only on FLUX.1-dev. It remains unclear whether explicit binding and routing would provide additional benefits on newer models with existing reference-image generation capabilities, such as FLUX.1 Kontext and FLUX.2-dev.

  3. Some figures are difficult to read, particularly Fig. 3 and Fig. 4. Larger figures or close-up views would improve readability.

Questions for authors
  1. Do the OminiControl baseline in Tables 1 and 2 use the same FLUX.1-dev backbone as your method?

  2. I am curious how the method compare with FLUX.1 Kontext and FLUX.2-dev on multi-reference subject generation? Have you tried applying the binding and routing components to these models?

  3. How were the router’s block and timestep ranges selected? How do performance and computational cost change when the boundaries are varied, or when routing is enabled across all blocks or the full denoising trajectory?

  4. In Table 4, all five values for Qwen-Image-Edit in the two-reference setting are identical to those for XVerse in the three-reference setting. Is this a typo?

5 comments
AuthorsOpenReview ↗

Response to reviewer pJps

Please refer to our revised manuscript and the Summary of Changes in the Revised Manuscript comment. Below is our response to the questions:

1. Baseline comparability. Yes, the OminiControl baseline also uses FLUX.1-dev and is trained on the same dataset as our method.

2. Backbone selection. During baseline reproduction, we also experimented with initializing OminiControl from FLUX.1 Kontext rather than FLUX.1-dev. The PSRBench results are summarized below:

SettingInitializationSubject ↑Semantic ↑Aesthetic ↑Average ↑
3-referenceFLUX.1 Kontext0.46530.76110.76280.6631
3-referenceFLUX.1-dev0.51900.86380.94690.7766
4-referenceFLUX.1 Kontext0.43940.73190.77640.6492
4-referenceFLUX.1-dev0.40260.85430.94620.7344

Except for Subject under the 4-reference setting, Kontext initialization yields lower scores in our baseline-reproduction experiments. These results motivated our choice of FLUX.1-dev, which is also the backbone adopted by OminiControl and several compared personalization methods. FLUX.2-dev represents a substantial change in architecture and scale: its transformer has 32B parameters, compared with approximately 12B for the FLUX.1-dev transformer. The transformer weights alone require approximately 64 GB in BF16, before accounting for training activations and other memory requirements. Adapting and training our framework on this backbone would therefore require substantial additional computation and implementation effort, beyond the scope of the current study.

3. Routing ranges. In Appendix A, we have added attention visualizations across different blocks and timesteps, which served as an important basis for selecting the routing ranges. We also evaluated different routing block and timestep ranges on the 3-reference subset of XVerseBench, with the results reported in Table 4 of Appendix A. We found that applying the router to all blocks throughout the entire denoising process yielded better performance and therefore adopted this configuration in the revised manuscript. An exhaustive search over all hyperparameter configurations is infeasible; reasonable adjustments to these hyperparameters do not alter our contributions or conclusions. Expanding routing from blocks [11, 40] and timesteps [0.3, 0.7] to all blocks and timesteps adds 0.3 s per image on the 3-reference subset, bringing the total generation time to 11.1 s per image (2.5 s more than the baseline for the full method). Peak GPU memory usage increases by approximately 4.3 MiB, which is negligible relative to the total usage of 35.9 GB.

4. Corrected XVerse results. Yes, we confirmed that the reported 3-reference results for XVerse on XVerseBench were mistakenly copied from the 2-reference results for Qwen-Image-Edit. We have corrected this error in the revised manuscript. The correct results are as follows:

MethodDPGAESIPIDAvg
XVerse80.9452.3658.8561.0263.29

As the number of references increases to three, XVerse’s DPG and AES scores on XVerseBench decline substantially, consistent with the collage-like artifacts observed in the qualitative results (Fig. 5). Thank you for pointing out this error.

5. Figure readability. Regarding Weakness 3, we have adjusted the image and text sizes in Figs. 3 and 4 in the revised PDF to improve readability. Thank you for the suggestion.

Reviewer_pJpsOpenReview ↗

Reply

Thank you for the clarification and additional experiments. Beyond the backbone initialization comparison, could you provide a direct quantitative and qualitative comparison with the original FLUX.1 Kontext?

AuthorsOpenReview ↗

Response on FLUX.1 Kontext zero-shot

Thank you for your response!

We have added the quantitative results for FLUX.1 Kontext in the zero-shot setting:

Method3-ref Subject3-ref Semantic3-ref Aesthetic3-ref Average4-ref Subject4-ref Semantic4-ref Aesthetic4-ref Average
FLUX.1 Kontext zero-shot0.48420.67050.78520.64660.43160.56700.78270.5938
FLUX.1 Kontext baseline0.46530.76110.76280.66310.43940.73190.77640.6492
FLUX.1-dev baseline0.51900.86380.94690.77660.40260.85430.94620.7344
Ours0.52750.89610.96390.79580.48920.86440.98790.7805

We have added the qualitative and quantitative comparisons in Appendix Fig. 11 and Table 8. FLUX.1 Kontext exhibits some ability to generate multiple subjects in the zero-shot setting, but it was primarily trained with single-reference conditioning. Under multi-reference inputs, its semantic alignment is substantially lower than that of the other methods in this comparison. Its training does not explicitly address semantic alignment in multi-subject generation. Examples in Fig. 11 further show that it often generates only one of the target subjects.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Changes in the Revised Manuscript

1. Routing scope and hyperparameter selection. We revised the hyperparameters of the final method, changing the spatial router’s scope from blocks [11,40] and timesteps [0.3,0.7] to the entire generation process. Following the reviewers’ suggestion, we expanded the hyperparameter comparison (see Appendix A, Table 4). The updated hyperparameters show more pronounced advantages across the metrics. We emphasize that the hyperparameters only change the scope over which the method operates. Since it is not feasible to test every possible combination experimentally, we cannot determine the globally optimal combination; we can only select the best result from the combinations we tested. The hyperparameter values do not affect the methodological contribution, and the overall conclusions remain unchanged. The only drawback is that the single-image generation time of our full method increases from 10.8 s to 11.1 s (+0.3 s) per image on the 3-reference subset, while peak GPU memory usage increases by approximately 4.3 MiB, which is negligible relative to the total usage of 35.9 GB. We consider this an acceptable trade-off, and we have confirmed the runtime in Section 4.1 and Appendix A.

2. Quantitative attention-alignment analysis. We added quantitative metrics for evaluating how our method addresses Text-Reference Misalignment and Subject Confusion in Section 4.3, Table 2: Pearson correlation and soft IoU, which measure the consistency between the text attention and reference image attention distributions. Figure 9 further illustrates this improvement through visualizations of the two attention distributions.

3. Expanded ablation studies. We provide more comprehensive ablation studies in Appendix D. We note that individual components of the complete method may tend to improve particular metrics. However, the full method addresses the Text-Reference Misalignment and Subject Confusion problems introduced in Section 1 and achieves a more balanced performance across metrics, rather than optimizing for a single metric.

4. Methodological and evaluation enhancements. We now explain how placeholders are inserted into text embeddings in Section 3.1 and Figure 3, rather than leaving this detail to Section 4.4. We also added details on training, parameter selection, and module design to Section 4.1 and Appendix A. In Section 4.3, we further clarified the calculation of the metrics, added standard deviations, and introduced CLIP-I and CLIP-T as additional evaluation metrics.

5. Corrections and minor edits. We corrected several minor errors, including the 3-reference result for XVerse on XVerseBench in Table 6 and the misattribution of Gal et al.’s textual inversion to Mokady et al.’s null-text inversion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.