acceptodds
Under review as a conference paper at ICLR 2027

DomainShuttle: Toward Generalizable Cross-Domain Subject-driven Video Generation

Abstract

Subject-driven text-to-video (S2V) generation has attracted growing interest in academia and industry. An ideal model should faithfully preserve reference appearance for in-domain generation while retaining intrinsic subject features under flexible cross-domain transformations. This generalizability should extend across generation paradigms, from bidirectional fixed-length to streaming autoregressive generation. Existing methods primarily optimize in-domain fidelity by indiscriminately preserving reference appearance, limiting cross-domain flexibility in real-to-fantasy, fantasy-to-real, and real-fantasy interaction scenarios. To address this limitation, we propose DomainShuttle, a novel framework for flexible and high-fidelity video personalization. Specifically, Domain-MoT decouples video and reference processing, and applies Domain-aware AdaLN for domain-specific modeling of reference images. Spatiotemporal Reference Anchor RoPE (STRA-RoPE) temporally anchors reference image tokens and spatially organizes reference tokens with subject-aware offsets, supporting both generation paradigms. For bidirectional training, Cross-Pair Consistent Loss aligns velocity predictions across reference sets to reduce interference from subject-irrelevant features. To further demonstrate its generalizability across generation paradigms, we extend the bidirectional model to streaming autoregressive generation through chunk-wise causal training. Extensive experiments demonstrate that DomainShuttle achieves strong video quality, text controllability, and subject consistency in both in-domain and cross-domain scenarios and on both bidirectional and streaming autoregressive architectures. The anonymous project page with videos is available at https://DomainShuttle.github.io/.

Primary area:
generative models
Keywords:
Video Generation, Cross Domain, Personalized Generation, Streaming Autoregressive Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KQQkOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

DomainShuttle is a method for cross-domain, subject-driven video generation. It enables the generation of videos, such as transforming real people into an anime style, converting virtual characters into tangible objects, or creating interactions between real and virtual subjects, while faithfully adhering to prompts and preserving the subject's identity.

Critical strengths

Strengths:

  1. The paper clearly distinguishes identity preservation from appearance replication and addresses practically relevant scenarios, including multiple subjects and maintains the structure of the body, alternating between different areas, etc.
  2. The in-domain and cross-domain demos on the project page are visually impressive and demonstrate the strong capabilities of DomainShuttle.
  3. The experimental evaluation is comprehensive, with comparisons against a broad range of leading baselines and state-of-the-art results across the reported metrics. The ablation studies are thorough and clearly establish the contribution of each module.
  4. The paper is well written, with a clear structure and a coherent presentation.
Critical weaknesses

Weaknesses

  1. Although the training paradigms differ, the following recent and closely related works should be cited and discussed to better position the proposed method:
    • [1] Li W, Zhang Y, Long F, et al. ReactID: Synchronizing Realistic Actions and Identity in Personalized Video Generation[C]//International Conference on Learning Representations. 2026, 2026: 62463–62485.
    • [2] Xu X, Li Y, You S, et al. Smrabooth: Subject and Motion Representation Alignment for Customized Video Generation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 16130–16141.
    • [3] Wang W, Huang M, Tu Y, et al. Dualreal: Adaptive Joint Training for Lossless Identity-Motion Fusion in Video Customization[C]//2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025: 16565–16575.
  2. The DOMAIN-MOT design intuitively provides an effective way to incorporate reference images into the model through in-context conditioning. However, the paper offers limited analysis of why this architecture is particularly suitable for customization, leaving the design somewhat insufficient beyond its empirical effectiveness. While I find the design promising, I would appreciate deeper architectural insights, supported by visualizations or additional analyses that explain how it facilitates reference conditioning and identity preservation.
  3. The introduction presents streaming autoregression as a core contribution, whereas the method section describes only the training of a bidirectional model and does not explain any training procedure specific to streaming generation. This creates a substantial inconsistency between the claimed contributions and the technical presentation. Although the supplementary material provides some implementation details (L886–894), these are insufficient to establish streaming autoregression as a core methodological innovation. The authors should either substantiate this claim in the main paper or revise the contribution statement accordingly.
  4. Style control appears to rely solely on text prompts. The paper does not explore conditioning on visual style references, such as a painting that exemplifies a particular artistic style. This limits the demonstrated scope of style customization, and a discussion or evaluation of reference-based style control would strengthen the work.
Questions for authors

See the above weakness. Overall, I find this paper promising, with visually impressive results and a comprehensive experimental evaluation. If the authors adequately address my concerns and provide convincing supporting evidence, I would keep my acceptance.

4 comments

Thank you for your positive assessment and constructive suggestions. We address your four concerns below.

W1: Related work and task differences.

We have cited these works in Related Work (L142) and discussed their differences from our method in the appendix (L1384–L1392). ReactID focuses on identity preservation and action generation, while SMRABooth and DualReal jointly customize subject identity and reference motion. Our task instead emphasizes preserving subject features under flexible cross-domain transformations. To this end, we propose DomainShuttle to improve subject consistency and generation flexibility across visual domains.

W2: Architectural motivation and analysis of Domain-MoT.

The independent reference branch of Domain-MoT allows reference features to be processed separately from video features, while joint attention connects reference information to video generation. Domain-aware AdaLN further conditions reference processing on each subject's target-domain attributes. This design follows our motivation to preserve recognizable subject features while allowing domain-dependent appearance to change. To examine this motivation, we have added two ablation studies to the main paper (L503–L530).

(1) Target-domain information versus its injection mechanism. We compare Domain-aware AdaLN with a control that adds learned domain embeddings to the corresponding subject's reference features.

SettingGMEScore ↑CD-CLIP ↑CD-Score ↑DINO-I ↑CLIP-I ↑
Dual Self-Attn0.6720.6090.7120.3640.678
Dual Self-Attn + Domain Embedding0.6750.6140.7270.3710.681
Domain-MoT (Dual Self-Attn + Domain-aware AdaLN)0.6860.6280.7810.3800.687

Adding domain embeddings improves CD-Score by 2.1%. Domain-aware AdaLN further increases it with a relative improvement of 7.4% over additive embeddings. This shows that the performance gains are mainly from Domain-aware AdaLN, rather than merely from additional information.

(2) Effect of shuffling domain attributes. We randomly shuffle domain attributes at inference to examine their role in generation.

Target-domain attributesCD-CLIP ↑CD-Score ↑Style Score ↑DINO-I ↑CLIP-I ↑
Randomly shuffled0.6270.7830.8170.3780.677
Correct0.6390.8540.8830.3810.681

Shuffling reduces CD-Score and Style Score by 8.3% and 7.5%, respectively, while DINO-I and CLIP-I decrease by only 0.79% and 0.59%. These results suggest that domain attributes mainly guide domain transformations, while the identity similarity remains almost unaffected by our domain-aware modeling.

W3: Clarifying the streaming autoregressive contribution.

We clarify streaming autoregression as an extension to verify reference modeling. The original manuscript already describes streaming training in L314–L320. We initialize the streaming model from the trained bidirectional checkpoint and perform an additional 5,000 steps of chunk-wise causal training, retaining Domain-MoT and STRA-RoPE. The construction of streaming cross-attention masks is described in L210–L230. Appendix A further details the reference-aware self-attention and cross-attention masks, together with KV caching for streaming inference.

Following your advice, we have revised the introduction (L94) to describe DomainShuttle as a framework for subject consistency and generation flexibility across domains, “which we further extend to streaming autoregressive generation.” We have also revised the contribution statements in L108–L117, presenting streaming generation only in the third contribution: “We further extend DomainShuttle to streaming autoregression through chunk-wise causal generation to verify our reference modeling.”

These revisions position streaming generation as an extension and validation of our reference modeling.

W4: Reference-based style customization.

We clarify that our model was not trained on a dataset for reference-image-based style transfer. Following your advice, we evaluate it on 10 cases with style reference images as visual inputs without additional training. We have added comparisons in Fig. 7 of the revised manuscript (L479–L500). DomainShuttle transfers the reference style to both subjects and backgrounds. In contrast, baselines either only transform the background style or fail to ensure that both the background and the subject adhere to the reference style. We use the main paper’s Style Score, computed using a simple text label describing the style of the reference image.

MethodStyle Score ↑
HuMo0.708
MAGREF0.726
VACE0.768
FFGO0.782
RefAlign0.814
Ours0.863

Our method achieves the highest Style Score, with a 6.02% relative improvement over baselines. Together, these results support the potential of DomainShuttle for reference-based style customization without task-specific training.

I thank the authors for their response. Their explanations have addressed my concerns, the experiments on Domain-MoT are sufficiently thorough, and the revisions to the manuscript are appropriate. I therefore maintain my recommendation to accept the paper and have adjusted my confidence score accordingly. I hope the authors will retain these revisions in the final version.

Dear Reviewer KQQk,

Thank you for your positive feedback and continued support. We appreciate your constructive suggestions, which helped us strengthen the experimental analysis and clarify the manuscript. We will retain these revisions in the final version. Thank you again for your time and thoughtful review.

Best regards,

The Authors

Since the rebuttal period is coming to an end, I would like to reiterate my positive assessment of this work. I believe the paper makes a valuable engineering contribution to the field, and the quality of the results is sufficiently impressive to warrant acceptance at ICLR. I also appreciate the authors' efforts to thoroughly address the reviewers' concerns and to demonstrate the system through compelling demos.

Although the other two reviewers have not yet responded, I am willing to raise my score to 4 to reflect my confidence in the work and my recognition of the substantial effort the authors have made during the rebuttal period.

As a researcher working on video stylization and personalized generation, I have particular confidence in the practical value and potential impact of this work. I believe the paper would be a valuable contribution to the community.

Reviewer_6Av5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets subject-driven video generation in which subjects must either be preserved as is (in-domain) or transformed across visual domains (real-to-fantasy, fantasy-to-real, real-fantasy interactions). It also aims to work in both bidirectional and streaming autoregressive generation. DomainShuttle combines three components:

  • Domain-MoT: separate video and reference branches in joint attention, with a reference AdaLN conditioned on a 4-way target-domain label that an MLLM annotates from the prompt.
  • STRA-RoPE: references anchored at temporal index 0, with subject- and image-specific spatial offsets.
  • A stop-gradient Cross-Pair Consistent Loss between predictions under two reference sets.

The bidirectional model is further fine-tuned into a chunk-wise causal streaming model. Evaluation uses 120 OpenS2V-Eval in-domain cases and 120 new cross-domain cases, scored with CLIP/DINO similarity, MLLM-judged CD-Score, VQAScore-based Relation Alignment, and Style metrics. Additional evidence comes from ablations (Table 2, 13), a 500-frame streaming comparison against Helios-based cascades (Fig. 3, Table 12), and two user studies (App. C.3).

Critical strengths
  1. Controlled, cumulative ablations with data controls. Table 2(a) moves from a naive concatenation baseline trained on the same data (CD-Score 0.696) to the full model (0.854), with each step isolated. Table 13 shows the method still leads without the Ditto-1M editing data. These are the main reasons I believe the cross-domain gains come from the design rather than from extra data or backbone differences.
  2. Practically relevant problem and broad comparison. The paper studies cross-domain S2V, including real-fantasy interactions and carrier binding. It compares against the closed-source Kling 1.6 and eight open-source baselines on Wan2.1/2.2, plus Helios-based streaming cascades. Within matched backbones, the method leads clearly on CD-Score and Style Score, which indicates a real improvement on an underexplored axis.
Critical weaknesses
  1. Limited methodological novelty, and an under-analyzed label input.

    Separate reference branches with joint attention follow MoT and Video-As-Prompt. For STRA-RoPE, the advantage of subject-aware grouping over offsetting every reference image along both axes is shown only in a single qualitative example (Fig. 12). CCL is a standard stop-gradient consistency term. The most distinctive element is conditioning the reference AdaLN on a per-subject 4-way target-domain attribute, which an MLLM annotates from the prompt at test time (Sec. 3.2.1, App. C.1). In Table 2(a), ID-1 and ID-2 differ only in this conditioning, and CD-Score rises from 0.712 to 0.781. Since the reference tokens in ID-1 already receive the text condition through cross-attention, the label helps beyond the prompt. However, because the label enters only through AdaLN, it remains unclear whether the gain comes from the per-subject domain information or from the AdaLN injection itself. A control that supplies the same label in another form, such as a learned token per subject, would separate the two. The paper also does not test robustness to wrong or ambiguous labels, which matters because the labels are produced automatically at test time. The novelty rests largely on this one component, and its analysis is incomplete.

Questions for authors
  1. Test-time domain labels (Sec. 3.2.1, App. C.1). Were the MLLM-annotated labels for the test set manually checked or corrected, and how often were they wrong? How does performance change when labels are shuffled or deliberately wrong at inference?

  2. Human evidence for the cross-domain claims (App. C.2, C.3). How is the system-level correlation between CD-Score and expert ratings computed? If it is over per-method means, it rests on only four data points, so please report it over individual generated videos. Can you also report the user study results for cross-domain cases separately?

  3. Benchmark release. The main claims rest on the self-constructed cross-domain set, but the only code the reproducibility statement provides is for bidirectional inference. Will you release the 120 cases (references, prompts, domain labels, edited references) and the evaluation scripts?

2 comments

Thank you for recognizing the practical relevance of cross-domain personalization. We address your concerns below.

W1 & Q1: Component contributions and target-domain conditioning.

We would like to clarify how each component contributes to our goal of preserving subjects while enabling cross-domain transformations.

  • Domain-MoT. As discussed in Appendix C.4.2, our use of separate branches with joint attention follows MoT and Video-As-Prompt. However, our conditioning design and task differ. For conditioning design, Domain-MoT introduces Domain-aware AdaLN into the reference branch to condition its processing on each subject's target-domain attributes, improving cross-domain generation beyond separate branches alone. For tasks, Video-As-Prompt uses reference videos to guide video generation, whereas we focus on cross-domain subject-driven generation from reference images.
  • STRA-RoPE. STRA-RoPE temporally anchors and spatially organizes reference tokens. In Table 2(a), IDs 2–3, we already compare it with Temporal-Offset RoPE to show our model's effectiveness. CD-Score increases from 0.781 to 0.808, and GMEScore from 0.686 to 0.693. To address your specific concern about subject-aware grouping, we additionally compare subject-decoupled and reference-decoupled spatial offsets below.
  • CCL. CCL reduces interference from subject-irrelevant features by aligning predictions under different reference sets paired with the same target video. With the same reference pool, adding CCL improves CD-Score from 0.808 to 0.854 (Table 2(a), IDs 3–4), supporting its benefit beyond cross-pair data alone.

To further address your concerns, we present three additional ablations below, examining subject-aware reference grouping, different target-domain information injection mechanisms, and shuffled domain attributes at inference.

(1) Quantitative analysis of subject-aware reference grouping.

Following your suggestion, we quantitatively compare the two spatial-offset strategies on 10 cross-domain cases:

Reference positioningGMEScore↑CD-CLIP↑CD-Score↑
Reference-Decoupled STRA-RoPE0.7010.5920.825
Subject-Decoupled STRA-RoPE (ours)0.7250.6100.850

Subject-Decoupled STRA-RoPE groups references of the same subject through a shared height offset, while Reference-Decoupled STRA-RoPE separates each image along both spatial axes. As shown in the table above, our method performs better across all three metrics. These results support our motivation: grouping multiple references of the same subject helps improve semantic consistency. The qualitative example in Appendix C.7 (Fig. 13) also shows that Subject-Decoupled STRA-RoPE better binds features from multiple references of the same subject, improving semantic consistency.

(2) Different target-domain information injection mechanisms.

Following your suggestion, we have added an ablation that uses the same target-domain attributes through learned embeddings. The embedding baseline maps the four domain attributes to learnable embeddings and adds each embedding to the corresponding subject's reference features.

SettingGMEScore ↑CD-CLIP ↑CD-Score ↑DINO-I ↑CLIP-I ↑
Dual Self-Attn0.6720.6090.7120.3640.678
Dual Self-Attn + Domain Embedding0.6750.6140.7270.3710.681
Domain-MoT (Dual Self-Attn + Domain-aware AdaLN)0.6860.6280.7810.3800.687

Adding domain embeddings only improves CD-Score by 2.1%, whereas Domain-aware AdaLN significantly increases it, with a relative improvement of 7.4%. This shows that the performance gains are mainly from Domain-aware AdaLN, rather than merely from additional information.

(3) Domain attribute accuracy and sensitivity to incorrect conditioning.

We manually checked the MLLM-generated test attributes and found an approximately 0.6% subject-level annotation error rate, which indicates that the MLLM’s annotation exhibits a relatively high degree of robustness. To address your question, we compare randomly shuffled domain attributes with correct attributes at inference, as shown below:

Target-domain attributesCD-CLIP ↑CD-Score ↑Style Score ↑DINO-I ↑CLIP-I ↑
Randomly shuffled0.6270.7830.8170.3780.677
Correct0.6390.8540.8830.3810.681

As shown in the table above, randomly shuffling the attributes has a minimal impact on subject similarity, with DINO-I and CLIP-I scores decreasing by only 0.79% and 0.59%, respectively. This demonstrates that our domain-aware modeling is robust with respect to subject similarity. Meanwhile, in cross-domain scenarios, shuffling reduces the CD-Score by 8.3% and the Style Score by 7.5%. These results support our motivation for target-domain guidance. Domain attributes mainly influence cross-domain performance, while identity similarity remains almost unaffected by our domain-aware modeling.

Q2: Human evidence for cross-domain performance.

The originally reported system-level Pearson correlations in Appendix C.2, 0.837 and 0.812, were computed over the four per-method means. Following your suggestion, we additionally compute correlations across individual generated videos, pairing each video’s CD-Score with its mean expert rating. The correlations are 0.541 for GPT-5.2 and 0.523 for Qwen3-VL-8B-Instruct, providing evidence of positive alignment with expert judgments at the individual-video level. For reference, VideoRealBench [1] reports a video-level Pearson correlation of 0.5707 for its trained evaluator on video realism assessment. Our correlations fall within a similar numerical range.

Cross-domain user study. We separately report below the cross-domain subsets of the two user studies described in Appendix C.3. In the five-method ranking study, the cross-domain subset is shown below. Our method shows the highest preference across all three criteria.

MethodOverall Video Quality ↑Text Controllability ↑Open-Domain Subject Consistency ↑
Ours4.414.574.81
Kling-1.63.633.313.16
SkyReels-V32.482.712.45
Phantom2.342.332.11
VACE-14B2.142.082.47

The cross-domain pairwise comparison with RefAlign is shown below:

CriterionRefAlign preferredTieOurs preferred
Overall Video Quality13.7%17.4%68.9%
Text Controllability16.8%12.5%70.7%
Open-Domain Subject Consistency22.9%11.3%65.8%

Both studies therefore support human preference for our method on cross-domain cases.

Q3: Benchmark release.

We plan to release all 120 cross-domain cases, including reference images, prompts, subject-level domain attributes, edited references, and evaluation scripts, subject to the required organizational review and approval.

[1] Min Yang et al. VideoRealBench: A Chain-of-Thought Realism Evaluation Benchmark for Generated Human-Centric Videos. CVPR 2026, pp. 18565–18575, Table 4.

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

This paper proposes DomainShuttle, a framework for generalizable subject-driven video generation across visual domains and generation paradigms. The framework aims to preserve subject identity while enabling flexible transformations between different visual domains, including real-to-fantasy, fantasy-to-real, and real-fantasy interaction scenarios.

DomainShuttle consists of three principal components. First, Domain-MoT employs separate processing branches for reference images and video tokens, incorporating target-domain information into the reference branch through domain-aware Adaptive Layer Normalization (AdaLN). Second, Spatiotemporal Reference Anchor RoPE (STRA-RoPE) introduces fixed temporal anchors and subject-aware spatial offsets to organize reference tokens, facilitating reference conditioning in both bidirectional and autoregressive generation. Third, Cross-Pair Consistent Loss (CCL) encourages consistent velocity predictions across different reference sets associated with the same target video, aiming to suppress subject-irrelevant variations while preserving identity-related information.

The authors conduct extensive experiments on Wan2.1 and Wan2.2 backbones, evaluating the proposed framework in both in-domain and cross-domain settings. The method is further extended to streaming autoregressive video generation through chunk-wise causal training. The reported results demonstrate improvements in cross-domain controllability, subject consistency, and long-form video generation compared with several existing approaches.

Critical strengths
  • A practically important and relevant problem formulation. The paper addresses an important limitation of existing subject-driven video generation methods: the difficulty of preserving subject identity while allowing domain-dependent appearance attributes to change according to textual instructions. This problem is particularly relevant to practical video personalization, where maintaining recognizable subject characteristics should not come at the expense of generative flexibility. The evaluation encompasses several challenging scenarios, including real-to-fantasy transformations, fantasy-to-real transformations, and interactions between real-world and fantasy subjects. These settings extend beyond conventional in-domain personalization and provide a broader perspective on the challenges associated with cross-domain subject consistency. The reported improvements across multiple evaluation metrics further demonstrate the practical potential of the proposed framework.
  • A coherent reference-conditioning framework supported by component-level evaluations and autoregressive applicability. The proposed framework integrates domain-aware reference modulation, structured positional encoding, and cross-reference consistency regularization into a unified architecture. These components collectively address different aspects of reference conditioning, including target-domain adaptation, reference-token organization, and robustness to variations in reference images. The ablation studies provide empirical evidence supporting the effectiveness of the individual components. Furthermore, the extension to streaming autoregressive generation demonstrates the potential applicability of the framework beyond fixed-length video synthesis. The reported long-form generation results suggest that maintaining structured reference conditioning throughout autoregressive generation can help mitigate subject inconsistency over extended temporal horizons.
Critical weaknesses
  • The conceptual novelty and the claimed disentanglement mechanism are not sufficiently established. Although DomainShuttle demonstrates encouraging empirical improvements, the methodological contributions appear largely incremental when considered individually. The principal components are closely related to established techniques, including mixture-of-transformers architectures, conditional adaptive normalization, reference-aware positional encoding, and consistency-based regularization. While their integration is technically coherent, the paper does not sufficiently articulate the fundamental methodological insight that distinguishes the proposed framework from existing reference-conditioning approaches. More importantly, the central motivation emphasizes disentangling intrinsic subject identity from domain-dependent appearance attributes. However, it remains unclear whether the proposed architecture explicitly achieves such disentanglement. In particular, Domain-MoT introduces target-domain information into the reference branch through domain-aware AdaLN, which effectively provides an additional conditioning mechanism. This architectural modification does not, by itself, establish that identity-related representations are separated from domain-specific variations.
  • The claimed generalization across generation paradigms is not sufficiently distinguished from architecture adaptation and additional autoregressive training. The paper presents generalization across bidirectional and streaming autoregressive generation as a major contribution. However, the experimental methodology does not clearly establish whether the proposed framework learns representations that inherently generalize across generation paradigms or whether the observed performance primarily results from dedicated adaptation to the autoregressive setting. Specifically, the streaming autoregressive model is initialized from the bidirectionally trained model and subsequently optimized through additional chunk-wise causal training. This procedure demonstrates that the proposed reference-conditioning architecture can be successfully adapted to autoregressive generation. However, such adaptability is conceptually different from demonstrating intrinsic cross-paradigm generalization of the learned representations or conditioning mechanisms.
Questions for authors

My primary questions concern the two weaknesses discussed above.

  • Could the authors clarify what distinguishes the proposed domain-aware reference conditioning from conventional conditional feature modulation at the representation level? In particular, what evidence supports the claim that Domain-MoT disentangles intrinsic subject identity from domain-dependent appearance, rather than simply introducing additional target-domain conditioning? A controlled analysis of learned representations or domain interventions would help clarify this distinction.
  • Could the authors clarify the intended definition of generalization across generation paradigms? Specifically, how can the benefits of the proposed reference-conditioning mechanism be distinguished from those arising from additional chunk-wise causal training? A comparison under more closely matched autoregressive training and reference-conditioning settings would help substantiate the claimed generalization.
1 comment

Thank you for recognizing the practical importance of our problem and the effectiveness of our framework. We clarify the role of domain-aware reference conditioning and the scope of our cross-paradigm claims below.

Q1: Domain-aware reference conditioning and domain interventions.

Domain-MoT uses Domain-aware AdaLN as a modulation mechanism. Its distinction lies in conditioning each subject’s reference processing on its desired target domain. At the representation level, each subject’s reference features are modulated by the diffusion timestep and its desired target-domain attributes, while the video branch remains timestep-conditioned. Here, we would like to clarify that “decoupling” in our paper refers to separate video and reference processing (L93–L95, L111–L112) and modulation (L232–L234). Following your advice, we have added two controlled ablations to the main text (L503–L530) to demonstrate that our approach is not simply introducing additional target-domain conditioning. The details are as follows:

(1) Target-domain information versus its injection mechanism. We compare Domain-aware AdaLN with learned embeddings added to each subject’s reference features, using the same target-domain attributes.

SettingGMEScore↑CD-CLIP↑CD-Score↑DINO-I↑CLIP-I↑
Dual Self-Attn0.6720.6090.7120.3640.678
+ Domain Embedding0.6750.6140.7270.3710.681
+ Domain-aware AdaLN0.6860.6280.7810.3800.687

Adding domain embeddings only improves CD-Score by 2.1%, whereas Domain-aware AdaLN improves CD-Score by 7.4% over additive embeddings supplied with the same domain information. This supports the importance of our injection mechanism. The performance gains are mainly from Domain-aware AdaLN, rather than simply from additional target-domain conditioning.

(2) Intervention on target-domain attributes.

We compare randomly shuffled domain attributes with correct attributes at inference, as shown below:

Target-domain attributesCD-CLIP↑CD-Score↑Style Score↑DINO-I↑CLIP-I↑
Randomly shuffled0.6270.7830.8170.3780.677
Correct0.6390.8540.8830.3810.681

As shown in the table above, randomly shuffling the attributes has a minimal impact on subject similarity, with DINO-I and CLIP-I scores decreasing by only 0.79% and 0.59%, respectively. This demonstrates that our domain-aware modeling process robustly preserves subject similarity. Meanwhile, in cross-domain scenarios, shuffling reduces the CD-Score by 8.3% and the Style Score by 7.5%. These results support our motivation for target-domain guidance. Domain attributes mainly influence cross-domain performance, while identity similarity remains almost unaffected by our domain-aware modeling.

Q2: Scope of generalization across generation paradigms.

We would like to clarify that generalization across generation paradigms refers to the applicability and effectiveness of our reference-modeling design in both bidirectional and streaming autoregressive generation, with additional adaptation training for the streaming autoregressive setting. We retain Domain-MoT and STRA-RoPE across both paradigms.

The original manuscript already includes ablations to distinguish the benefit of reference modeling from that of additional causal training in Table 2(b) of the main text, IDs 6–8. All three variants undergo 5,000 steps of chunk-wise causal training from their corresponding bidirectional checkpoints. Under the same adaptation budget, the full model achieves a CD-Score of 0.773, compared with 0.713 without Domain-MoT and 0.747 without STRA-RoPE. These results support the benefit of our reference-modeling design under comparable autoregressive adaptation conditions.

Streaming settingCD-Score↑DINO-I↑CLIP-I↑
W/o Domain-MoT0.7130.3470.670
W/o STRA-RoPE0.7470.3750.681
Full model0.7730.3820.688

Since each variant retains its configuration across both training stages, these results evaluate the contribution of reference modeling across the complete training pipeline. Together with the bidirectional ablations in Table 2(a), they support the effectiveness of our reference-modeling design in both paradigms.

We have revised the abstract and introduction (L94) to position streaming generation as an extension of DomainShuttle. The revised contribution statements (L108–L117) describe it only in the third contribution:

We further extend DomainShuttle to streaming autoregression through chunk-wise causal generation to verify our reference modeling.

Together, these results support the applicability and effectiveness of reference modeling in both bidirectional and streaming autoregressive generation. The streaming ablations further show that the design provides benefits when all variants receive the same number of causal-training steps.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response of Submission790

Dear Reviewers and Area Chairs,

We sincerely thank all reviewers and the Area Chair for their time and comments. We especially appreciate the following feedback:

  • Reviewer KQQk raised both their rating and confidence from 3 to 4. They confirmed that our explanations addressed their concerns and that the Domain-MoT experiments were sufficiently thorough. We especially thank them for recognizing the practical value, potential impact, and contribution of our work to the community.
  • Reviewer 6Av5 gave a positive assessment and recommendation for acceptance. We thank them for recognizing the practical relevance of our work and suggesting further controlled experiments.
  • Reviewer ubEM recognized the practical importance of our problem and the coherent reference-conditioning framework. We thank them for suggestions that motivated additional domain interventions and clearer explanations of our architectural design and streaming contribution.

We have provided detailed responses and summarized the additional experiments and clarifications below.

Additional Experiments

Target-domain information versus its injection mechanism. Domain-aware AdaLN outperforms additive embeddings with the same target domain attributes. This demonstrates that the performance improvements are mainly from Domain-aware AdaLN, rather than simply from additional target-domain conditioning.

Interventions on domain attributes. Shuffling target-domain attributes reduces cross-domain and style performance, while the measured subject-similarity scores change only slightly. These results support our motivation for target-domain guidance. Domain attributes mainly influence cross-domain performance, while identity similarity remains almost unaffected by our domain-aware modeling.

Reference-based style customization. Without additional training, DomainShuttle achieves the highest Style Score among the compared methods when directly conditioned on style reference images. We add qualitative comparisons in Fig. 7 of the revised manuscript and quantitative results in the comment to Reviewer KQQk. The added qualitative results support our model's potential for reference-based style customization.

Subject-aware reference grouping ablation. Subject-Decoupled STRA-RoPE outperforms Reference-Decoupled STRA-RoPE, which supports that grouping multiple references of the same subject helps improve semantic consistency.

Human evidence for cross-domain performance. The additional analysis shows positive correlations between CD-Score and expert ratings. We also separately reported the cross-domain subsets of both user studies.

Clarifications and Explanations

Architectural motivation. All components of our model share the goal of improving cross-domain personalized video generation. Domain-MoT decouples video and reference processing and applies Domain-aware AdaLN for domain-specific modeling of reference images; STRA-RoPE temporally anchors and spatially organizes reference tokens; and CCL reduces interference from subject-irrelevant features. Following the reviewers’ suggestions, the additional comparison of target-domain information injection mechanisms and interventions on domain attributes further supports this design motivation and effectiveness.

Scope of the streaming contribution. We revised the main text to present streaming autoregression as an extension validating our reference-modeling design, with explicit additional causal adaptation. The original streaming ablations show that the full model outperforms variants without Domain-MoT or STRA-RoPE under the same adaptation budget. Together with the bidirectional results, this supports the design’s applicability and effectiveness in both generation paradigms.

Related work and test set. We added citations and discussion of more related work, clarifying their task differences from ours. We also clarified our plan to release the cross-domain test set, including reference images, prompts, domain attributes, edited references, and evaluation scripts, subject to the required approvals.

We hope these additional results and clarifications address the reviewers’ concerns and strengthen the paper. Thank you again for your time and thoughtful consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.