acceptodds
Under review as a conference paper at ICLR 2027

ReBind: Multi-Reference Video Editing via Structured Instructions with Explicit Reference Relationships

Abstract

Multi-reference video editing integrates content from multiple reference images while preserving relevant content from a source video. However, operation-centric instructions often leave preservation requirements and content-source relationships underspecified. Even detailed instructions may associate entities or attributes with the wrong visual sources, introducing inconsistent training supervision. To address these limitations, we introduce ReBind, a supervision framework that combines comprehensive editing and preservation semantics with accurate source attribution. As it’s core, ReBind-Instruct generates comprehensive, reference-consistent instructions through a two-stage training process. To learn these capabilities, Structured Instruction Learning establishes instruction generation through supervised fine-tuning, while Reference Attribution Optimization applies GRPO with content-level and relation-level rewards to improve semantic completeness and correctness and penalize source assignment errors. The resulting annotations provide scalable supervision for ReBind-Edit, which extends a pretrained text-to-video model to multi-reference video editing through joint visual conditioning on the source video and reference images. Controlled ablations show that corrupting source assignments degrades editing performance despite largely retained instruction quality, demonstrating the value of accurate attribution. ReBind-Instruct achieves the highest combined Instruction Quality and Reference Accuracy score among evaluated MLLMs, while ReBind-Edit achieves state-of-the-art open-source performance on IntelligentVBench and UniVBench across single- and multi-reference editing. We will publicly release both models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Mluti-Reference Video Editing, Explicit Reference Relationship

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

This paper introduces ReBind, a supervision framework for multi-reference video editing that combines explicit editing and preservation requirements with accurate source attribution. It addresses underspecified preservation constraints and incorrect assignments of entities or attributes to reference images. First, supervised fine-tuning trains ReBind-Instruct to generate structured instructions from source and edited video pairs together with reference images. A second stage applies GRPO, using content-level and relation-level rewards to improve semantic completeness and source consistency, with a length reward to limit verbosity. The resulting annotations train ReBind-Edit, an LTX-2.3-based editor that jointly conditions on the source video and multiple reference images. Benchmark comparisons and controlled ablations show that detailed instructions benefit editing and that corrupting reference assignments reduces editing quality even when descriptive content is largely preserved. On UniVBench and IntelligentVBench, ReBind-Edit achieves the best overall performance among the open-source methods evaluated.

Critical strengths
  1. The paper addresses an important challenge in multi-reference video editing: accurately associating edited attributes with their corresponding visual references while preserving source-video content. The proposed ReBind supervision provides a clear and structured way to represent editing requirements, preservation constraints, and reference relationships.

  2. ReBind-Edit achieves strong performance on IntelligentVBench and UniVBench, outperforming the compared open-source baselines. In particular, the controlled attribution-corruption experiment in Table 4 provides useful evidence that accurate reference assignments contribute to improved video editing performance.

Critical weaknesses
  1. The overall training pipeline mainly follows the conventional SFT + GRPO paradigm. Although the proposed reference attribution supervision is task-specific, its methodological novelty beyond existing instruction tuning and reward optimization approaches is not sufficiently established. It is unclear why the proposed lengthy ReBind instructions are necessary compared with simpler explicit reference bindings or structured entity–attribute–source representations. Moreover, the Dual Checklist Reward is relatively straightforward, and the authors do not demonstrate why separate preservation and modification checklists are more effective than a single unified checklist.

  2. Large multimodal models are extensively involved in data construction, reward computation, and evaluation. In particular, GPT-5.1 is used for both reward-related processing and final evaluation, potentially introducing evaluator-specific biases. The learned instructions may favor the detailed descriptions preferred by these models rather than those most effective for video editing. Independent evaluation would strengthen the conclusions.

  3. ReBind-Instruct is evaluated using target videos, whereas the actual editing pipeline uses Qwen3-Omni to rewrite user instructions without access to target videos (Section 4.3). Although the complete editing pipeline is evaluated, the reported instruction quality scores do not directly establish the reliability of the deployment-time rewriter. A separate analysis is needed to determine whether rewriting introduces errors or alters the original editing intent.

  4. Section 5.5.1 reports that a subset of reference assignments is corrupted but does not specify the proportion of affected samples or assignments, or how the subset is selected. Without these details, the 0.27-point performance decrease in Table 4 is difficult to interpret and the experiment is not fully reproducible.

Questions for authors
  1. In Section 4.3 , has the Qwen3-Omni rewriter been evaluated separately without target videos? Please report any available checks on whether rewritten instructions alter the original editing intent or preservation requirements, and clarify their impact on the final outputs. This would help assess the inference-time step that the target-conditioned instruction evaluation in Appendix C.2 does not directly test.

  2. For the attribution-corruption experiment in Section 5.5.1, what proportion of training samples and image-reference assignments were modified, and how was the affected subset selected? These details are needed to relate the 0.27-point drop in editing Overall to the annotation-error rate and to reproduce the controlled experiment.

  3. Appendix B.1configures eight reference slots but uses only one to five reference images during training. Does the eight-slot configuration describe input capacity only, or are editing results with six to eight references available? Clarifying this distinction would establish the tested operating range without equating architectural capacity with demonstrated editing reliability.

6 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer QhBm (1/3)

We sincerely thank you for your constructive review, especially for your positive comments on our motivation, method design, and quantitative results. We address your concerns point by point below.

W1. Contribution, instruction supervision, and checklist design

Thank you for these constructive comments!

1. Source attribution supervision and targeted optimization

ReBind contributes a supervision formulation and targeted optimization for multi-reference video editing, with SFT and GRPO serving as the optimization tools. The formulation jointly specifies editing requirements, preservation constraints, and the visual sources of entities and attributes. ReBind-Instruct learns to produce these annotations through complementary objectives for content quality and reference consistency.

Our key insight is that descriptive completeness and source-attribution correctness are distinct requirements for effective editor supervision. Table 4 provides controlled evidence for this distinction: corrupting selected image-reference assignments while preserving descriptive text, instruction format, visual training data, and editor training settings leaves IQ largely unchanged, but reduces Editing Overall by 0.27 points, with declines across all eight dimensions. Thus, detailed descriptions remain insufficient when their source assignments are incorrect.

Table 5 further supports optimizing attribution separately from descriptive quality: adding the relation reward improves RA, even though IQ decreases relative to the setting using the content reward alone. Together, these findings support explicit source attribution and its targeted optimization as complementary to descriptive enrichment, which constitutes the methodological focus of ReBind.

2. Instruction quality, attribution accuracy, and downstream editing

We evaluate instructions along two dimensions: Instruction Quality (IQ) measures semantic completeness and correctness, while Reference Accuracy (RA) measures source-attribution accuracy.

To examine the value of detailed descriptions beyond explicit reference bindings, we compare full ReBind instructions with compact versions derived from the same annotations. The compact instructions retain the original reference assignments and edit/preserve targets while removing detailed appearance descriptions. Both settings use the same visual training data, editor architecture, and training configuration, with 50K examples and 5K training steps.

Instruction supervisionInstruction evaluationEditor evaluation
IQ ↑RA ↑Sub. ↑Back. ↑Act. ↑Cam. ↑Col. ↑Light. ↑Style ↑Rel.Pos. ↑Overall ↑
Compact explicit bindings44.8558.424.865.305.727.445.856.484.836.645.89
ReBind instructions61.6859.605.085.385.767.466.076.575.066.676.01

The two variants achieve similar RA scores, while full ReBind instructions achieve substantially higher IQ. Full instructions improve Editing Overall by 0.12 points, with gains across all eight editing dimensions, particularly Subject, Color, and Style. These results support retaining detailed descriptions alongside explicit source bindings to improve fine-grained editing performance. We will include this additional comparison and its construction details in Section 5.5.1.

3. Purpose of the dual checklist design

The dual checklist design makes both preservation and modification requirements explicit in the content reward. Correctly describing the requested changes is insufficient if the instruction omits source content that should remain unchanged.

As described in Section 4.2.2 and illustrated in Figure 3, these requirements can concern different attributes of the same entity: its identity and clothing may change, while its motion and interactions should be preserved. We therefore derive keypoints for both aspects from the ground-truth instruction and jointly evaluate the generated instruction for missing, incorrect, and hallucinated content. This allows the reward to penalize omitted preservation requirements even when the requested modifications are accurately described.

Here, “dual” refers to the two complementary categories used to construct the supervision, rather than two independent evaluations. A unified checklist retaining the same requirements and scoring rules can implement the same design. We will clarify this distinction and the joint evaluation procedure in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal for Reviewer QhBm (2/3)

W2. Evaluator bias and independent evaluation

Thank you for raising this important concern! We address it through additional instruction evaluations with two alternative judges and independent assessments of downstream video editing performance.

We re-evaluate all models in Table 3 using Claude-Opus-5.5 and Gemini-4.0 Argon. The results are reported below:

ModelGPT-5.1 (original)Claude-Opus-5.5Gemini-4.0 Argon
IQ ↑RA ↑Overall ↑IQ ↑RA ↑Overall ↑IQ ↑RA ↑Overall ↑
Closed-Source Models
Gemini-2.5-Pro46.3763.8555.1155.0460.7357.8955.7038.8247.26
Gemini-3.1-Pro59.2452.9056.0764.9549.8457.3968.5032.1150.30
Open-Source Models
Qwen2.5-Omni-3B10.1021.8615.9819.2415.1317.1919.095.5912.34
Qwen2.5-Omni-7B14.3235.3724.8419.3732.4325.9019.8511.5315.69
Qwen3-Omni-Captioner22.0139.4730.7440.9636.8438.9043.2717.0530.16
Qwen3-Omni-Instruct31.8249.5940.7043.3450.0946.7143.4427.3035.37
Gemma4-31B-it49.4936.7143.1056.2431.2543.7558.3117.0237.66
ReBind-Instruct (Ours)61.6859.6060.6467.9552.7160.3370.4232.5851.50

ReBind-Instruct achieves the highest IQ and Overall among all evaluated models under both alternative judges, and ranks first among open-source models on all three metrics. Its Overall score exceeds its Qwen3-Omni-Instruct base by 13.62 points under Claude and 16.13 points under Gemini. These results provide support across judges for the instruction-level improvements, reducing the concern that the reported gains depend specifically on GPT-5.1.

Our evidence also includes independent evaluations of the edited videos themselves. On the single-reference IntelligentVBench benchmark, editing performance is evaluated using Gemini-2.5-Pro, where ReBind-Edit achieves the highest average score among all evaluated methods (Table 2). Separately, our human study on multi-reference editing asks participants to rank anonymized outputs based on instruction following, source-video preservation, reference-image preservation, and visual quality. ReBind-Edit receives the highest preference scores across all four criteria (Figure 4). These assessments directly evaluate editing outcomes and provide complementary evidence beyond instruction quality assessed by automated judges.

We will include these additional results and clarify the evaluation models and protocols in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal for Reviewer QhBm (3/3)

W3 & Q1. Reliability and impact of automatic instruction rewriting

Thank you for this valuable comment! We agree that the target-conditioned instruction evaluation in Table 3 does not directly establish the reliability of deployment-time rewriting. Automatic rewriting can introduce inaccuracies or deviations from the original editing intent, potentially affecting downstream editing performance. We therefore compared downstream editing performance using instructions generated by Qwen3-Omni-30B-A3B-Instruct and Claude-Opus-5.5, together with Claude instructions reviewed and corrected by human annotators.

Both automatic rewriters used only the original user instruction, source video, and reference images, without access to target videos. We evaluated all three instruction sources on the same 100 UniVBench MRVE cases, keeping the ReBind-Edit checkpoint, generation settings, and per-case random seeds fixed. All generated videos were evaluated against the original user instructions and benchmark visual inputs, rather than the rewritten instructions.

Instruction sourceSub. ↑Back. ↑Action ↑Cam. ↑Color ↑Light. ↑Style ↑Rel.Pos. ↑Overall ↑
Qwen3-Omni-30B-A3B-Instruct6.207.417.128.217.397.856.547.707.30
Claude-Opus-5.56.247.347.548.387.127.746.457.707.31
Human-corrected instructions6.337.677.438.617.377.796.687.947.48

Qwen3-Omni achieves an Overall score of 7.30, closely matching Claude-Opus-5.5 at 7.31 and falling 0.18 points below the human-corrected reference. These results show that the default rewriter maintains downstream editing performance close to the human-corrected condition on the evaluated cases, while manual correction provides further improvement.

We retain Qwen3-Omni as the default open-source rewriter to facilitate community use without requiring a proprietary rewriting API. We will include these results and the experimental details in the revised manuscript and investigate further improvements to deployment-time instruction rewriting.

W4 & Q2. Reference attribution corruption protocol

Thank you for requesting this clarification. We identified eligible multi-reference training samples containing correct image-reference assignments with incompatible alternative references in the same sample. We randomly selected eligible samples accounting for 30.0% of the training set.

Within each selected sample, we redirected each verified correct image-reference assignment with an incompatible alternative to one such alternative from the same sample, choosing randomly when multiple alternatives were available. The descriptive text, instruction format, and all other assignments, including <Video_1> references, remained unchanged. In total, 38.0% of image-reference assignments across the full training set were modified.

Both settings used the same visual training data, editor architecture, and hyperparameters. Under this corruption setting, Editing Overall decreased by 0.27 points, with declines across all eight editing dimensions while IQ remained nearly unchanged. We will add these statistics and the corruption procedure to Section 5.5.1.

Q3. Eight-slot input capacity and evaluated reference range

Thank you for asking us to clarify this distinction! The eight-slot configuration was preset during implementation to provide additional input capacity. After data construction and filtering, our training samples contained one to five reference images, while the original eight-slot configuration was retained. As described in Appendix B.1, unused slots are masked during attention.

Our reported evaluations cover one to five reference images; the benchmarks do not include cases with six to eight references. The model can process up to eight reference images without structural modification. In a small number of manual tests, we observed satisfactory editing results in some eight-reference cases. However, these successes have not yet translated into consistently reliable performance across cases.

We will clarify the distinction between input capacity, training coverage, and the evaluated reference range in Appendix B.1 and Section 5. Extending training coverage and improving robustness with six to eight references remain future work.

Reviewer_98mpOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ReBind for multi-reference video editing. It trains an instruction model through SFT and GRPO to describe both edits and preservation requirements, with explicit reference assignments. These instructions then supervise an LTX-based video editor. The paper evaluates instruction generation and downstream editing quality.

Critical strengths
  • The paper combines data construction, human annotation, instructor training, and downstream editor training. The controlled experiment in Table 4 shows that the improved instruction supervision benefits editing performance under the same data and training budget.

  • Evaluation covers single-reference and multi-reference editing, two benchmarks, and a human study. These experiments provide evidence for the usefulness of the complete pipeline and models.

Critical weaknesses
  • Could the authors clarify how Image_x and Image_y in the instruction are associated with the corresponding reference images? Section 4.3 states that references are ordered accordingly, while Appendix B.2 uses positional encoding that does not encode the reference index. How does the editor distinguish these indices?

  • The downstream benefit of the relation reward is not fully isolated. Table 5 compares reward variants through instruction scores, but does not report editing results for editors trained on annotations from the SFT-only and content-reward-only instructor variants.

Questions for authors
  • The planned release of both models is valuable. Will the authors also release the editing dataset and human-annotated instructions used for training and evaluation? Making these resources available would substantially increase the value of this work to the community.

  • With the rewritten instruction and random seed fixed, does swapping two reference images change the output according to their identifiers? This would help clarify how reference binding is implemented and working.

5 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer 98mp (1/2)

We sincerely thank you for your constructive review, especially for your positive comments on our motivation, method design, and quantitative results. We address your concerns point by point below.

W1. Reference binding and the discrepancy in Appendix B.2

Thank you for identifying this discrepancy. We apologize for the confusion. All reported ReBind-Edit results were obtained using models trained and evaluated with the slot-specific temporal indices described below. Appendix B.2 inadvertently retained an outdated shared-index description; correcting this description does not change the implementation or the reported results.

Reference slots are distinguished through slot-specific positional encoding. We retain the backbone's 3D RoPE and spatial patch coordinates (h, w). For the ordered reference inputs, we assign p_ref,k(h, w) = (-k, h, w) for k = 1, ..., K. Source and edited video frames use p_src(f, h, w) = p_edit(f, h, w) = (f, h, w) for f = 0, ..., F - 1.

Each reference slot therefore has a distinct positional encoding, and adding reference images does not change the video-frame coordinates. Unused reference slots are masked during attention. These positional indices are separate from the diffusion timesteps, which are zero for clean reference and source tokens and σt\sigma_t for edited-video tokens.

The association between textual identifiers and visual slots is learned through instruction-conditioned training. Throughout training and inference, <Image_k> denotes the reference placed in the kk-th indexed slot. The slot-specific positional encoding distinguishes the visual inputs, while consistently paired instructions and visual conditions teach the editor to associate each identifier with the corresponding reference content.

We will correct Eq. 10 and the accompanying shared-index description in Appendix B.2 and explicitly clarify this reference-index encoding and its association with textual identifiers in Section 4.3.

W2. Downstream editing ablation across instructor variants

Thank you for this constructive suggestion! To directly evaluate the downstream benefit of the relation reward RCR_C, we conducted an additional controlled ablation using the four instructor variants from Table 5.

We selected a fixed subset of 10K visual training triplets and generated separate instruction annotation sets using each of the four instructor checkpoints. We then trained a separate ReBind-Edit model on each annotation set while keeping the visual triplets, editor initialization, hyperparameters, and training budget identical across runs. Below we report the UniVBench MRVE editing results under identical sampling settings alongside each variant's Table 5 instruction metrics.

ModelInstructor EvaluationEditor Evaluation
IQ ↑RA ↑Sub. ↑Back. ↑Action ↑Cam. ↑Color ↑Light. ↑Style ↑Rel.Pos. ↑Overall ↑
SFT only56.8453.114.254.445.267.165.356.024.245.525.28
Content reward (RDR_D)63.8154.154.384.585.447.245.496.164.395.605.41
Content + relation rewards (RD+RCR_D + R_C)60.7460.074.825.065.587.325.826.364.706.025.71
Full (RD+RC+RLR_D + R_C + R_L)61.6859.604.865.115.627.355.866.404.746.065.75

Adding the relation reward RCR_C improves MRVE Overall by 0.30 points from 5.41 to 5.71 over training with the content reward alone, with gains across all eight editing dimensions. These results demonstrate the downstream benefit of incorporating RCR_C into instructor training.

Together with the corrupted source attribution experiment in Table 4, these results provide complementary evidence for the importance of accurate source attribution in editor learning. We will include these results and training details in Section 5.5.2.

Q1. Open-source release plan

Thank you for this valuable suggestion! We will publicly release both model checkpoints, relevant code and materials, together with the corresponding configurations and documentation, to ensure reproducibility.

AuthorsOpenReview ↗

Rebuttal for Reviewer 98mp (2/2)

Q2. Effect of swapping reference images

Thank you for this diagnostic question.

We distinguish two settings: jointly swapping the reference images and their textual identifiers and swapping only the images while keeping the instruction fixed.

1. Jointly swapping images and textual identifiers. When two reference images and all corresponding <Image_k> identifiers are swapped together, each target entity remains associated with the same reference content. The intended reference-to-entity assignments and editing requirements therefore remain unchanged.

2. Swapping only images under a fixed instruction. When only the images are exchanged, each unchanged identifier refers to the image newly occupying its corresponding slot. In the cases we inspected, the generated subjects switched according to these new assignments between identifiers and images. Because ReBind instructions also contain image-specific attribute descriptions, an image-only swap may additionally introduce inconsistencies between the frozen descriptions and the newly assigned images.

We evaluate the second setting on 100 randomly sampled multi-reference cases from UniVBench, keeping the rewritten instruction, source video, checkpoint, and random seed fixed. During evaluation, the benchmark instruction remains unchanged, and the reference images are provided in the same order used for generation.

Reference orderEditing Overall ↑
ReBind6.41
Swapped5.67

The observed subject switches provide evidence that the editor follows the new assignments between identifiers and images in the inspected cases. The decrease in Overall additionally measures the downstream impact of image-only reordering under the frozen instructions. We will include these observations, quantitative results, and evaluation details in the revised manuscript.

AuthorsOpenReview ↗

Seeking Your Feedback on Our Clarifications

Dear Reviewer QhBm,

We would like to express our appreciation for the valuable insights you provided in your review. Your comments have been instrumental in helping us clarify several important aspects of our work.

Following our detailed responses, we would be particularly interested in hearing your perspective on the explanations we have provided. Your feedback would help us better understand whether any substantive concerns remain or whether additional clarification would be beneficial.

As the discussion window is closing soon, we would sincerely appreciate any further thoughts you might be willing to share.

Thank you once again for your time and engagement.

Best regards,

The Authors

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

This paper introduces ReBind, a supervision framework for multi-reference video editing that comprehensively specifies both editing and preservation semantics, together with entity- and attribute-level source attribution. Methodologically, it proposes two components: ReBind-Instruct, a multimodal instruction generator trained through supervised fine-tuning (SFT) and GRPO with content-level and reference-attribution rewards; and ReBind-Edit, a video editor based on LTX-2.3 trained with the structured instructions from ReBind-Instruct. The paper constructs a synthetic dataset of approximately 100K editing samples and reports improvements over existing open-source methods on IntelligentVBench and UniVBench.

Critical strengths
  1. The paper is clearly organized and easy to follow.
  2. The motivation is well established. ReBind explicitly models entity–attribute–reference relationships and preservation requirements, effectively addressing the ambiguity of operation-centric instructions.
  3. The solid experiments convincingly demonstrate the effectiveness of the approach and ablation studies validate each component.
Critical weaknesses
  1. Although the paper includes a human study with 15 participants on 50 cases, the study remains relatively limited in scale. Broader human evaluation or cross-evaluator agreement analysis would provide stronger evidence.
  2. The training data pipeline relies on multiple generative models during inverse editing, synthetic reference-image generation, and model-based filtering periods. These steps may introduce model-specific biases and artifacts that accumulate across successive processing stages.
Questions for authors
  1. Can the authors clarify what generalizable insight distinguishes ReBind from conventional instruction enrichment and reward-based instruction optimization? In particular, which part of the explicit source-attribution formulation is essential to the observed editing gains, beyond simply providing longer and more comprehensive instructions?
  2. Will the 100K training samples, ReBind instruction annotations, and data-construction code be released? Open-sourcing would significantly benefit community research and reproducibility.
8 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer wWsJ (1/2)

We thank you for your constructive review and positive feedback. We address your concerns below.

W1. Scale of Human Evaluation and Inter-Annotator Agreement

Thank you for this constructive suggestion. To strengthen the human evaluation, we expanded it from 50 to 100 randomly sampled cases from the UniVBench multi-reference subset. Each of the five participants independently evaluated all 100 cases across all four dimensions: Instruction Following, Source Video Preserving, Reference Image Preserving, and Visual Quality. For each case, participants viewed the source video, reference images, and editing instruction, and ranked the anonymized outputs from the four compared models separately for each dimension.

The expanded evaluation results are presented below. We additionally report Krippendorff’s α\alpha [1] for each dimension to quantify inter-rater agreement.

MethodInstruction Following ↑Source Video Preserving ↑Reference Image Preserving ↑Visual Quality ↑
OmniWeaving23.417.526.820.6
Kiwi-Edit24.330.420.428.8
UniVideo57.159.057.154.9
ReBind-Edit (Ours)95.293.295.795.7
Krippendorff’s α\alpha0.740.660.750.72

ReBind-Edit outperforms all compared baselines across the four evaluation dimensions, with the largest margins over the strongest baseline in Visual Quality and Reference Image Preserving. These findings support its advantages in producing visually appealing edits while preserving reference appearance on the expanded evaluation set. Krippendorff’s α\alpha values range from 0.66 to 0.75, indicating good inter-rater agreement.

We will update Section 5.3 and Figure 4 with these results, scoring protocol, and inter-rater agreement details.

[1] Andrew F. Hayes and Klaus Krippendorff. Answering the Call for a Standard Reliability Measure for Coding Data. Communication Methods and Measures, 1(1):77–89, 2007.

W2. Model-Specific Biases and Artifact Accumulation in the Training Data Pipeline

Thank you for raising this important concern!

First, we select videos from two publicly available video datasets, OpenVid-1M and AVSET-10M, as supervision targets (VeditV_{\mathrm{edit}}) and construct the corresponding synthetic source videos (VsrcV_{\mathrm{src}}) and reference images (IrefI_{\mathrm{ref}}) as conditioning inputs. The selected dataset videos remain the supervision targets throughout this process. Thus, artifacts introduced during condition generation are not directly incorporated into the target content used to train the editor. We further apply multi-stage filtering to remove low-quality pairs and flawed reference candidates, as described in Appendix A.1.

To examine the retained synthetic inputs, we additionally conduct a detector-based data assessment using Forensic-Chat, following the detector-based realism assessment adopted in RealGen [2]. We compute a fake score from normalized real/fake token logits, where lower scores indicate weaker evidence of synthesis. Each image receives one score; for videos, we average scores over sampled frames within each video and then equally across videos. The results are presented below:

ModalityDataset / real-data score ↓Synthetic-condition score ↓Difference Δ\Delta
Videos(Vedit)=0.1242s(V_{\mathrm{edit}})=0.1242s(Vsrc)=0.1349s(V_{\mathrm{src}})=0.1349+0.0107
Images(Ireal)=0.1524s(I_{\mathrm{real}})=0.1524s(Iref)=0.1639s(I_{\mathrm{ref}})=0.1639+0.0115

Here, s(⋅)s(\cdot) denotes the mean Forensic-Chat fake score. The video comparison set comprises target videos sampled from OpenVid-1M, and IrealI_{\mathrm{real}} denotes the real-image comparison set. Δ\Delta is the synthetic-condition score minus its corresponding comparison score.

The results show that the retained synthetic source videos and reference images have mean detector scores numerically close to those of their respective comparison sets. This provides complementary evidence regarding detectable synthetic traces in the retained conditioning inputs.

Finally, we evaluate downstream generalization on external public benchmarks. As reported in Tables 1 and 2, ReBind-Edit outperforms the evaluated open-source baselines on both UniVBench and IntelligentVBench, which were constructed independently of our training pipeline. These results provide evidence that ReBind-Edit generalizes to external evaluation data.

Together, dataset-video supervision, detector scores, and external benchmark results address concerns about synthetic artifacts affecting the editor. We will include this analysis and its protocol in Appendix A.

[2] Junyan Ye et al. RealGen: Photorealistic Text-to-Image Generation via Detector-Guided Rewards. arXiv:2512.00473, 2025.

AuthorsOpenReview ↗

Rebuttal for Reviewer wWsJ (2/2)

Q1. Contribution beyond instruction enrichment and generic reward optimization

Thank you for encouraging us to clarify this distinction!

ReBind supports concise, reusable source references and explicitly optimizes their attribution correctness. This is particularly useful in multi-reference video editing, where the same visual source may supply attributes or preservation constraints across several instruction clauses.

Recent work such as Images in Sentences [3] establishes local binding by inserting visual features directly at corresponding positions within an instruction. Retaining this direct visual binding at multiple mentions of the same source would require repeating the visual-feature blocks at those positions. ReBind instead provides visual inputs as shared conditions and refers to them through short textual identifiers. Repeated source references therefore require only repeated identifiers, without duplicating the visual conditioning input for each mention. This separates reference reuse from visual-feature insertion and provides a compact way to express multiple requirements tied to the same source.

Crucially, reusable identifiers must also be assigned correctly. Table 4 tests this requirement by preserving the descriptive text and reference-tag format while corrupting selected image-reference assignments. With the visual training data and editor training configuration held fixed, IQ remains largely unchanged, but Editing Overall decreases by 0.27 points, with declines across all eight dimensions. These results show that detailed descriptions and explicit identifiers alone are insufficient when the correspondence between content and sources is incorrect.

Our referential consistency reward directly optimizes this correspondence. It evaluates assignments linking entities, descriptions, and sources, so a correct description attributed to the wrong reference does not count as a correct relation. Table 5 shows that adding this reward improves RA by 5.92 points despite lower IQ than optimization with the content reward alone. The additional matched editor training ablation further demonstrates a downstream improvement of 0.30 points from adding the relation reward.

Our contribution therefore combines reusable source references with targeted attribution supervision, supporting repeated reference use while addressing source assignment errors that descriptive enrichment alone does not resolve. We will clarify this distinction and include the comparison with visual-token interleaving in the revised manuscript.

[3] Zhang et al. Images in Sentences: Scaling Interleaved Instructions for Unified Visual Generation, 2026.

Q2. Open-source release plan

Thank you for this valuable suggestion! We will publicly release both model checkpoints, relevant code and materials, together with the corresponding configurations and documentation, to ensure reproducibility.

Reviewer_wWsJOpenReview ↗

Rebuttal Acknowledgement by Reviewer wWsJ

I appreciate the authors' efforts in preparing the rebuttal, which addresses several of my previous concerns. However, I still have several concerns regarding the limitations of the proposed method that remain insufficiently addressed.

  1. Limited evaluation of complex multi-reference scenarios. I think the proposed source-attribution mechanism is particularly promising for mitigating cross-reference confusion in multi-entity scenarios. However, although the paper claims that the training dataset contains samples with up to five reference images, the qualitative examples mainly demonstrate cases involving no more than three references. It remains unclear whether the method can reliably scale to scenarios with more reference images and a larger number of interacting entities, where cross-reference confusion and attribute attribution become substantially more challenging.

  2. Limited modeling of temporally evolving entity relationships. The current entity- and attribute-level source assignments do not explicitly account for temporal changes in entity relationships. For example, consider a video in which a man initially holds a cup and then passes it to a woman. Although the cup's visual source remains unchanged, its interaction and ownership relationships evolve over time. It is unclear how such dynamic relationships are represented and supervised within the current framework, and whether the model can consistently tackle these temporal transitions in content-source relationships. I consider this a limitation of the current model.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.