acceptodds
Under review as a conference paper at ICLR 2027

Reveris: Adaptive Reference Conditioning for Few-Step Visual Effect Transfer

18 views

Abstract

Advances in visual generative models have expanded digital content creation, where creators seek to reuse distinctive effects across subjects and scenes. Reference videos provide concrete demonstrations of transformations that are difficult to specify through text alone, but also entangle the intended effect with reference-specific content. Faithful visual effect transfer therefore requires capturing the relevant change evidence while adapting its influence to the target. We introduce Reveris, a four-stage training framework spanning reference adaptation, modulation calibration, multi-source guidance, and few-step distillation. Within a shared image-to-video backbone, we combine clean-reference temporal variation with caption-conditioned content matching and target-conditioned modulation to adapt reference evidence to new content. We further coordinate target and reference conditions through multi-source classifier-free guidance, and distill this guided behavior through distribution matching into a four-step generator. We also construct a curated dataset of 49,194 videos spanning 350 effect types in 13 families to support learning shared effect behavior across diverse subjects and scenes. Experiments demonstrate strong overall transfer performance, with consistent improvements in effect fidelity, source preservation, and visual quality. Our project demo is available at https://reveris123.github.io/reveris.

Primary area:
generative models
Keywords:
Visual Effect Transfer, Image-to-Video Generation, Video Diffusion Models

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_9xacOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Reveris, an efficient visual effect transfer method based on video diffusion models. It aims to transfer dynamic visual effects from a reference video to a target scene while preserving the target subject and scene structure and preventing leakage of irrelevant reference content. The authors construct a large-scale dataset containing 49,194 target videos and 350 types of visual effects, and design a four-stage training framework. First, adaptive reference conditioning modeling is adopted to extract effect information. Second, modulation calibration is used to control effect strength and spatial coverage. Third, multi-source conditional guidance is incorporated to boost generation quality. Finally, distribution-matching distillation compresses diffusion sampling from 50 steps to only 4 steps. In addition, the authors propose a multi-dimensional evaluation system combining VLMs and evidence-based validation. Experiments show that Reveris outperforms existing methods overall in visual effect transfer quality, subject preservation and reference content isolation, while achieving approximately 9.4× inference speedup.

Critical strengths
  1. The research problem carries clear practical significance. This work focuses on reference-video-driven visual effect transfer. Its goal is to achieve dynamic effect transfer across different subjects and scenes while preserving the original subject and scene structure. Existing methods suffer from issues such as reference content leakage, destruction of subject structure, and low generation efficiency. This paper proposes a relatively systematic solution addressing these challenges.
  2. It presents a complete and targeted four-stage training framework. Reference Adaptation improves the accuracy of reference effect extraction via content matching and temporal variation modeling. Modulation Calibration introduces signed modulation magnitudes and target-conditional gating mechanisms, enabling visual effect transfer to adapt to diverse target scenes. Multi-source Guidance leverages different conditional information comprehensively to enhance the controllability of generation outputs. Few-step Distillation compresses diffusion sampling into only 4 steps, boosting inference efficiency while maintaining generation quality.
  3. A large-scale dataset with rich visual effect types is constructed. The dataset contains 49,194 target videos covering 350 kinds of visual effects across multiple effect categories and diverse scenes. Built through automated generation, filtering and validation pipelines, the dataset provides data support for training reference-conditioned visual effect transfer models. Dataset construction itself constitutes an important contribution of this paper.
  4. The experiments are comprehensive, including extensive comparisons and ablation analyses. Quantitative and qualitative comparisons are conducted against VAP, VFXMaster, ReVFX and their variants. Ablation studies are performed on key components including reference modulation, multi-source guidance and distillation. The appendix further analyzes the impacts of different guidance weights, evaluator selections and reference conditions, helping clarify the function of each module.
  5. The method exhibits notable advantages in generation efficiency. Few-step distillation reduces sampling steps from 50 to 4, achieving approximately 9.4× inference speedup. Compared with conventional methods requiring numerous diffusion steps, this improvement enhances usability for practical video editing tasks.
  6. A relatively comprehensive automated evaluation system is proposed. It considers multiple dimensions beyond visual quality, including effect fidelity, subject preservation, temporal consistency and reference content leakage. An evaluation mechanism combining VLM judgments and evidence-based validation is introduced, attempting to reduce uncertainty introduced by relying solely on subjective VLM scores.
Critical weaknesses
  1. The core method exhibits relatively limited novelty and mainly consists of the combination and optimization of existing techniques. The four-stage training framework adopts reference conditioning adaptation, feature modulation, multi-condition CFG, distribution-matching distillation and other techniques. Most of these underlying ideas have been explored in prior works. Although each module is adapted for visual effect transfer, the paper does not sufficiently elaborate the fundamental methodological innovations of these designs compared with prior arts. In particular, it remains to be clarified whether the main performance gains stem from the novel reference conditioning modeling mechanism, or from the multi-stage training strategy and larger-scale dataset.
  2. The evaluation over-relies on automated VLM assessment and lacks human subjective evaluation. The paper mainly uses GPT-5.6 and other VLMs to evaluate generation outputs. Despite the evidence-based validation mechanism, evaluation results may still be affected by model bias, prompt design and visual understanding capabilities. The naturalness, visual appeal and perceptual realism of video visual effects are highly subjective and difficult to fully measure solely via VLMs. The absence of human evaluation constitutes a notable limitation in experimental validation, especially given the relatively modest quantitative improvements.
  3. The dataset is built primarily via an automated synthesis pipeline; its authenticity and diversity remain to be verified. The authors construct large-scale training data using generative models and automatic filtering pipelines, yet synthesized visual effects may carry shared generative biases. Although there are many effect categories, it is unclear whether the data possesses sufficient diversity in visual styles, motion patterns and physical interactions. The paper lacks systematic generalization experiments on real-captured visual effects, complex physical interactions and out-of-distribution scenarios, making it hard to fully assess its real-world applicability.
  4. There is a lack of systematic analysis of failure cases and the applicability boundary of the method. The paper demonstrates numerous successful examples but provides limited discussion of failure scenarios. For instance, it remains unclear whether the method works stably for visual effects involving complex occlusions, drastic camera motion, multi-subject interactions, precise spatial constraints or sophisticated physical changes. The lack of systematic failure-case analysis hinders a comprehensive understanding of the method’s limitations and directions for future improvements.
  5. Minor formatting issues exist. Several subtle formatting problems are present in the manuscript. For example, the vertical spacing between the captions of some tables and the top borders of the tables is noticeably too small, which deviates from the standard layout specified in the official ICLR template. The authors are advised to further check the formatting throughout the paper to ensure that details such as table captions and spacing comply with the requirements of the official ICLR template.
Questions for authors
  1. Regarding methodological novelty: The four-stage framework of this paper combines multiple existing techniques. Could the authors further clarify the key innovations of Reference Adaptation and Modulation Calibration compared with existing reference conditioning modeling and feature modulation methods? Could the authors further specify which performance gains mainly come from reference conditioning adaptation, and which come from multi-source guidance or distillation? Under a fixed data scale and training computation budget, do the relative contributions of each module still hold?
  2. Regarding dataset quality and generalization ability: Training data is mainly constructed via an automated generation pipeline. How do the authors ensure that the dataset does not inherit systematic biases from the base generative model? For reference visual effects from real-captured videos that exhibit a clear distribution shift from the training data, can Reveris maintain comparable transfer quality?
  3. Regarding evaluation reliability: The current evaluation relies primarily on VLMs. Have the authors considered adding human user studies and analyzing the correlation between human preferences and automatic evaluation metrics? Especially for small performance gaps among different methods, can VLM scores accurately reflect human perceptual experience?
  4. Regarding the applicability boundary of the method: How does Reveris perform on visual effects with complex occlusions, multi-subject interactions, drastic motion, or scenarios requiring precise physical interactions? Can representative failure cases be supplemented, together with analysis of the primary causes of failure?
6 comments
AuthorsOpenReview ↗

Response to Q1, Subquestion 1: Methodological Novelty

Reference Adaptation and Modulation Calibration address different aspects of reference conditioning.

Reference Adaptation establishes reference–target interaction within a shared pretrained I2V backbone. Target and reference tokens interact through selected self-attention blocks while retaining their respective text and image conditions. Compared with VAP’s separate reference expert coupled to a frozen backbone, our design reuses the existing backbone parameters and adapts its interaction self-attention without introducing new parameters in this stage.

Modulation Calibration augments content-based reference matching with an explicit temporal-variation prior computed from clean reference latents. The prior enters target-to-reference attention as a position-dependent bias. Learned signed amplitudes allow different attention heads to favor or downweight stronger reference variation, while gates conditioned on the source image and target prompt set its strength for each target. Reference-caption-conditioned features provide semantic context for matching.

Compared with a scalar conditioning-strength bias, such as that used in OminiControl, our modulation uses the spatial and temporal distribution of observed reference changes and adapts their influence to the target. Stage II calibrates this mechanism by freezing the adapted backbone and training only the added amplitudes and gates, with the attention bias initialized to zero. Reference Adaptation learns the reference–target interactions; Modulation Calibration learns how reference-change evidence should influence those interactions for different targets.

AuthorsOpenReview ↗

Response to Q1, Subquestion 2: Sources of Performance Gains

The following ablation examines where the teacher’s quality gains come from. All four configurations are evaluated before distillation.

Training configurationEF ↑SP ↑TC ↑RD ↑VQ ↑EL ↑
Stage I only2.9273.3633.3434.9723.1673.572
Stage I + II3.0233.4743.2634.9713.1233.650
Stage I + III, without adaptive modulation3.0223.4843.2604.9743.1113.499
Full teacher: Stage I + II + III3.0733.6283.4074.9773.2493.734

Reference Adaptation establishes the reference-conditioned model represented by Stage I. Adding Modulation Calibration in Stage II raises EF from 2.927 to 3.023, SP from 3.363 to 3.474, and EL from 3.572 to 3.650. These improvements in effect fidelity, source preservation, and edit locality are obtained through reference-conditioning training, before multi-source guidance or DMD.

Multi-source guidance training further improves how the transferred effect is balanced with the target content. Compared with Stage I + II, the full teacher increases SP from 3.474 to 3.628, TC from 3.263 to 3.407, and VQ from 3.123 to 3.249, while EF rises from 3.023 to 3.073. The Stage I + III model also improves EF and SP over Stage I, but the complete teacher achieves the highest scores across all six metrics. This shows the benefit of combining calibrated reference conditioning with multi-source guidance, which improves both effect fidelity and output quality.

DMD primarily contributes inference efficiency. The quality improvements above are already present in the 50-step teacher. Distillation reduces inference to four steps and provides a measured 22.8× speedup (Table 2 in the manuscript). It retains comparable effect fidelity, with EF of 3.110 versus the teacher’s 3.073, while further improving SP by 0.072 and VQ by 0.143 (Table 1).

For the continued teacher-training comparisons, we fixed the training data and batch size and matched the total number of updates after the same saved Stage I checkpoint, counting both Stage II and Stage III toward the full teacher’s budget. The Stage I-only row is the starting-point baseline. The full teacher’s advantage over both Stage I + II and Stage I + III therefore persists under this matched data and update budget.

AuthorsOpenReview ↗

Response to Q2: Dataset Quality and Generalization Ability

We evaluated our final four-step model on a manually curated set of real-captured effect references from Kinetics-700-2020, paired with source images from our dataset, without additional fine-tuning. We also evaluated it on 500 cases from the external EffectData dataset (Appendix D).

Evaluation setEF ↑SP ↑TC ↑RD ↑VQ ↑EL ↑
Real-captured references2.9794.0833.6254.9793.3753.291
Our test set3.1103.7003.4684.9823.3923.892
EffectData3.1763.4283.3704.9923.0663.392

With real-captured references, Reveris maintains comparable effect fidelity and visual quality: EF is 2.979 versus 3.110 on our original test set, and VQ is 3.375 versus 3.392. Source preservation and temporal consistency reach 4.083 and 3.625, respectively. Appendix F shows real fire-breathing and welding effects transferred to different target subjects and scenes. The generated videos reproduce the flame or welding arc and sparks while retaining the target subjects and surroundings. These results complement the EffectData evaluation and show successful transfer under a change in reference distribution.

To reduce reliance on a single generator, our training data combine videos from multiple video generators and cover 350 effect types across 13 families, with diverse subjects and scenes. We check both effect realization and visual quality, rejecting absent or incorrect effects, shot changes, severe structural artifacts, and dominant unrelated transformations. Captions describe the effects actually observed in the generated videos, including their onset, temporal progression, and final appearance, improving correspondence between the visual and textual training signals.

Reference–target pairs combine different videos within each effect category, exposing the model to the same effect across different subjects and backgrounds. These construction choices encourage transferable effect representations, while the external and real-reference evaluations assess generalization beyond the training corpus.

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

This paper presents Reveris, a four-stage framework for transferring visual effects from a reference video to a source image. Its main contribution is adaptive reference modulation, which combines temporal variation in clean reference latents with learned signed amplitudes and target-conditioned gates to bias reference attention. The framework further uses multi-source classifier-free guidance and distribution matching distillation to obtain a four-step generator. The authors also construct a dataset of 49,194 videos covering 350 effect types. Evaluation on 2,004 test cases and 500 additional EffectData cases supports improved transfer quality, while the distilled model achieves a reported 22.8× speedup over its 50-step teacher.

Critical strengths
  1. The proposed adaptive reference modulation offers a degree of novelty and substantially improves computational efficiency.

  2. The manuscript is well written and clearly organized.

Critical weaknesses
  1. Adjacent latent differences can reflect camera motion, ordinary subject movement, and occlusion as well as effect-related changes. The existing ablations support adaptive modulation, and Appendix C.4 provides useful qualitative evidence through trained mean-prior and inference-time perturbation controls. However, these diagnostics do not quantify how consistently the spatiotemporal change map improves over generic reference-strength modulation, particularly when substantial motion is unrelated to the intended effect. Matched quantitative comparisons and controlled interference tests would strengthen the evidence for this central contribution.

  2. The paper specifies ten uniformly spaced interaction blocks but does not evaluate alternative counts or placements, or whether a separate modulation-calibration stage is necessary. This leaves uncertainty about sensitivity to the selected configuration and training schedule.

Questions for authors
  1. It is recommended to add additional experiments on temporal prior: 1) quantitatively compare the full 50-step teacher with the trained “Mean prior + learned gate” model under identical sampling and guidance settings; 2) report performance on cases with substantial camera or subject motion unrelated to the intended effect.

  2. Please clarify the architecture and training details. For example, what motivated selecting blocks 0, 4, …, 36 for reference interaction? Is there evidence that performance is robust to alternative block selections, and that a separate modulation-calibration stage offers benefits over joint training with a matched update budget?

  3. Appendix C.6 reports that the opposite-flight reference pair produces leftward outputs even with a shared, direction-neutral caption. How should this result be interpreted relative to the paired-reference control in Figure 5? Could you quantify this behavior across more examples and clarify under what conditions the visual reference controls attributes not specified by the text?

6 comments
AuthorsOpenReview ↗

Response to Q1, Subquestion 1: Comparison with the Trained “Mean Prior + Learned Gate” Model

We trained the requested “Mean prior + learned gate” baseline and compared it with our full 50-step teacher under identical sampling and guidance settings. In this baseline, we average the token-wise temporal prior within each reference video and broadcast the mean to all reference tokens, retaining the learnable signed amplitudes and input-conditioned gates. The mean prior is used during both training and inference. Both models use 50 sampling steps and multi-source guidance scales of (4, 5, 2).

MethodEF ↑SP ↑TC ↑RD ↑VQ ↑EL ↑
Mean prior + learned gate3.0273.4933.2834.9733.1223.508
Ours (full 50-step teacher)3.0733.6283.4074.9773.2493.734

Our full teacher scores higher on all six metrics, improving source preservation by 0.135, temporal consistency by 0.124, visual quality by 0.127, and edit locality by 0.226, alongside an effect-fidelity gain of 0.046.

These results support retaining token-level temporal-change information alongside learned gating. The gate adjusts the strength of reference modulation, while the token-wise prior differentiates reference positions according to their observed changes. Averaging the prior removes this differentiation.

AuthorsOpenReview ↗

Response to Q1, Subquestion 2: Unrelated Camera or Subject Motion

We have added Appendix E, “Effect Transfer with Incidental Reference Motion,” with five test-set examples generated by our four-step model. They cover subject movement and camera reframing beyond the intended effect: walking during ribbon wrapping and stone transformation, approaching the camera during an outfit change, a camera push-in with pose changes, and downward reframing during falling petals.

In the ribbon-wrapping example, the reference man enters from the left and continues walking while ribbons wrap around his body. The target woman remains at her original position as colorful ribbons form and interlace around her upper body, preserving her appearance and indoor setting. The output transfers the effect without reproducing the reference subject’s entrance or lateral displacement.

The camera-motion examples show the demonstrated stylization appearing on the target girl and falling petals surrounding the target statue, with their respective scenes retained. These examples provide qualitative evidence of effect transfer in the presence of substantial incidental reference motion.

AuthorsOpenReview ↗

Response to Q2, Subquestion 1: Architecture and Training Details

We chose blocks 0, 4, …, 36 to distribute reference interaction across the 40-block backbone. This allows target features to incorporate reference information at different network depths while limiting reference-token processing to ten blocks. Reference and target tokens undergo joint self-attention in these blocks; only target tokens are updated in the remaining blocks.

We also trained a model with reference interaction in the first ten blocks, 0–9, retaining the token-wise temporal prior and learned gates. Both configurations contain ten interaction blocks and use identical 50-step sampling and guidance settings.

ConfigurationEF ↑SP ↑TC ↑RD ↑VQ ↑EL ↑
First ten blocks (0–9)2.8863.5533.2994.9823.1573.532
Our full teacher (0, 4, …, 36)3.0733.6283.4074.9773.2493.734

The first-ten-block model achieves EF of 2.886 and SP of 3.553. Our full teacher scores higher on five metrics, particularly effect fidelity (+0.187), temporal consistency (+0.108), and edit locality (+0.202). Source preservation and visual quality also improve, while reference disentanglement remains nearly unchanged. These results support our default configuration, with reference interactions distributed throughout the backbone.

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

Reveris transfers visual effects from reference videos to new subjects through adaptive reference conditioning and four-step distillation. A temporal-change prior and target-conditioned gates modulate reference attention, while multi-source guidance coordinates textual and visual conditions. The paper also constructs 49,194 videos covering 350 effects. Results support improved transfer fidelity and substantially faster inference.

Critical strengths

1.The paper addresses a useful and challenging problem. Adapting reference effects to new content while preserving the target subject and scene is a meaningful contribution.

2.Combining promising transfer quality with substantially faster generation. Comparisons, ablations, and evaluation on an external dataset provide substantial support for its effectiveness.

Critical weaknesses
  1. The extensive automated evaluation supports the results, but human validation would make the perceptual improvements more convincing. Can a blinded human evaluation on a representative subset?

  2. The flexibility of the fast model needs further clarification. The four-step model learns a fixed balance between target-content guidance and reference-content guidance. Although the teacher model allows this balance to be adjusted, I would also like to know whether users can similarly control effect strength and content preservation in the fast model without retraining. This is only a minor consideration.

Questions for authors

see the weakness

3 comments
AuthorsOpenReview ↗

Response to Q1: Human Evaluation

Following your suggestion, we added a blinded human evaluation covering 50 test cases and 250 outputs from five configurations: Reveris, Reveris Teacher, VFXMaster, RefVFX+CausVid, and VAP. For each case, the source image, reference video, target prompt, and five anonymized outputs were presented together. Five independent raters assigned integer scores on a 0–5 scale for effect fidelity (EF), source preservation (SP), temporal consistency (TC), visual quality (VQ), and edit locality (EL). This yields 1,250 rater–output records and 6,250 individual scores. We excluded reference disentanglement (RD) from the human evaluation because its automatic scores are near the ceiling and provide little separation among methods. RD remains part of the paper’s automatic evaluation.

We first averaged the five ratings for each output, then averaged across the 50 test cases. We report all five dimensions separately and compute their equally weighted composite:

Composite = (EF + SP + TC + VQ + EL) / 5.

Human ratings on a 0–5 scale. Higher is better; bold indicates the highest mean in each column. The composite excludes RD.

MethodEFSPTCVQELComposite
Reveris (4 steps)3.4884.2244.1203.7163.5963.829
Reveris Teacher (50 steps)3.3964.1364.1483.5643.4443.738
VFXMaster3.2163.9123.8603.3363.3283.530
RefVFX+CausVid3.2244.0683.8963.4723.2443.581
VAP2.9283.6403.7363.4282.7403.294

Reveris has the highest mean EF, SP, VQ, EL, and composite scores, while the teacher has a slightly higher mean TC score. We estimate uncertainty using 5,000 paired reference-group bootstrap replicates, keeping all model outputs and ratings together within each of the 50 groups.

The composite differences relative to VFXMaster, RefVFX+CausVid, and VAP are +0.298 [0.138, 0.452], +0.248 [0.112, 0.384], and +0.534 [0.373, 0.690], respectively. These are exploratory 95% percentile intervals without multiple-comparison adjustment.

The corresponding EF differences are +0.272 [−0.012, 0.560], +0.264 [−0.020, 0.548], and +0.560 [0.276, 0.864]. The EF interval excludes zero for VAP and includes zero for VFXMaster and RefVFX+CausVid.

Compared with the teacher, the EF difference is +0.092 [−0.236, 0.424], and the composite difference is +0.091 [−0.052, 0.237]. The student–teacher intervals include zero for all five dimensions, so we report these as differences in mean ratings without claiming statistical superiority or equivalence. The full protocol and per-dimension uncertainty estimates will be included in the revision.

AuthorsOpenReview ↗

Response to Q2: Flexibility

The four-step student has no additional inference-time control parameters for adjusting effect strength and content preservation. DMD incorporates the teacher’s selected guidance balance into the student’s weights (Section 4.4 and Appendix A.3), so the current student cannot provide the same guidance-weight adjustment without retraining.

This reflects a trade-off between generation speed, quality, and inference-time flexibility. Our DMD configuration prioritizes high-quality four-step generation, achieving a measured 22.8× speedup (Table 2 in the manuscript), with comparable effect fidelity and improved source preservation and visual quality relative to the teacher (Table 1). The student operates at the distilled guidance balance, while the teacher retains adjustable guidance for applications requiring finer control.

Reviewer_ZpDJOpenReview ↗

good luck

Thank you for the authors' response. I believe a score of 3 is appropriate for this paper, and I will therefore maintain my rating.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Thanks for all reviewers.

We sincerely thank all reviewers for their valuable feedback and insightful questions. In response, we have conducted additional experiments and analyses and added several sections to the appendix of the revised manuscript to present the corresponding visual results. The titles of the newly added appendix sections are highlighted in red. Our point-by-point responses refer to these sections where relevant, and we invite reviewers to consult the revised appendix for the supporting visualizations and analysis.

AuthorsOpenReview ↗

Best wishes to everyone

We are glad that our responses and additional experiments have addressed all reviewers's main concerns. Your suggestions have helped us clarify the method, strengthen the experimental evidence, and better explain its applicability. We appreciate your positive assessment of our work and the time and care you have devoted to the review process. Thank you again, and best wishes to everyone!

AuthorsOpenReview ↗

Summary of Author Responses and Revisions

Dear AC,

Thank you and the three reviewers for the time you devoted to our submission. The reviewers highlighted the practical value of transferring an effect to a new subject and scene (9xac, ZpDJ), the design of the adaptive reference modulation (kXry), and the efficiency of four-step generation (all three reviewers). Their questions led to the following additions and clarifications.

  • Human evaluation (9xac, ZpDJ). Five independent raters scored 250 anonymized videos from five methods on 50 test cases. On a 0–5 scale, four-step Reveris had the highest mean composite score (3.829), followed by the 50-step teacher (3.738) and RefVFX+CausVid (3.581). The paired 95% bootstrap intervals for the composite gain excluded zero against each external baseline and included zero against the teacher. Human and VLM rankings agreed most closely on effect fidelity (Spearman ρ = 0.714).

  • What adaptive modulation adds (9xac, kXry). At identical 50-step sampling and guidance settings, preserving where and when the reference changed improved edit locality from 3.508 to 3.734 over a trained mean-prior baseline with the same learned gating mechanism. With the training data and update budget held fixed after Stage I, the full teacher also scored highest on all six automatic metrics among the tested variants. These gains were already present before distillation.

  • Transfer to real footage and limits of control (all three reviewers). Appendix F shows transfer from real fire-breathing and welding footage without further fine-tuning. Visual quality was close to the in-domain score (3.375 vs. 3.392), while edit locality was lower (3.291 vs. 3.892). Appendices E, G, and H show when incidental motion is ignored, when an effect attaches to the wrong subject, and when flight direction needs to be specified in the prompt. The four-step student runs 22.8× faster than the teacher at a fixed guidance balance. Users can adjust that balance at inference in the teacher.

We believe our responses, additional experiments, and revisions addressed the concerns raised by all three reviewers. All three maintained their positive ratings after the discussion.

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.