acceptodds
Under review as a conference paper at ICLR 2027

Soap2Soap: Long Cinematic Video Remaking via Multi-Agent Collaboration

Abstract

We study series-level cinematic remaking, a long-horizon video-to-video generation problem that localizes full episodes or films via stylization or actor replacement while strictly preserving narrative structure, motion choreography, and character identity across hundreds of shots. Existing video generation and editing pipelines often break down in this regime due to compounding identity drift, background mutation, and semantic erosion under large camera motions and viewpoint changes. We propose Soap2Soap, a multi-agent framework that enforces long-term language–visual consistency through a Dual-Bridge Consistency mechanism: a scene-aware JSON screenplay serving as a persistent semantic backbone, and dynamically allocated visual reference anchors at both scene and shot levels. To suppress drift before video synthesis, we introduce batch keyframe consistency, jointly generating multiple keyframes in a shared latent context via a grid-based formulation. A closed-loop verification agent further audits identity, stability, and alignment to trigger selective regeneration. Experiments on SoapBench demonstrate strong improvements over commercial video generation APIs in long-term consistency and narrative fidelity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Series-Level Video Remaking, Long-Term Visual Consistency, Multi-Agent Video Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DEuNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Soap2Soap looks at cinematic remaking, which mean changing how the characters look or the visual style while keeping the source video's story and shot structure. To do this it put together a structured screenplay, visual references that persists across shots, keyframes generated jointly in grid, and verification loop. The SoapBench benchmark test understanding on 607 shots and remaking on ten movie excerpt. Both the automatic metrics and a blinded human study comes out in favor of the proposed system.

Critical strengths

I find the task useful, and the motivation is clear. If you change the cast or the style of a sequence but still keep its story, you end up with real consistency problem. Keeping the narrative instructions apart from the visual references that persists across shots seem like a practical way to organize it. The results also doesn't rely only on the automated evaluator. Table 1 show big gains in identity consistency, and the randomized, anonymous comparison with 20 participant gives decent evidence that viewers actually prefers these outputs too.

Critical weaknesses

1.The remaking evaluation only use ten excerpts, each 1.5 to 5 minute long, and none of them have more than 42 consecutive shot. The 607-shot number is for understanding part, not for continuous remaking. So the results do back up multi-shot remaking, but they doesn't yet show that the method stay consistent at the scale of an episode or a film, with hundreds of shots. 2.In the main comparison, the generation models changes along with the workflow. The ablations takes out memory allocation or verification, but there's no separate test of grid generation, and no comparison with simpler pipeline that use the same generators. The paper also don't say how many retry were actually used. Because of these gap, it's hard to tell how much of the gain come from the proposed design and how much from stronger models and extra generation attempts.

Questions for authors

1.Is there any evidence that the remaking work beyond the 42-shot sequences reported in paper? If longer sequence weren't evaluated, would the authors consider scaling back the claims so they matches what the experiments actually show? 2.If you keep the same Nano Banana 2 and Veo 3 models, the same references, and same generation budget, how does Soap2Soap does against a simpler pipeline that just condition on the references? It would also help to know how much grid generation contribute on its own, and how many retry are needed on average.

Reviewer_D3gxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Soap2Soap, a training-free multi-agent framework for long-form cinematic video remaking. The framework combines structured JSON screenplays, dynamically allocated visual references, grid-based keyframe generation, and iterative verification to maintain character, scene, and narrative consistency. The authors also introduce SoapBench and evaluate the framework against existing video generation and editing systems.

Critical strengths

The paper addresses cinematic video remaking, which requires preserving character identities, scene appearances, and narrative structure across multiple shots.

The authors construct SoapBench by collecting existing movie clips and screenplay-based annotations to evaluate video understanding and remaking.

Critical weaknesses

Section 3.5 and Appendix B.4 state that the framework uses Veo 3 for image-to-video generation conditioned on multiple reference images, including character references, scene references, and generated keyframes. However, Google's official Veo API documentation(https://ai.google.dev/gemini-api/docs/veo) specifies that the referenceImages functionality is supported by Veo 3.1 rather than Veo 3. This discrepancy is particularly concerning because reference-based conditioning is a central component of the proposed Visual Bridge. The authors should provide the exact model identifiers, API parameters, and generation procedures used in their experiments.

Figure 2 explicitly identifies the Video Understanding Agent as GEMINI-3-PRO-PREVIEW, whereas Section 5.1 states that the framework uses Gemini 3 Flash for video understanding. The authors must clarify which model was actually used to obtain the reported results.

Section 2.2 cites Ortega et al. (2020) in the discussion of streaming and autoregressive video generation. However, the corresponding reference is DMD: A Large-Scale Multi-Modal Driver Monitoring Dataset, which is unrelated to the claimed video generation methodology. Repeated or substantive citation misrepresentations would raise broader research-integrity concerns under the ICLR Code of Ethics.

These issues go beyond minor presentation problems. In particular, the unclear model specifications and unsupported API functionality directly affect whether the core generation pipeline can be reproduced as described.

The paper repeatedly emphasizes movie-scale remaking, consistency across hundreds of shots, and preservation of motion choreography. However, the actual remaking benchmark contains only ten video segments of 1.5–5 minutes, with the longest continuous sequence consisting of 42 shots. The 607 shots reported for the understanding benchmark are collected across ten movies and should not be interpreted as evidence of successful generation over hundreds of consecutive shots. Consequently, the experiments do not adequately support the paper's central claims regarding movie-scale or extremely long-horizon consistency.

Questions for authors

The quantitative evaluation relies heavily on Gemini 3 Flash as a VLM-based judge. However, the same model is also used for video understanding and the Verification Agent. This creates a potential evaluation bias, since the generated outputs are explicitly optimized through feedback from the same model family used for the final assessment. The paper does not provide sufficient independent evaluation to rule out this concern.

Reviewer_Pc8MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Soap2Soap is a multi-agent framework which remakes source videos through several components: a structured screenplay, character and scene references, jointly generated keyframes, and verification-driven regeneration. The framework aims to preserve identity, narrative, and cinematic structure across long sequences.

Critical strengths

I like the practical focus on remaking an existing sequence, which should be a useful application scenario. The qualitative examples show convincing cross-shot appearance consistency, and the identity improvements are shown to be substantial: CLIP-I increases from 0.632 baseline for Kling to 0.842. Meanwhile, I think the memory and verification ablations provide useful evidence, while the 20-participant study complements the evaluation.

Critical weaknesses

First, I found the demonstrated horizon does not support the strongest scale claims. The remaking evaluation contains ten 1.5–5 minute segments, with at most 42 consecutive shots. And the only 607-shot figure concerns understanding rather than generation. Thus, I think these results do not establish stability across hundreds of generated shots or full episodes.

What I concern most is the contribution centered on pure organizing the backbones, and will also lead to high generation budget. The main baselines use different generation systems, while Soap2Soap combines Nano Banana, Veo, and selective regeneration. Also the actual retry counts and costs are not reported. While Table 2 provides same-backbone ablations of memory allocation and verification, but it leaves grid synthesis and the individual bridges untested. Currently I still cannot confirm the proposed pipeline’s effectiveness. I think a simpler reference-conditioned pipeline with comparable generation budgets would help isolate the proposed design’s benefit.

Finally, the evaluation does not establish strict source fidelity. The task definition requires preservation of motion choreography, camera language, and audiovisual coherence, while the metrics primarily assess identity, scene appearance, and semantic plot consistency. I think some important perspectives, like Direct motion and camera correspondence are not measured or without uncertainty estimates.

Questions for authors

Please refer to the Weakness section

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.