MSEditor: Toward Consistent Multi-Shot Video Editing

Kunyu Feng; Yue Ma; Bingyuan Wang; Yuefeng Wang; Zhiyuan Qin; Hao Cheng; Hao Li; Qifeng Chen

cs.CVAug 18, 2026

MSEditor: Toward Consistent Multi-Shot Video Editing

Authors: Kunyu Feng, Yue Ma, Bingyuan Wang, Yuefeng Wang, Zhiyuan Qin, Hao Cheng, Hao Li, Qifeng Chen,

Organizations: HKUST (Guangzhou) · HKUST · Baidu Inc. · Beijing Innovation Center of Humanoid Robotics ·

Abstract

In this paper, we tackle the problem of performing consistent, unified modifications to a multi-shot video sequence. This task is particularly challenging because multi-shot videos consist of discontinuous temporal segments that vary significantly in viewpoint, camera scale, and subject pose, leading to severe identity drift and cumulative error propagation. Achieving coherent edits requires establishing reliable cross-shot semantic awareness to maintain stable subject appearance and visual continuity across these disjointed boundaries. To address this, we propose MSEditor, the first framework designed specifically for consistent multi-shot video editing. To overcome the scarcity of high-quality multi-shot training data, we repurpose existing multi-view video datasets to provide robust cross-shot supervision. Architecturally, we introduce a Supervisory Adapter that injects this cross-shot information into the diffusion backbone, enabling the model to learn identity-consistent representations. Furthermore, to effectively mitigate cumulative errors and ensure long-range temporal coherence, we design a Cross-Shot Packing strategy that dynamically aggregates information from semantically related shots within the self-attention window. Extensive experiments demonstrate that MSEditor significantly outperforms existing methods on our curated multi-shot video editing benchmark in terms of identity preservation, temporal stability, and overall visual quality.

Figures & tables

Figure 1: Showcase of MSEditor. Given a multi-shot input video, our method produces consistent edits across shots with diverse scene compositions, maintaining appearance, motion, and cross-shot consistency.

Figure 2: Comparison with the previous single-shot editing framework . (Top) Existing single-shot editing baselines partition multi-shot sequences into independent temporal chunks. This fragmented inference leads to cumulative errors and cross-shot inconsistency, while being computationally time-consuming due to multiple inference passes. (Bottom) Our MSEditor performs a single inference pass, achieving high-fidelity consistency and superior efficiency.

Figure 3: Overview of our method. Top: Given an input multi-shot video and corresponding editing masks, our method first trains a Supervisory Adapter to inject multi-shot information into the diffusion backbone. We also introduce two key strategies: Cross-Shot Packing and Sparse Cross-Attention to keep the consistency during training. Bottom: During inference, editing is performed on the first frame and produces coherent and identity-preserving multi-shot video output.

Figure 4: Comparison of multi-shot mask extraction pipeline for training data construction. (a) Using segmentation models (e.g., SAM2 [ Ravi et al., 2025 ] , Grounded-SAM [ Ren et al., 2024 ] ) naively on the per-shot video often results in inconsistent results. The model may successfully segment the subject in one shot but fail in another due to a complex background. (b) Our proposed method bypasses this limitation by leveraging the multi-view data’s camera and depth information. We use a complete mask from one shot and re-project it into another’s view, rendering an accurate mask. This process ensures high-quality, cross-shot, consistent annotations.

Figure 5: Overview of the Cross-Shot Packing Strategy and Sparse Cross Attention Module. (a) Given multi-shot inputs, the Cross-Shot Packing Strategy addresses the challenge of maintaining consistency in multi-shot video editing. (b) The Sparse Cross Attention Module integrates textual guidance. Visual queries from each shot (e.g., $Q_{\text{shot-1}}$ ) sparsely attend only to the concatenation of the shared global text embeddings ( $K_{\text{global}},V_{\text{global}}$ ) and their corresponding shot-specific embeddings ( $K_{\text{shot-1}},V_{\text{shot-1}}$ ).

Figure 6: Gallery of our proposed method, which displays four-shot and two-shot editing examples . These results demonstrate the effectiveness of our model in applying diverse semantic modifications while maintaining high visual fidelity and cross-shot consistency.

Figure 7: Qualitative comparison results with the state-of-the-art methods. We compare our multi-shot approach against leading single-shot video editing baselines on a four-shot editing task. The results show that our method successfully executes the complex subject replacement, demonstrating superior cross-shot temporal consistency.

Method	Video Quality	Inter-shot Consistency		Intra-shot Consistency		Semantic Consistency
Method	Aesthetic Score $\uparrow$	Subject $\uparrow$	Background $\uparrow$	Subject $\uparrow$	Background $\uparrow$	Global $\uparrow$	Shot $\uparrow$
TokenFlow [ Geyer et al., 2024 ]	5.51	0.9181	0.9279	0.9029	0.9173	0.1302	0.1415
InsV2V [ Cheng et al., 2024 ]	4.89	0.9013	0.9307	0.8962	0.8941	0.1433	0.1137
VACE [ Jiang et al., 2025 ]	5.91	0.9237	0.9372	0.8805	0.9365	0.1649	0.1738
Ditto [ Bai et al., 2026 ]	5.73	0.9287	0.9341	0.9125	0.9331	0.1514	0.1569
Ours	6.05	0.9370	0.9447	0.9501	0.9462	0.1912	0.1886

Table 1: Quantitative Results of Comparative Studies. Red and Blue denote the best and second best results, respectively.

Method	Video Quality	Inter-shot Consistency		Intra-shot Consistency		Semantic Consistency
Method	Aesthetic Score $\uparrow$	Subject $\uparrow$	Background $\uparrow$	Subject $\uparrow$	Background $\uparrow$	Global $\uparrow$	Shot $\uparrow$
VACE [ Jiang et al., 2025 ]	5.93	0.9301	0.9395	0.9057	0.9382	0.1728	0.1753
VideoPainter [ Bian et al., 2025 ]	6.01	0.9237	0.9396	0.9061	0.9377	0.1687	0.1732
Ours	6.05	0.9370	0.9447	0.9501	0.9462	0.1912	0.1886

Table 2: Quantitative Results of Comparative Studies. For fair comparison, we train recent methods VACE and VideoPainter on our dataset. Red and Blue denote the best and second best results, respectively.

Figure 8: More Qualitative Results after Training VACE and VideoPainter on Our Dataset. For fair comparison, we train VACE and VideoPainter on our dataset, and compare the performance on the two-shot editing task. The results demonstrate that our proposed method has superior performance.

Method	Video Quality	Inter-shot Consistency		Intra-shot Consistency		Semantic Consistency
Method	Aesthetic Score $\uparrow$	Subject $\uparrow$	Background $\uparrow$	Subject $\uparrow$	Background $\uparrow$	Global $\uparrow$	Shot $\uparrow$
w/o First Frame editing	6.01	0.9321	0.9357	0.9336	0.9305	0.1752	0.1729
w/o Cross-Shot Packing strategy	5.83	0.8919	0.9132	0.8937	0.9008	0.1738	0.1521
w/o Sparse Cross attention	5.70	0.8931	0.9256	0.8651	0.9126	0.1691	0.1432
Ours	6.05	0.9370	0.9447	0.9501	0.9462	0.1912	0.1886

Table 3: Quantitative Results of the Ablation Study. Red and Blue denote the best and second best results, respectively.

Figure 9: Qualitative Results of the Ablation Study. We ablate the proposed modules in two-shot editing tasks. Although the first shot is successfully edited, (1) Without the Cross-Shot Packing Strategy, the result lacks inter-shot consistency, leading to visual artifacts (e.g., a distorted face in the close-up shot). (2) Without the Sparse Cross Attention, the model fails in the second shot, indicating a failure to process the per-shot semantic guidance. (3) Without the First Frame Editing, the model successfully transforms the target but exhibits subtle deviations in fine-grained details, highlighting its role as a high-fidelity visual anchor rather than a prerequisite for structural consistency.

Video Editing

Video editing uses AI to alter, compose, relight, or reorganize moving footage while controlling how changes persist across frames.

See all 178 papers

Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning
Aug 27, 2026Chenyang Wu, Fuchen Long, Binyuan Huang +4Agentic Video Editing Video Editing
VCIP, CS, Nankai University · Smart Creation Platform Department, Online Video BU, Tencent
IDAG-Edit: Multi-Object Video Editing via Instance-Decoupled Attention and Guidance
Jun 20, 2026Yuan-Zhih Lin, Huu-Thang Nguyen, Huu-Phu Do +2Diffusion Model Guidance Video Editing
Department of Computer Science, National Yang Ming Chiao Tung University, Taiwan
Beyond Consistency: Preserving Temporal Structure in Zero-Shot Video Editing
Jun 7, 2026Deyin Liu, Yisheng Ding, Zhe Jin +3Video Diffusion Models Video Editing
Anhui University, No 111, Jiulong Road, Hefei, 230601, Anhui Province, China · Surrey Institute for People-Centred Artificial Intelligence, University of Surrey, Guildford, GU2 7XH, Surrey, UK · University of Warwick, Gibbet Hill Road, Coventry, CV4 7AL, UK

Aug 27, 2026J/K move · Enter open · S save

Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning

Chenyang Wu, Fuchen Long, Binyuan Huang, Xinlong Sun+3

VCIP, CS, Nankai University · Smart Creation Platform Department, Online Video BU, Tencent

Agentic Video Editing Video Editing Instruction-Based Video Editing Video Editing Benchmarks

Open paper AskPDF arXiv

While generative AI has significantly advanced video editing, existing methods primarily focus on single-shot or short video clips. Editing long videos with multiple instructions remains a formidable challenge. Naive chunking strategies, e.g., fixed-duration segmentation, often lead to entity fragmentation, severe editing hallucinations, and disrupted temporal continuity. To bridge this gap, we introduce the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task, which is structured around three core objectives: Cross-Shot Editing Consistency (CSEC), Multi-Instruction Decoupling (MID), and Zero-Destruction on Spatiotemporal Structure (ZDSS). To tackle these three unique challenges, we introduce an agentic editing framework that leverages the synergy of Large Language Models (LLMs) and Vision-Language Models (VLMs) to achieve shot-level video decoupling and precise instruction parsing. Furthermore, to comprehensively evaluate this task, we construct MMLVE-Bench, which is an MMLVE-focused dataset characterized by complex real-world spatiotemporal dynamics, high-density heterogeneous instructions, and sparse, random entity distributions. Three MMLVE-focused evaluation metrics are further exploited to assess the quality of the editing results. Extensive experiments demonstrate that our MMLVE-Agent outperforms existing closed-source SOTA approaches (e.g., Seedance 2.0), successfully eliminating editing hallucinations, preserving cross-shot editing consistency, and attaining seamless spatiotemporal transitions.

Abstract

Video Editing

Explore similar work

Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning

IDAG-Edit: Multi-Object Video Editing via Instance-Decoupled Attention and Guidance

Beyond Consistency: Preserving Temporal Structure in Zero-Shot Video Editing