ThinkV2V: Unleashing the Reasoning Capability of MLLMs for Instruction-Guided Video Editing

Donghao Zhou; Haoyang He; Fan Zhang; Hao Yang; Guisheng Liu; Xin Gao; Zhongwei Wan; Xingyuan Bu

cs.CVSep 29, 2026

ThinkV2V: Unleashing the Reasoning Capability of MLLMs for Instruction-Guided Video Editing

Authors: Donghao Zhou, Haoyang He, Fan Zhang, Hao Yang, Guisheng Liu, Xin Gao, Zhongwei Wan, Xingyuan Bu, +5 more

Organizations: The Ohio State University

Abstract

Instruction-guided video editing has made significant progress, yet existing methods use multimodal large language models (MLLMs) primarily as semantic encoders, so they often fall short in working with implicit edits that require causal or semantic reasoning. To bridge this fundamental gap in video editing, we propose ThinkV2V, a reasoning-driven framework for complex instruction-guided video editing, explicitly activating MLLM thinking before visual generation. At its core, ThinkV2V builds on a practical MLLM-to-DiT architecture to turn explicit thinking over the source video and instruction into refined conditioning signals for video editing. Further, we equip it with a dedicated training and inference recipe, combining Progressive Curriculum Training, which gradually cultivates the model from basic editing to reasoning-intensive cases, with Inference-Time Thinking Scaling, which iteratively refines candidate prompts and selects the most reliable one, to better elicit reasoning in challenging editing scenarios. We also curate the ThinkV2V-150K dataset and introduce ThinkV2V-Bench to support training and evaluation of video editing with implicit intent and causal reasoning. Experimental results demonstrate the state-of-the-art performance of ThinkV2V on both complex and standard editing scenarios, in which our 5B-scale DiT model substantially outperforms larger 10B-scale baselines.

Figures & tables

Fig. 1: Overview of ThinkV2V. (a) Documentation on video editing with complex instructions, illustrating the capabilities of ThinkV2V. (b) Conceptual comparison between prior methods that mainly use MLLMs as semantic encoders and our method that activates the MLLM as an explicit thinker. (c) Quantitative results showing the strong overall performance of ThinkV2V on complex and standard editing scenarios. Zoom in for better view.

Fig. 2: Pipeline of ThinkV2V. The ThinkV2V framework consists of three key parts: (a) Model architecture ( Section III-A ) uses an MLLM for explicit thinking, a learnable-query connector for feature alignment, and a multi-condition DiT for final video editing. (b) Training and inference recipe ( Section III-B ) includes Progressive Curriculum Training and Inference-Time Thinking Scaling to better elicit reasoning for complex editing. (c) Dataset and benchmark construction ( Section III-C ) for ThinkV2V-150K and ThinkV2V-Bench aims to support reasoning-oriented training and evaluation.

Fig. 3: Statistics and example of ThinkV2V-150K and ThinkV2V-Bench. (a) Word cloud of complex instructions, illustrating the semantic diversity of our reasoning-oriented editing data. (b) Distribution of source-video frame counts, showing the temporal diversity of the collected videos. (c) A representative benchmark sample, including the source video and the corresponding instructions.

Method	#Param.	Global Style $\uparrow$	Bg. Change $\uparrow$	Local Change $\uparrow$	Local Remove $\uparrow$	Local Add $\uparrow$	Overall $\uparrow$
ThinkV2V-Bench
Evaluation on Seed-1.6-VL
VACE [ 35 ]	14B	1.33	1.17	1.41	1.00	1.05	1.19
OmniVideo [ 8 ]	1.3B	1.92	1.40	1.91	1.96	1.87	1.82
ReCo [ 36 ]	1.3B	2.25	1.62	2.50	3.13	2.39	2.26
InsVIE [ 37 ]	2B	2.08	1.18	1.48	1.04	1.24	1.40

TABLE I: Quantitative comparison. We conduct quantitative evaluation on both ThinkV2V-Bench and OpenVE-Bench [ 33 ] , using two judge models, Seed-1.6-VL [ 34 ] and Gemini-2.5-Pro [ 31 ] . “#Param.” denotes the parameter count of the video editor. The results show that our method achieves competitive overall performance across both benchmarks and all evaluated settings, indicating its consistent effectiveness across diverse editing scenarios.

Fig. 4: Qualitative comparison. We provide qualitative results of our method and several competitive methods on examples from ThinkV2V-Bench across diverse reasoning-oriented editing cases, demonstrating the superior ability of ThinkV2V to produce more intention-consistent edits. Zoom in for better view.

Method	Global Style $\uparrow$	Bg. Change $\uparrow$	Local Change $\uparrow$	Local Remove $\uparrow$	Local Add $\uparrow$	Overall $\uparrow$
Ablation on Model Architecture
w/o Thinking	2.89	2.71	3.05	1.46	2.42	2.51
w/ Thinking + All Features	2.95	2.68	3.09	1.94	2.46	2.63
w/ Thinking + Answer Features	2.93	2.63	3.02	2.34	2.47	2.68
Ablation on Progressive Curriculum Training
Simple Only	2.76	2.66	3.27	1.22	2.63	2.52

TABLE II: Quantitative ablation studies. We conduct ablations based on Seed-1.6-VL [ 34 ] evaluation from three key aspects, including model architecture, Progressive Curriculum Training, and Inference-Time Thinking Scaling. These results further verify the effectiveness of the proposed techniques.

Fig. 5: Qualitative ablation on explicit thinking. Enabling explicit thinking leads to more accurate edits, while disabling it often results in misaligned modifications under reasoning-intensive instructions.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. 6: More qualitative comparison. We provide additional qualitative results of ThinkV2V and other competitive methods on more examples from ThinkV2V-Bench, demonstrating the superior ability of ThinkV2V to produce more intention-consistent edits. Zoom in for better view.

Fig. 7: ThinkV2V on a challenging reasoning-intensive case. A detailed example showing how ThinkV2V handles a complex instruction with implicit target identification, fine-grained scene understanding, and strong consistency requirements. Zoom in for better view.

Fig. 8: ThinkV2V on a challenging reasoning-intensive case. A detailed example showing how ThinkV2V handles a complex instruction with implicit target identification, fine-grained scene understanding, and strong consistency requirements. Zoom in for better view.

Explore similar work

Reasoning to Align: Implicit Reasoning in Diffusion Transformers for Video Editing
May 23, 2026Yan Li, Lin Liu, Xiaopeng Zhang +1Video Editing Diffusion-Based Image Editing
†The Hongkong University of Science and Technology · ‡Huawei Inc.
MiVE: Multiscale Vision-language features for reference-guided video Editing
May 14, 2026Tong Wang, Meng Zou, Chengjing Wu +4Video Editing Vision Encoders
MT Lab, Meitu Inc., Beijing 100083, China · Beijing University of Posts and Telecommunications, Beijing 100876, China · Department of Computer Science and Technology, BNRist, IDG/McGovern Institute for Brain Research, Tsinghua University, Beijing 100084, China
One Editor, Many Edits: A Unified Training-Free Framework for Diverse Video Editing
Sep 3, 2026Adheesh Sunil Juvekar, Onkar Kishor Susladkar, Kiet A. Nguyen +6Video Editing Edit
University of Illinois Urbana-Champaign

May 23, 2026, cs.CVJ/K move · Enter open · S save

Reasoning to Align: Implicit Reasoning in Diffusion Transformers for Video Editing

Yan Li, Lin Liu, Xiaopeng Zhang, Qi Tian

†The Hongkong University of Science and Technology · ‡Huawei Inc.

Video Editing Diffusion-Based Image Editing Diffusion Transformers Cross-Attention Transformer Architectures Reasoning Skills

Open paper AskPDF arXiv

Instruction-based video editing requires transforming a source video according to a natural-language instruction while preserving irrelevant content and remaining temporally coherent. We argue that existing Diffusion Transformer (DiT) editors struggle with this task for two structural reasons. First, conditioning signals are fed undifferentiated into all transformer blocks, forcing a single token stream to encode both global editing intent and fine-grained visual evidence. Second, the cross-attention patterns that govern the edit are supervised only indirectly through pixel-level reconstruction, leaving the model's internal reasoning process under-constrained. To address both limitations, we propose RVEDiT, an implicit Reasoning Video Editing DiT framework built around two complementary components. The first, Granularity-Routed Token Conditioning, introduces learnable editing tokens distilled from a multimodal LLM and routes them to shallow blocks, while reserving native visual and textual tokens for deeper blocks, thereby inducing a coarse-to-fine editing process inside the backbone. The second, Reference-Anchored Attention Alignment, employs a parameter-sharing reference branch during training and maximizes the mutual information between the attention features of the editing and reference branches, regularizing the model's internal reasoning without incurring any additional inference cost. Experiments on standard instruction-based video editing benchmarks show that RVEDiT consistently outperforms state-of-the-art baselines, with particularly strong gains on localized and compositional edits.

Abstract

Explore similar work

Reasoning to Align: Implicit Reasoning in Diffusion Transformers for Video Editing

MiVE: Multiscale Vision-language features for reference-guided video Editing

One Editor, Many Edits: A Unified Training-Free Framework for Diverse Video Editing