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 and 5 others
The Chinese University of Hong Kong · Zhejiang University · ByteDance · The Ohio State University
Published September 29, 2026 · on the feed October 1, 2026 · ▲ 30 · arXiv 2609.38541
The 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.
The Brief
2:32A short podcast summary of the paper. Written and voiced by AI.
Transcript
ThinkV2V. Unleashing the Reasoning Capability of MLLMs for Instruction-Guided Video Editing. By Donghao Zhou and twelve others, from the Chinese University of Hong Kong, Zhejiang University, and two other institutions. Published on the twenty-ninth of September. Its code repository is worked out: about two hundred source files, with a readme and a licence.
Editing a video from a written instruction has come a long way. But the multimodal large language models inside these systems are mostly used as semantic encoders, and that leaves them short on implicit edits, where carrying out the instruction requires reasoning about causes or meaning before anything can be drawn. The authors call this a fundamental gap in video editing, and it is the gap this work sets out to bridge.
ThinkV2V is a reasoning-driven framework for complex instruction-guided video editing, and its central move is to make the language model think explicitly before any visual generation happens. Thinking first, then editing.
The architecture is a practical one, running from a multimodal language model into a DiT. The model thinks explicitly over the source video and the instruction, and that thinking is turned into refined conditioning signals for the editor. Around the architecture sits a training and inference recipe with two halves. Progressive Curriculum Training gradually moves the model from basic editing toward cases that are heavy on reasoning. Inference-Time Thinking Scaling iteratively refines candidate prompts and picks the most reliable one. Both aim at drawing out reasoning in the hardest editing scenarios.
To train and to measure this kind of editing, the authors also curate a dataset named ThinkV2V-150K, and introduce ThinkV2V-Bench, aimed at edits with implicit intent and causal reasoning.
Experiments show state-of-the-art performance for ThinkV2V, in complex editing scenarios and in standard ones alike. Its DiT is at the five billion parameter scale, and it substantially outperforms baselines at the larger ten billion scale. That pairing is the result to notice. The comparison is not between models of equal size. It is the smaller generator coming out ahead of larger ones, on the complex edits the framework was built for and on ordinary ones as well.
The design choice running through ThinkV2V is about the role of the language model. Used as an encoder, it passes the instruction along. Used as a thinker, it reasons over the video and the instruction first, and the conditioning the editor receives is refined by that reasoning. The curriculum and the test-time refinement exist to coax that thinking out, both while the model learns and while it edits.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Some edit requests never name what they want changed. That is the wall here.
You build a system that edits a video from a written request. Direct requests work fine. But users often write things like 'remove the container used for carrying the food for this outdoor meal'.
The road ahead
- 1 · The problem ~7 min Editors handle direct requests well, but fail when a request describes its target indirectly, because their language model only encodes the words.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Let the multimodal language model think explicitly about the video and request first, then turn that thinking into conditions that guide the video generator.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min ThinkV2V runs in three steps: reason over video and request, turn that reasoning into conditions, then let the DiT generate the edit.
- 6 · The experiment ~6 min You’ll guess: How did ThinkV2V's 5B-scale DiT model compare with larger 10B-scale baselines?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.