Research map / Video / Video-language understanding
V-Stylist: Video Stylization via Collaboration and Reflection of MLLM Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces V-Stylist, a multi-agent MLLM/LLM system (video parser, style parser with tree-of-thought model search, style artist with multi-round self-reflection on ControlNet weights) for open-query video stylization, plus the TVSBench benchmark.
Problem
Video stylization from open-ended user style queries struggles with complex multi-shot transitions, vague style preferences, and fixed ControlNet detail control.
Approach
Mimic video professionals with a collaborative pipeline of MLLM/LLM agents: shot-level prompting via shot detection and captioning, tree-of-thought search over a style model tree (from community checkpoints) instead of training, and an MLLM scorer that iteratively reassigns ControlNet weights.
Key results
- V-Stylist achieves overall average 0.6011 on TVSBench, versus 0.5405 for FRESCO and 0.5560 for ControlVideo.
- Paper reports V-Stylist surpasses FRESCO and ControlVideo by 6.05% and 4.51% in overall average metrics.
- Ablation on TVSBench-highlight shows adding the Style Artist improves video quality and style alignment, with the full system reaching the best scores across all metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Video Parser for shot-level captions and prompts (ablation) | V-Stylist · TVSBench-highlight | CLIP-T: 0.2556 -> 0.2627 | Helped |
| add the Style Parser for style-model selection (ablation) | V-Stylist with Video Parser · TVSBench-highlight | CLIP-T: 0.2627 -> 0.2655 | Helped |
| add Style Artist with multi-round control reflection (ablation) | V-Stylist with Video Parser and Style Parser · TVSBench-highlight | CLIP-T: 0.2655 -> 0.2662 | Helped |
Surprising findings
- Adding the Video Parser slightly reduces style alignment because style words make up a smaller share of the prompt.
Limitations
- Authors plan to optimize system efficiency, implying high inference cost from multi-round rendering and agent calls.
- Evaluation relies on a single self-built benchmark of 50 videos and 17 styles.
- Style Artist uses an MLLM (Qwen2-VL) as scorer, so reflection quality depends on the MLLM.
Experimental setup
Datasets: TVSBench
Benchmarks: TVSBench TVSBench-highlight
Metrics: CLIP-T CLIP-W Structure Semantics Aesthetic-I Aesthetic-V Distortion-I Distortion-V Overall average
Baselines: ControlVideo Control-A-Video FLATTEN Rerender-A-Video FRESCO
Builds on: AnimateDiff ControlNet Diffutoon DiffusionGPT AutoShot
Closest papers by meaning
- Generative Human Motion Stylization in Latent Space (ICLR 2024 · similarity 0.81)
- Local Consistency Guidance: Personalized Stylization Method of Face Video (Student Abstract) (AAAI 2024 · similarity 0.80)
- AgenticVBench: Can AI Agents Complete Real-World Video Production Tasks? (NeurIPS 2026 · similarity 0.79)
- StyleTailor: Towards Personalized Fashion Styling via Hierarchical Negative Feedback (AAAI 2026 · similarity 0.79)
- VRAgent-R1: Boosting Video Recommendation with MLLM-based Agents via Reinforcement Learning (AAAI 2026 · similarity 0.79)
- EvoGround: Self-Evolving Video Agents for Video Temporal Grounding (NeurIPS 2026 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.