Video2GUI: Synthesizing Large-Scale Interaction Trajectories for Generalized GUI Agent Pretraining
Abstract
A large-scale GUI dataset was created by automatically extracting interaction trajectories from internet videos, enabling improved performance in GUI agents through pre-training on this diverse collection.
Recent advances in multimodal large language models have driven growing interest in graphical user interface (GUI) agents, yet their generalization remains constrained by the scarcity of large-scale training data spanning diverse real-world applications. Existing datasets rely heavily on costly manual annotations and are typically confined to narrow domains. To address this challenge, we propose Video2GUI, a fully automated framework that extracts grounded GUI interaction trajectories directly from unlabeled Internet videos. Video2GUI employs a coarse-to-fine filtering strategy to identify high-quality GUI tutorial videos and convert them into structured agent trajectories. Applying this pipeline to 500 million video metadata entries, we construct WildGUI, a large-scale dataset containing 12 million interaction trajectories spanning over 1,500 applications and websites. Pre-training Qwen2.5-VL and Mimo-VL on WildGUI yields consistent improvements of 5-20% across multiple GUI grounding and action benchmarks, matching or surpassing state-of-the-art performance. We will release both the WildGUI dataset and the Video2GUI pipeline to support future research of GUI agents.
Community
🚀 Excited to share Video2GUI — a fully automated framework that turns unlabeled YouTube videos into grounded GUI interaction trajectories at internet scale.
Existing GUI agent datasets rely on costly manual annotation and are limited to narrow domains, which bottlenecks generalization. We ask: can we instead mine the massive supply of GUI tutorials already on the web? Starting from 500M+ YouTube videos, our coarse-to-fine filtering + VLM-driven trajectory extraction + spatial grounding pipeline produces WildGUI: 12.7M trajectories, 124.5M screenshots, 1,500+ apps and websites across web, mobile, and desktop — the largest open-source GUI pretraining dataset to date.
Pretraining Qwen2.5-VL and Mimo-VL on WildGUI yields 5–20% gains across ScreenSpot-Pro, OSWorld-G, AndroidControl, CAGUI, OSWorld, and AndroidWorld — matching or surpassing SOTA. On ScreenSpot-Pro, accuracy improves from 41.2 → 56.9, a nearly 38% relative gain. Happy to discuss the pipeline design and how internet-scale video mining can power the next generation of GUI agents. Dataset and pipeline will be released!
the trajectory extraction that turns gui tutorials into grounded instruction-trajectory pairs is a neat idea. my main worry is grounding robustness when ui elements drift across app versions or dynamic layouts, because a single misgrounding can cascade in long-horizon plans. it would be great to see an ablation that isolates versioning or accounts for layout changes, to quantify how brittle the grounding actually is. btw the arxivlens breakdown helped me parse the method details, there's a nice walkthrough that aligns with their two-stage pretraining, see https://arxivlens.com/PaperView/Details/video2gui-synthesizing-large-scale-interaction-trajectories-for-generalized-gui-agent-pretraining-2682-0bbf37b1. overall i still think this is a promising direction for scaling gui agents, as long as the ground-truth grounding holds up under real-world ui drift.
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper