Vision2Web
Multimodal design-to-code tasks combining text, screenshots, interaction videos and reusable assets to reconstruct working interfaces.
Overview
Vision2Web trains and evaluates models on understanding multimodal design inputs and turning them into runnable websites. This showcase focuses on its dynamic-interaction subset: each task combines text requirements, page screenshots, interaction-reference videos and reusable assets to specify layouts, key states and continuous behaviors such as scrolling, hovering, dragging and switching views.
Tasks span commerce, interest communities, client project workspaces, art showcases and city exploration. Input materials, implementation code, runtime and evaluation environments, and production trajectories preserve the process from design understanding to interaction implementation. Visual and functional evaluations assess reconstruction fidelity and actual behavior separately.
Built for
- Multimodal coding-agent post-training
- Reconstructing continuous interactions and states
- Joint visual and functional evaluation
Featured samples
Loading preview…
ORBIT · Fashion
& independent designers
Interactive · Web · Loading
Switch between spatial displays and product grids, then explore designers, editorial stories and shopping-bag flows.
Build the dataset your model is missing.
Every product here is built with domain experts and delivered to your training or evaluation spec. Tell us what you need.
Discuss your project