Generative Tutorial: Towards Live Contextualized Visual Instructions for Physical Tasks
An AR prototype generated task guidance that matched a user’s actual workspace and outperformed pre-authored instructions in a 24-person lab study.
The paper frames “Generative Tutorial” as live visual instruction for physical tasks, using observed workspace context to show intended outcomes and actions in place. The authors first evaluated current image and video generation across 15 tasks, finding both useful affordances and failure modes. Their prototype proactively produced goal images and demonstration videos tied to the user’s task flow. Participants showed higher task performance quality, stronger perceived workspace correspondence, and shorter step-confirmation intervals than with pre-authored guidance. ArXiv · AI/CL/LG's note
The paper frames “Generative Tutorial” as live visual instruction for physical tasks, using observed workspace context to show intended outcomes and actions in place. The authors first evaluated current image and video generation across 15 tasks, finding both useful affordances and failure modes. Their prototype proactively produced goal images and demonstration videos tied to the user’s task flow. Participants showed higher task performance quality, stronger perceived workspace correspondence, and shorter step-confirmation intervals than with pre-authored guidance. ArXiv · AI/CL/LG's note
score 4