A new research system called OmniScribble, presented by Hien Trinh, Tan-Hiep To, and Trung-Nghia Le, tackles a problem familiar to anyone who’s used an AI tool that “just does it”: when a generative system decides everything at once, it’s hard to understand why it chose what it chose — or to fix just the part that’s wrong.
The core idea: decisions as artifacts, not black boxes
OmniScribble animates multi-object sketches by breaking the process into two explicit, editable stages: object decomposition (figuring out what the separate moving parts of a sketch are) and motion planning (deciding how those parts move). Rather than generating a finished animation end-to-end, the system pauses at verification checkpoints, presenting its proposals for each stage as something the author can accept, repair, or regenerate — not a finished result to take or leave.
Both the manual and AI-assisted paths run through the same intermediate representation and rendering pipeline, which is a deliberate research design choice: it isolates exactly what AI delegation changes, rather than comparing two different tools built different ways.
What the study found
In a within-subject study with eight participants, AI-assisted authoring produced a median 71.4% reduction in completion time compared to manual authoring — without participants reporting a drop in perceived control or how well the result matched their intended motion. Interaction traces and interviews showed people adapting to AI mistakes in three ways: repairing local issues directly, regenerating mismatched proposals, or simply accepting ones that worked.
Why “inspectable and repairable” matters more than “automated”
The paper’s framing is pointed at a broader problem in AI creative tools generally: speed is easy to deliver, but speed that removes the author’s ability to understand or redirect the work tends to erode trust fast. OmniScribble’s contribution isn’t a faster animation engine — it’s evidence that keeping AI decisions as inspectable, editable artifacts can preserve both speed and control, instead of trading one for the other.
The paper is published on arXiv (2611.04284) under Human-Computer Interaction.