Reference-free Human-Object Interaction Editing

InteractEdit is a reference-free human-object interaction editing system built for a problem that sits beyond the usual image-editing buckets. Instead of changing color, style, or swapping an object outright, it tries to transform an existing interaction into a new one while keeping the same subject and object identities intact. That makes it especially relevant to any pipeline that needs controllable scene manipulation rather than one-off generative cleanup.

The core idea is to break a scene into subject, object, and background components so the model can disentangle the interaction from the identities involved. It pairs that with selective inversion and Selective-Rank Adaptation (SeRA), which are used to preserve visual identity from the source image while still leaning on pretrained interaction priors. In practice, that’s an attempt to avoid the common failure mode where an edit works structurally but drifts too far from the original characters or props.

The project also adds IEBench, a benchmark specifically for HOI editing, along with a metric that scores both interaction change and identity preservation together. That matters because these systems can look good in isolation while quietly breaking the thing developers care about most: keeping the right character, prop, or silhouette recognizable after the edit. The paper reports that InteractEdit beats 23 existing methods, positioning it as a strong baseline for future work.

For game developers, the immediate relevance is in tooling and content iteration. Techniques like this could eventually help with concept...