Creating a realistic, animatable 3D avatar from a single photograph is a long-standing problem in computer vision and graphics. The new arXiv paper "OMEGA-Avatar" tackles this by proposing a feed-forward method that outputs a 360-degree Gaussian avatar directly from one image, without per-subject optimization.
The authors identify three desirable attributes for avatar generation: the method should be feed-forward, it should model a full 360-degree view, and—though the abstract excerpt cuts off before the third attribute is stated—the work centers on a Gaussian-based representation for rendering. The approach is positioned as a step toward practical one-shot avatar creation.
Because only the abstract's opening lines are available in the source, details on the third attribute, training data, and evaluation are not covered here. The paper's emphasis on feed-forward inference and 360-degree coverage suggests a focus on speed and completeness compared to methods that require multiple views or optimization.