Every image that moves through Wearly Studio now passes through an annotation layer. The guiding principle is to spend as little as possible on the decisions that are cheap, so that we can afford to be careful on the decisions that are not.
The cheap decisions come first. A compact classifier assigns each image a role, such as front, back, detail, on-model, or flat lay, at a negligible cost per image. The same class of models flags which images contain readable brand labels and printed marks, then routes them on for detection and reading. Because these classifiers are small and fast, the entire history of the platform can be re-annotated in hours rather than weeks whenever a model improves.
The careful work is segmentation. We separate the garment from the model wearing it and from the scene around it, at pixel accuracy. Masks are useful in two places. At generation time they let us condition and constrain models on the garment region specifically. At evaluation time they let us measure whether the garment, rather than the lighting or the pose, survived generation intact.
The step that makes the rest valuable is linkage. Every image is grouped to the physical SKU it depicts. This joins the original product photos with every image generated from them. The outcome is a graph that holds one garment, every reference of it, and every generated depiction of it, with a human-verified core layered on top.
None of this is research for its own sake. Role classification powers better defaults in the studio. Label detection protects the most commerce-sensitive region of a garment during generation. Masks improve conditioning today. At the same time, the graph that the layer produces is the training and evaluation substrate for the rest of the research program, which is why we treat the data layer as the first research investment rather than the last product feature.