Alterbute: Editing Intrinsic Attributes of Objects in Images

Alterbute is a new diffusion method for intrinsic object editing: it targets properties like color, texture, material, and shape without blowing away the object’s perceived identity. The paper argues that prior methods either preserve identity poorly or are too constrained to produce useful variations, so this work relaxes training while tightening inference with the original background and object mask.

For game developers, the appeal is obvious for concept art, reference generation, and rapid asset exploration. The method also introduces Visual Named Entities (VNEs), such as specific car models, to group visually similar objects while still allowing variation in intrinsic attributes. The authors say they automatically extract VNE labels and attribute descriptions from a large public image dataset using a vision-language model, which is the scalable part that matters if you want identity-preserving supervision without hand-labeling everything. The paper was submitted Jan. 15, 2026...