CoatFusion: Controllable Material Coating in Images

Material Coating is a novel image editing task that simulates applying a thin material layer to a target surface (substrate) while preserving the substrate's geometry and shading. Material coating is fundamentally different from material transfer, which typically reinterprets an object as being made of a new material and can modify its shape or overwrite fine details. We introduce CoatFusion , a diffusion-based framework for controllable material coating in images. Since real paired before/after coating data is difficult to acquire, we construct DataCoat110K , a synthetic paired dataset of approximately 110K examples rendered with physically motivated coating layers, including opaque, metallic, and transmissive appearances. CoatFusion formulates coating as visual in-context learning over a triptych containing the substrate image, a canonical 2D albedo texture, and the coated result, and uses cycle-consistent compose/decompose supervision to encourage substrate preservation and geometry-aware texture wrapping. Learned material tokens provide continuous control over roughness and thickness, and discrete on/off control over metalness and transmission. Across held-out synthetic evaluations, real-image examples, ablations, and a human study, CoatFusion produces more realistic and controllable coatings than material transfer/editing baselines and classical blending operations. Code and data for this paper are available at https://github.com/sagielevy/coatFusion.