LayeredFashion: Recovering Complete Garment Layers through Layerwise Image Decomposition

A clothed person is a stack of garments, yet human images are usually processed as flat pixels. We present LayeredFashion, which decomposes a clothed-person image and outfit description into an editable PSD-style stack of a variable number of named RGBA garments over a person underlay. We reverse dressing: Extraction recovers each garment, and Restoration reconstructs the clothing or body beneath it for the next step. To correct missing regions, false positives, and boundary errors, we introduce the RGB-Semantic Alpha Refiner (RSAR), combining current-image RGB, garment semantics, and whole-image context to refine alpha while preserving RGB. We train extraction and restoration LoRAs on reference and model-generated intermediate states to accommodate inherited errors. For stepwise supervision, we pair over 22K outfits with nearly 100K complete garment layers, together with person underlays and composition-consistent intermediate images. We achieve the best mean among compared general-purpose decomposition methods on all twelve garment, underlay, and recomposition metrics. The editable layers also support text-to-PSD composition and provide a basis for downstream layered 3D reconstruction.