A flattened image can look finished while being difficult to edit: changing one object may disturb nearby content, and a separate layer decomposition step does not necessarily preserve the intent of the original edit. The key insight here is to train editing and layer generation together, while also checking that the predicted layers still compose into an image consistent with the edit.
Key Findings
- One token sequence covers both outputs. In-Context RGB-Alpha Token Modeling represents the edited canvas alongside an ordered stack of RGBA layers. This links the semantic instruction to both the visible result and its editable components.
- The composite is part of the learning target. Compositional Edit-Prior Distillation transfers the pretrained editing prior to the canvas and to the image composed from predicted layers. This addresses a practical failure mode: individually plausible layers may not reconstruct the intended edit when combined.
- The data supports layered supervision. LayerEditor-100K pairs image edits with ordered RGBA annotations, and the work introduces benchmarks for layered editing. The contribution is therefore not only a model design, but also a training and evaluation setup for this task.
Method and Positioning
The approach treats layer production as part of instruction-driven editing rather than as an independent decomposition stage. That makes it distinct from workflows that first edit a flat image and then attempt to recover layers, or decompose an image without conditioning the process on an edit instruction. Its central quality criterion is joint: the edit should be faithful, the layers should be useful individually, and their composition should remain coherent.
Who It’s For—and the Trade-off
A good fit for research and creative workflows that need localized edits, reusable assets, or layer-level control after generation. It is less compelling when a flat image is sufficient, since layered outputs add supervision and representation complexity; the results also depend on the quality of ordered RGBA annotations and on how well the predicted layers recombine.