LayerEditor turns edit prompts into reusable RGBA layers
TL;DR
- LayerEditor outputs both an edited canvas and an ordered RGB-Alpha layer stack within a single token sequence, enabling layer-level manipulation after an instruction edit.
- Compositional Edit-Prior Distillation transfers the pretrained editing prior to the edited canvas and the image composed from predicted layers to preserve edit quality.
- The authors release LayerEditor-100K, a paired image-editing dataset with ordered RGBA layer annotations, plus dedicated benchmarks for layered editing.
LayerEditor, described in a new arXiv paper, couples instruction-guided image editing with the generation of reusable RGBA layers, so an edit arrives as a stack a designer can rework rather than a flat composite. The authors call the core technique In-Context RGB-Alpha Token Modeling, which "jointly represents the edited canvas and an ordered sequence of RGB-Alpha layers within a single token sequence."
To stop the model losing its editing skill once retrained to emit layers, the team add Compositional Edit-Prior Distillation, which "transfers its editing prior to both the edited canvas and the image composed from the predicted layers." Training and evaluation lean on LayerEditor-100K, which the paper describes as "a large-scale paired image-editing dataset with ordered RGBA layer annotations", plus purpose-built benchmarks for layered editing.
The abstract reports "strong editing fidelity while generating accurate RGBA layers" but publishes no numeric comparisons against baselines; those sit in the full paper rather than the summary.
Originally reported by paper
Read the original article →Original headline: LayerEditor Generates Reusable RGBA Layers From Edit Instructions, Trained on 100K New Pairs