LayerEditor turns one text edit into separate image layers in 0.9 seconds

Researchers unveiled LayerEditor, a diffusion model that outputs ready-to-use RGBA foreground, background and mask layers plus a reusable edit-prior from a single text prompt. It reaches 0.78 IoU on COCO-Edit versus 0.55 for MRT 20B, with inference taking 0.92 s on an RTX 4090.

• 0.78 IoU on COCO-Edit, a 23% gain over the best baseline
• Three RGBA layers plus an edit-prior in 0.92 s on RTX 4090
• Edit-Recall of 0.84 versus 0.63 for MRT 20B
• The ~256 KB edit-prior can be reused on other images