One name typed, two pictures back
Nothing was masked or painted by hand. SAM 3 finds the thing from its name. Qwen Image 2.1 takes it out along with its shadow (most of the time) and reflection. Then only the thing's area and what changed with it are pasted back, so the rest of the picture is the original.
Both results come out the same size, so they stack as two layers.
About 12 seconds a run on an RTX 5090, about 6 with the optional speed LoRA.