Open background removal benchmark#
Bench v0 uses the fixed image set from the Clipping Magic compare page (~37 images across people, motion, plants, objects, animals, and film). Every model runs on the same inputs. Ground-truth alpha mattes are processed to match each source resolution; compare pages show MAE, edge quality (MGE), and connectivity alongside side-by-sides.
Methodology
- Same source images for every client (Bench v0 stems).
- High-res GT mattes are resized to the inference input size before scoring.
- Metrics use full-resolution predicted alpha when available (raw response).
- Display assets on the site are max long side 600px.
- withoutBG hosts this hub — category callouts matter more than a single overall crown when margins are small.