Compare#

Side-by-side comparisons of withoutBG Open Weights and Pro against other background removal tools. For the open-model directory, see /models and the rembg hub.

Looking for a remove.bg alternative? See the open-source and API alternatives to remove.bg: licenses, how to run each one, and accuracy ranks.

Leaderboard

Mean error against ground-truth alpha mattes on Bench v0 (60 images, same inputs for every model). Lower is better. Pick a metric to rank by.

#	Model
1	withoutBG Pro Model API	0.0240	0.0274	0.0536
2	withoutBG Pro Model Self-Host	0.0276	0.0369	0.0608
3	withoutBG Open Weights	0.0292	0.0452	0.0672
4	BiRefNet General	0.0426	0.0495	0.0925
5	BiRefNet (legacy)	0.0439	0.0544	0.0976
6	BiRefNet General Lite	0.0522	0.0538	0.1090
7	BiRefNet DIS	0.0575	0.0585	0.1167
8	BiRefNet Massive	0.0714	0.0559	0.1204
9	BiRefNet Portrait	0.0714	0.0647	0.1455
10	IS-Net	0.0940	0.0612	0.1614
11	BiRefNet HRSOD	0.1069	0.0714	0.1638
12	U²-Net	0.1283	0.0798	0.2380
13	Silueta	0.1344	0.0845	0.2446
14	U²-NetP	0.1405	0.0803	0.2527

MAE is overall alpha accuracy; MGE (gradient error) and band MAE (error near the matte edge) measure edge quality. withoutBG routes hard opaque objects, flat scenes and vehicles to BiRefNet (17 of 60 images here). Methodology and per-image data: open background removal benchmark.

withoutBG Open Weights

withoutBG Pro Model API

withoutBG Pro Model Self-Host

BiRefNet General

Open and commercially permissive models.

U²-Net

Open and commercially permissive models.

U²-Net vs U²-NetP

remove.bg

Closed, paid cloud service. Side by side only: remove.bg has no Bench v0 scores. All remove.bg alternatives.