Lucida

Background removal that keeps what matters: glass, camouflage, text, glow and line art.

Try it live (GPU demo) Model weights (MIT) GitHub — code & benchmark Run the web UI locally (Docker)

Benchmark — 191 images, 8 categories (MAE, lower is better)

category	lucida-v13	inspyrenet	ideogram*	rmbg-2.0
design (print art)	0.0254	0.0587	0.0518	0.0478
camouflage	0.0227	0.0582	0.1179	0.1405
transparent	0.0338	0.0725	0.0343	0.0741
text / logos	0.0103	0.0181	0.0123	0.0173
illustration	0.0082	0.0242	0.0215	0.0125
complex	0.0465	0.0110	0.1046	0.0241
thin	0.0321	0.0166	0.0521	0.0211
hair	0.0093	0.0069	0.0112	0.0045

* ideogram = fal.ai commercial API, used as the quality reference. Full table & methodology on GitHub.

Examples

camouflage comparison — Camouflage — body paint in magnolia petals: Lucida finds the subject; the best open competitor keeps the whole image.

transparency comparison — Glass — real alpha in the lens/vessel areas, ahead of the commercial reference on this image.

text comparison — Text & logos — lettering with soft drop shadow preserved.

illustration comparison — Illustration — anime rider: 5x better than the commercial reference on this image.

Use it in three lines

from transformers import AutoModelForImageSegmentation
model = AutoModelForImageSegmentation.from_pretrained(
    "egeorcun/lucida", trust_remote_code=True)

MIT license. Fine-tuned from BiRefNet_HR. If Lucida is useful to you, a ⭐ on GitHub helps a lot.