Lucida: open-source background removal for glass and hair

Lucida is an open-source image matting model built specifically to handle the edge cases where conventional background removers fall apart. It cleanly cuts away backgrounds while preserving tricky elements like semi-transparent glass, camouflaged subjects, soft drop shadows, glow effects, and fine hair strands.

Most general-purpose models either turn glass containers into opaque blocks or slice through complex outlines entirely. Lucida is a fine-tune of the BiRefNet architecture, engineered specifically for semi-transparent objects, logos and typography with soft shadows, illustrations, and print-style designs like stickers and t-shirt prints.

What it does and how it compares

In the developer's benchmark across 203 images spanning 9 categories, Lucida achieved the lowest overall mean absolute error at 0.0257, beating InSPyReNet at 0.0295 and the commercial fal.ai Ideogram reference. It is the first Lucida version to lead the overall table.

Its strengths shine brightest on edge cases. On camouflage imagery, Lucida cut error rates by more than half compared to InSPyReNet and performed 4.4 times better than Ideogram. On print and sticker designs, its 0.0235 score was more than twice as accurate as competing models, easily isolating distressed text and soft smoke edges.

The benchmark also highlights its honest limitations. Ideogram still holds a razor-thin lead in sheer transparency error at 0.0343 versus Lucida's 0.0358. For complex multi-object product shots and thin wiry structures, InSPyReNet remains superior, while RMBG-2.0 retains a slight edge on pure hair segmentation.

Simply put: if your daily workload revolves around complex machinery or multi-object product arrangements, existing models like InSPyReNet might still serve you best. But for glassware, typography, camouflage, and illustrations, Lucida is the strongest open tool available.

Specs and getting it running

The entire project is released under the permissive MIT license. The model weights, code, training recipes, and datasets are all fully accessible on GitHub and Hugging Face.

At under 900MB, Lucida is remarkably lightweight. It runs comfortably on consumer-grade GPUs at its native 1024x1024 resolution without hogging system resources. It requires Python 3.12 or newer, and you can run it via uv or pip, complete with an edge-refinement pass and color decontamination. For quick tests, there is also a free ZeroGPU demo on Hugging Face that finishes runs in just a few seconds.

Why it matters

For designers, sticker creators, and solo developers, Lucida solves one of the most frustrating bottlenecks in automated image pipelines. Until now, pulling semi-transparent glassware, VFX glows, or delicate drop shadows out of complex backgrounds required either manual pen work or expensive proprietary APIs.

We think the open approach here sets a great example. Releasing the dataset and an honest benchmark that shows exactly where the model wins and loses makes it instantly practical. A sub-900MB model that fixes known edge cases is far more valuable to production workflows than a bloated generalist.