Back to Ege Orcun
Ege Orcun /

LU
Lucida

Open WeightsSpecializedimageUpdated May 1, 2026

Lucida: Precision Background Removal That Keeps What Matters

Model Overview

Lucida is a state-of-the-art BiRefNet-based image matting model fine-tuned specifically for the cases where general-purpose background removers fail: semi-transparent glass, camouflaged subjects, logos and typography with soft shadows, glow/VFX effects, illustrations, and print-style designs. Trained iteratively across 7 versions on ~52,882 image/alpha pairs spanning 9 specialized categories, Lucida v7 achieves the lowest overall MAE (0.0257) of any model measured — including commercial APIs.


📊 Benchmark Results (MAE — Lower is Better)

Specification Table
Category (n)Lucida v7InSPyReNetIdeogram*RMBG-2.0BiRefNet-HR
Camouflage (25)0.02700.05820.11790.14050.0752
Transparent (25)0.03580.07250.03430.07410.0687
Complex (29)0.04840.01100.10460.02410.0385
Thin (36)0.03220.01660.05210.01800.0196
Hair (40)0.00930.00690.01120.00450.0048
Text (12)0.00910.01810.01230.01730.0207
FX (12)0.01800.02690.01650.02680.0272
Illustration (12)0.00920.02420.02150.01250.0157
Design (12)0.02350.05870.05180.04780.0544
OVERALL (203)0.02570.02950.05070.04010.0346

*Ideogram = fal.ai Ideogram remove-background (commercial API)


✨ Key Features

Specification Table
FeatureDescription
Best Overall AccuracyMAE 0.0257 — lowest error of any measured model, commercial or open
Camouflage Champion2x better than best open model; 4.4x better than commercial reference
Print Design & Text Logos2x+ better than all competitors including commercial APIs
9 Specialized CategoriesGlass, camouflage, complex, thin, hair, text, VFX, illustration, print designs
ComfyUI IntegrationDistributed in ComfyUI official BiRefNet collection as lucida.safetensors

🔧 Usage

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

Recommended input resolution: 1024×1024. Loadable with HuggingFace transformers.


🔗 Resources

Specification Table

📜 License

MIT License — weights are released under MIT, suitable for commercial use (see repo for training data license notes).

Key Features

Best Overall Accuracy: MAE 0.0257 overall — lowest error of any model measured, beating commercial reference Ideogram (0.0507) and open models RMBG-2.0 (0.0401), BiRefNet-HR (0.0346)

Feature 01

Camouflage Category: MAE 0.0270 — 2x better than best open competitor (InSPyReNet 0.0582) and 4.4x better than commercial reference

Feature 02

Print Design & Text: MAE 0.0235 design, 0.0091 text — 2x+ better than all competitors including commercial APIs

Feature 03

9 Specialized Categories: Transparent glass, camouflage, complex scenes, thin structures, hair, text/logos, VFX/glow effects, illustration, print designs

Feature 04

ComfyUI Integration: Distributed in ComfyUI's official BiRefNet model collection as lucida.safetensors

Feature 05

You might also want to compare

Verified Sources

Tags

image-segmentationbackground-removalmattingbirefnetcomfyui

Model Specs

open-weights

Parameters

Undisclosed

Context Window

unknown

License

MIT

Deployment

self-hostableapi-only

Resources & Links

Curator Notes

BiRefNet fine-tune by Ege Orcun. v7 achieves best overall MAE (0.0257) of all measured models including commercial APIs. MIT license. Available on HuggingFace and ComfyUI official collection.

Compare Specs

Compare parameters, context windows, modalities, and benchmark scores of this model side-by-side with others.

Compare Model