The Background Removal Library Landscape — BiRefNet, ViTMatte, MatAnyone, and Friends

Mon, 13 Apr 2026 00:00:00 +0900

Overview

Building popcon-matting-bench forced a survey of every credible open-source matting library. The space breaks into three eras: classical algorithms (pymatting, FBA), trimap-free deep models (BiRefNet, ViTMatte), and the new generation of stable video matting (MatAnyone). This post maps the landscape and notes which model wins for which job.

Today’s Exploration Map

graph TD
 A[Background Removal Need] --> B{Trimap available?}
 B -->|Yes| C[Classical: pymatting / FBA]
 B -->|No| D{Image or Video?}
 D -->|Image| E[BiRefNet / ViTMatte]
 D -->|Video| F[MatAnyone]
 E --> G[Toon-style?]
 G -->|Yes| H[MatteoKartoon BiRefNet fork]
 G -->|No| I[ZhengPeng7 BiRefNet]

BiRefNet — High-Resolution Dichotomous Segmentation

ZhengPeng7/BiRefNet (CAAI AIR 2024) is the model nearly every recent background-removal demo, including birefnet.top, is built on. It targets dichotomous image segmentation — high-resolution binary foreground/background masks — and it does so with a bilateral reference design: two streams (one for the source image, one for a reference) cross-attend through the U-Net decoder.

Two things make BiRefNet stand out:

Resolution. Most segmentation models top out at 1024×1024; BiRefNet has weights for 2048×2048 and the architecture handles arbitrary aspect ratios well. For e-commerce or asset extraction, this is decisive.
Generalization. The default general checkpoint handles humans, products, animals, and abstract shapes. Specialized variants (portrait, matting, dis5k_general) are available on Hugging Face if you need accuracy on a specific domain.

MatteoKartoon/BiRefNet is a fork called ToonOut that fine-tunes BiRefNet on toon/sticker datasets — relevant for any product generating animated emoji or cartoon assets. The fork mostly changes the training data and the evaluation harness; the core model is unchanged.

ViTMatte — ViT Backbone, Trimap Input

hustvl/ViTMatte (Information Fusion vol.103, March 2024) takes a different bet: a Vision Transformer backbone with explicit trimap input. The trimap (foreground / background / unknown regions) is a hard requirement, which makes ViTMatte less plug-and-play than BiRefNet but significantly more accurate on hair, fur, and translucent edges when you can supply one. The pipeline pattern is: BiRefNet produces an initial mask → erode/dilate to a trimap → ViTMatte refines the alpha at sub-pixel quality.

MatAnyone — Stable Video Matting (CVPR 2025)

pq-yang/MatAnyone targets the hardest matting problem: temporal stability. Frame-by-frame matting on video produces flicker — the alpha mask jitters by a pixel or two between frames, which the human eye picks up immediately. MatAnyone introduces memory-augmented region propagation: the model carries a memory bank of past frames’ high-confidence regions and uses them to constrain the current frame’s mask. The result is video matting that doesn’t shimmer.

This matters for popcon’s animated-emoji pipeline: extracting a clean alpha across 30 frames requires either MatAnyone or a hand-rolled temporal smoother on top of BiRefNet.

pymatting and FBA — The Classical Baselines

pymatting/pymatting (1.9k stars, MIT) implements every classical alpha matting method worth knowing — Closed-Form, KNN, Large Kernel, Random Walk, Shared Sampling — plus Fast Multi-Level Foreground Estimation. It requires a trimap but runs entirely on CPU (with optional CuPy/PyOpenCL acceleration for foreground estimation). The library is also the foundation of Rembg, the most widely deployed open-source background removal tool.

MarcoForte/FBA_Matting is the official “F, B, Alpha” matting paper repo — predicts foreground color, background color, and alpha jointly, which gives much cleaner composites when the foreground and background colors differ subtly.

The classical methods aren’t obsolete. For high-throughput batch processing where a trimap is available (e.g., chroma-key footage, scanned documents), they’re often 10-100× faster than deep models with comparable quality.

Architecture Pattern for popcon-matting-bench

graph LR
 A[Input Image] --> B[BiRefNet <br/> coarse mask]
 B --> C[Trimap generation <br/> erode/dilate]
 C --> D[ViTMatte <br/> or pymatting]
 D --> E[FBA Foreground <br/> estimation]
 E --> F[Composite output]

The benchmark repo’s job is to score each model on standard datasets (DIS-5K, AIM-500, RealWorldPortrait636) and produce a comparison harness. Key metrics: SAD, MSE, Grad, Conn for alpha quality; mIoU for binary segmentation; latency per 1024×1024 image on a single A100.

Insights

The matting space has bifurcated cleanly: BiRefNet owns high-resolution segmentation, ViTMatte owns trimap-refined alpha, MatAnyone owns video, and pymatting/FBA own the classical CPU path. There’s no single model that wins everywhere — production pipelines almost always cascade two or three. The interesting business question is no longer which model but what trimap workflow you want: zero-shot (BiRefNet alone) trades quality for ergonomics, while two-stage (BiRefNet → ViTMatte) trades latency for hair-grade accuracy. ToonOut shows the path forward for verticalized matting — the base model is good enough that fine-tuning on niche datasets is a low-risk play.

Pymatting on ICE-ICE-BEAR-BLOG