feat: Add MODNet portrait matting (#108 )

* feat: Add MODNet portrait matting * docs: Update docs and example of portrait matting * fix: Fix linting issue
feat: Add EdgeFace recognition model (#105 )
2026-05-18 06:37:47 +00:00 · 2026-04-11 23:30:32 +09:00 · 2026-04-04 20:11:28 +09:00 · 2026-04-04 18:28:27 +09:00 · 2026-04-03 10:10:51 +09:00
67 changed files with 2473 additions and 1762 deletions
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -18,6 +18,13 @@ repos:
      - id: debug-statements
      - id: check-ast

+  # Strip Jupyter notebook outputs
+  - repo: https://github.com/kynan/nbstripout
+    rev: 0.9.1
+    hooks:
+      - id: nbstripout
+        files: ^examples/
+
  # Ruff - Fast Python linter and formatter
  - repo: https://github.com/astral-sh/ruff-pre-commit
    rev: v0.14.10
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -0,0 +1,6 @@
+<!-- Cursor agent instructions — shared with CLAUDE.md -->
+<!-- See CLAUDE.md for full project instructions for AI coding agents. -->
+
+# AGENTS.md
+
+Please read and follow all instructions in [CLAUDE.md](./CLAUDE.md).
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -0,0 +1,81 @@
+# CLAUDE.md
+
+Project instructions for AI coding agents.
+
+## Project Overview
+
+UniFace is a Python library for face detection, recognition, tracking, landmark analysis, face parsing, gaze estimation, age/gender detection. It uses ONNX Runtime for inference.
+
+## Code Style
+
+- Python 3.10+ with type hints
+- Line length: 120
+- Single quotes for strings, double quotes for docstrings
+- Google-style docstrings
+- Formatter/linter: Ruff (config in `pyproject.toml`)
+- Run `ruff format .` and `ruff check . --fix` before committing
+
+## Commit Messages
+
+Follow [Conventional Commits](https://www.conventionalcommits.org/) with a **capitalized** description:
+
+```
+<type>: <Capitalized short description>
+```
+
+Types: `feat`, `fix`, `docs`, `style`, `refactor`, `perf`, `test`, `build`, `ci`, `chore`
+
+Examples:
+- `feat: Add gaze estimation model`
+- `fix: Correct bounding box scaling for non-square images`
+- `ci: Add nbstripout pre-commit hook`
+- `docs: Update installation instructions`
+- `refactor: Unify attribute/detector base classes`
+
+## Testing
+
+```bash
+pytest -v --tb=short
+```
+
+Tests live in `tests/`. Run the full suite before submitting changes.
+
+## Pre-commit
+
+Pre-commit hooks handle formatting, linting, security checks, and notebook output stripping. Always run:
+
+```bash
+pre-commit install
+pre-commit run --all-files
+```
+
+## Project Structure
+
+```
+uniface/            # Main package
+  detection/        # Face detection models (SCRFD, RetinaFace, YOLOv5, YOLOv8)
+  recognition/      # Face recognition/verification (AdaFace, ArcFace, EdgeFace, MobileFace, SphereFace)
+  landmark/         # Facial landmark models
+  tracking/         # Object tracking (ByteTrack)
+  parsing/          # Face parsing/segmentation (BiSeNet, XSeg)
+  gaze/             # Gaze estimation
+  headpose/         # Head pose estimation
+  attribute/        # Age, gender, emotion detection
+  spoofing/         # Anti-spoofing (MiniFASNet)
+  privacy/          # Face anonymization
+  stores/           # Vector stores (FAISS)
+  constants.py      # Model weight URLs and checksums
+  model_store.py    # Model download/cache management
+  analyzer.py       # High-level FaceAnalyzer API
+  types.py          # Shared type definitions
+tests/              # Unit tests
+examples/           # Jupyter notebooks (outputs are auto-stripped)
+docs/               # MkDocs documentation
+```
+
+## Key Conventions
+
+- New models: add class in submodule, register weights in `constants.py`, export in `__init__.py`
+- Dependencies: managed in `pyproject.toml`
+- All ONNX models are downloaded on demand with SHA256 verification
+- Do not commit notebook outputs; `nbstripout` pre-commit hook handles this
--- a/README.md
+++ b/README.md
@@ -26,10 +26,11 @@
 ## Features

 - **Face Detection** — RetinaFace, SCRFD, YOLOv5-Face, and YOLOv8-Face with 5-point landmarks
- **Face Recognition** — ArcFace, MobileFace, and SphereFace embeddings
+- **Face Recognition** — AdaFace, ArcFace, EdgeFace, MobileFace, and SphereFace embeddings
 - **Face Tracking** — Multi-object tracking with [BYTETracker](https://github.com/yakhyo/bytetrack-tracker) for persistent IDs across video frames
 - **Facial Landmarks** — 106-point landmark localization module (separate from 5-point detector landmarks)
 - **Face Parsing** — BiSeNet semantic segmentation (19 classes), XSeg face masking
+- **Portrait Matting** — Trimap-free alpha matte with MODNet (background removal, green screen, compositing)
 - **Gaze Estimation** — Real-time gaze direction with MobileGaze
 - **Head Pose Estimation** — 3D head orientation (pitch, yaw, roll) with 6D rotation representation
 - **Attribute Analysis** — Age, gender, race (FairFace), and emotion
@@ -40,6 +41,39 @@

 ---

+## Visual Examples
+
+<table>
+  <tr>
+    <td align="center"><b>Face Detection</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/detection.jpg" width="100%"></td>
+    <td align="center"><b>Gaze Estimation</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/gaze.jpg" width="100%"></td>
+  </tr>
+  <tr>
+    <td align="center"><b>Head Pose Estimation</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/headpose.jpg" width="100%"></td>
+    <td align="center"><b>Age &amp; Gender</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/age_gender.jpg" width="100%"></td>
+  </tr>
+  <tr>
+    <td align="center" colspan="2"><b>Face Verification</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/verification.jpg" width="80%"></td>
+  </tr>
+  <tr>
+    <td align="center" colspan="2"><b>106-Point Landmarks</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/landmarks.jpg" width="36%"></td>
+  </tr>
+  <tr>
+    <td align="center" colspan="2"><b>Face Parsing</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/parsing.jpg" width="80%"></td>
+  </tr>
+  <tr>
+    <td align="center" colspan="2"><b>Face Segmentation</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/segmentation.jpg" width="80%"></td>
+  </tr>
+  <tr>
+    <td align="center" colspan="2"><b>Portrait Matting</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/matting.jpg" width="100%"></td>
+  </tr>
+  <tr>
+    <td align="center" colspan="2"><b>Face Anonymization</b><br><img src="https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/anonymization.jpg" width="100%"></td>
+  </tr>
+</table>
+
+---
+
 ## Installation

 **Standard installation**
@@ -156,6 +190,32 @@ for face in faces:

 ---

+## Example (Portrait Matting)
+
+```python
+import cv2
+import numpy as np
+from uniface.matting import MODNet
+
+matting = MODNet()
+
+image = cv2.imread("portrait.jpg")
+matte = matting.predict(image)  # (H, W) float32 in [0, 1]
+
+# Transparent PNG
+rgba = cv2.cvtColor(image, cv2.COLOR_BGR2BGRA)
+rgba[:, :, 3] = (matte * 255).astype(np.uint8)
+cv2.imwrite("transparent.png", rgba)
+
+# Green screen
+matte_3ch = matte[:, :, np.newaxis]
+bg = np.full_like(image, (0, 177, 64), dtype=np.uint8)
+result = (image * matte_3ch + bg * (1 - matte_3ch)).astype(np.uint8)
+cv2.imwrite("green_screen.jpg", result)
+```
+
+---
+
 ## Jupyter Notebooks

 | Example | Colab | Description |
@@ -172,6 +232,7 @@ for face in faces:
 | [10_face_vector_store.ipynb](examples/10_face_vector_store.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/10_face_vector_store.ipynb) | FAISS-backed face database |
 | [11_head_pose_estimation.ipynb](examples/11_head_pose_estimation.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/11_head_pose_estimation.ipynb) | Head pose estimation (pitch, yaw, roll) |
 | [12_face_recognition.ipynb](examples/12_face_recognition.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/12_face_recognition.ipynb) | Standalone face recognition pipeline |
+| [13_portrait_matting.ipynb](examples/13_portrait_matting.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/13_portrait_matting.ipynb) | Portrait matting with MODNet |

 ---

@@ -212,6 +273,7 @@ https://yakhyo.github.io/uniface/concepts/execution-providers/
 | Recognition | MS1MV2 | MobileFace, SphereFace |
 | Recognition | WebFace600K | ArcFace |
 | Recognition | WebFace4M / 12M | AdaFace |
+| Recognition | MS1MV2 | EdgeFace |
 | Gaze | Gaze360 | MobileGaze |
 | Head Pose | 300W-LP | HeadPose (ResNet, MobileNet) |
 | Parsing | CelebAMask-HQ | BiSeNet |
@@ -243,10 +305,12 @@ If you plan commercial use, verify model license compatibility.
 | Detection | [yolov8-face-onnx-inference](https://github.com/yakhyo/yolov8-face-onnx-inference) | - | YOLOv8-Face ONNX Inference |
 | Tracking | [bytetrack-tracker](https://github.com/yakhyo/bytetrack-tracker) | - | BYTETracker Multi-Object Tracking |
 | Recognition | [face-recognition](https://github.com/yakhyo/face-recognition) | ✓ | MobileFace, SphereFace Training |
+| Recognition | [edgeface-onnx](https://github.com/yakhyo/edgeface-onnx) | - | EdgeFace ONNX Inference |
 | Parsing | [face-parsing](https://github.com/yakhyo/face-parsing) | ✓ | BiSeNet Face Parsing |
 | Parsing | [face-segmentation](https://github.com/yakhyo/face-segmentation) | - | XSeg Face Segmentation |
 | Gaze | [gaze-estimation](https://github.com/yakhyo/gaze-estimation) | ✓ | MobileGaze Training |
 | Head Pose | [head-pose-estimation](https://github.com/yakhyo/head-pose-estimation) | ✓ | Head Pose Training (6DRepNet-style) |
+| Matting | [modnet](https://github.com/yakhyo/modnet) | - | MODNet Portrait Matting |
 | Anti-Spoofing | [face-anti-spoofing](https://github.com/yakhyo/face-anti-spoofing) | - | MiniFASNet Inference |
 | Attributes | [fairface-onnx](https://github.com/yakhyo/fairface-onnx) | - | FairFace ONNX Inference |

@@ -270,3 +334,6 @@ Questions or feedback:
 ## License

 This project is licensed under the [MIT License](LICENSE).
+
+> **Disclaimer:** This project is not affiliated with or related to
+> [Uniface](https://uniface.com/) by Rocket Software.
--- a/assets/demos/age_gender.jpg
+++ b/assets/demos/age_gender.jpg
--- a/assets/demos/anonymization.jpg
+++ b/assets/demos/anonymization.jpg
--- a/assets/demos/detection.jpg
+++ b/assets/demos/detection.jpg
--- a/assets/demos/gaze.jpg
+++ b/assets/demos/gaze.jpg
--- a/assets/demos/headpose.jpg
+++ b/assets/demos/headpose.jpg
--- a/assets/demos/landmarks.jpg
+++ b/assets/demos/landmarks.jpg
--- a/assets/demos/matting.jpg
+++ b/assets/demos/matting.jpg
--- a/assets/demos/parsing.jpg
+++ b/assets/demos/parsing.jpg
--- a/assets/demos/segmentation.jpg
+++ b/assets/demos/segmentation.jpg
--- a/assets/demos/src_friends.jpg
+++ b/assets/demos/src_friends.jpg
--- a/assets/demos/src_man1.jpg
+++ b/assets/demos/src_man1.jpg
--- a/assets/demos/src_man2.jpg
+++ b/assets/demos/src_man2.jpg
--- a/assets/demos/src_man3.jpg
+++ b/assets/demos/src_man3.jpg
--- a/assets/demos/src_meeting.jpg
+++ b/assets/demos/src_meeting.jpg
--- a/assets/demos/src_portrait1.jpg
+++ b/assets/demos/src_portrait1.jpg
--- a/assets/demos/verification.jpg
+++ b/assets/demos/verification.jpg
--- a/docs/concepts/overview.md
+++ b/docs/concepts/overview.md
@@ -26,6 +26,7 @@ graph TB
        HPOSE[Head Pose]
        PARSE[Parsing]
        SPOOF[Anti-Spoofing]
+        MATT[Matting]
        PRIV[Privacy]
    end

@@ -42,6 +43,7 @@ graph TB
    end

    IMG --> DET
+    IMG --> MATT
    DET --> REC
    DET --> LMK
    DET --> ATTR
@@ -115,11 +117,12 @@ def detect(self, image: np.ndarray) -> list[Face]:
 ```
 uniface/
 ├── detection/      # Face detection (RetinaFace, SCRFD, YOLOv5Face, YOLOv8Face)
-├── recognition/    # Face recognition (AdaFace, ArcFace, MobileFace, SphereFace)
+├── recognition/    # Face recognition (AdaFace, ArcFace, EdgeFace, MobileFace, SphereFace)
 ├── tracking/       # Multi-object tracking (BYTETracker)
 ├── landmark/       # 106-point landmarks
 ├── attribute/      # Age, gender, emotion, race
 ├── parsing/        # Face semantic segmentation
+├── matting/        # Portrait matting (MODNet)
 ├── gaze/           # Gaze estimation
 ├── headpose/       # Head pose estimation
 ├── spoofing/       # Anti-spoofing
--- a/docs/contributing.md
+++ b/docs/contributing.md
@@ -47,6 +47,38 @@ pre-commit run --all-files

 ---

+## Commit Messages
+
+We follow [Conventional Commits](https://www.conventionalcommits.org/):
+
+```
+<type>: <short description>
+```
+
+| Type         | When to use                                      |
+|--------------|--------------------------------------------------|
+| **feat**     | New feature or capability                        |
+| **fix**      | Bug fix                                          |
+| **docs**     | Documentation changes                            |
+| **style**    | Formatting, whitespace (no logic change)         |
+| **refactor** | Code restructuring without changing behavior     |
+| **perf**     | Performance improvement                          |
+| **test**     | Adding or updating tests                         |
+| **build**    | Build system or dependencies                     |
+| **ci**       | CI/CD and pre-commit configuration               |
+| **chore**    | Routine maintenance and tooling                  |
+
+**Examples:**
+
+```
+feat: Add gaze estimation model
+fix: Correct bounding box scaling for non-square images
+ci: Add nbstripout pre-commit hook
+docs: Update installation instructions
+```
+
+---
+
 ## Pull Request Process

 1. Fork the repository
--- a/docs/index.md
+++ b/docs/index.md
@@ -36,7 +36,7 @@ ONNX-optimized detectors (RetinaFace, SCRFD, YOLO) with 5-point landmarks.

 <div class="feature-card" markdown>
 ### :material-account-check: Face Recognition
-AdaFace, ArcFace, MobileFace, and SphereFace embeddings for identity verification.
+AdaFace, ArcFace, EdgeFace, MobileFace, and SphereFace embeddings for identity verification.
 </div>

 <div class="feature-card" markdown>
--- a/docs/license-attribution.md
+++ b/docs/license-attribution.md
@@ -20,5 +20,6 @@ UniFace is released under the [MIT License](https://opensource.org/licenses/MIT)
 | SphereFace | [yakhyo/face-recognition](https://github.com/yakhyo/face-recognition) | MIT |
 | BiSeNet | [yakhyo/face-parsing](https://github.com/yakhyo/face-parsing) | MIT |
 | MobileGaze | [yakhyo/gaze-estimation](https://github.com/yakhyo/gaze-estimation) | MIT |
+| MODNet | [yakhyo/modnet](https://github.com/yakhyo/modnet) | Apache-2.0 |
 | MiniFASNet | [yakhyo/face-anti-spoofing](https://github.com/yakhyo/face-anti-spoofing) | Apache-2.0 |
 | FairFace | [yakhyo/fairface-onnx](https://github.com/yakhyo/fairface-onnx) | CC BY 4.0 |
--- a/docs/models.md
+++ b/docs/models.md
@@ -156,6 +156,24 @@ Face recognition using angular softmax loss.

 ---

+### EdgeFace
+
+Efficient face recognition designed for edge devices, using EdgeNeXt backbone with optional LoRA compression.
+
+| Model Name      | Backbone | Params | MFLOPs | Size  | LFW    | CALFW  | CPLFW  | CFP-FP | AgeDB-30 |
+| --------------- | -------- | ------ | ------ | ----- | ------ | ------ | ------ | ------ | -------- |
+| `XXS` :material-check-circle: | EdgeNeXt | 1.24M  | 94     | ~5 MB | 99.57% | 94.83% | 90.27% | 93.63% | 94.92%   |
+| `XS_GAMMA_06`   | EdgeNeXt | 1.77M  | 154    | ~7 MB | 99.73% | 95.28% | 91.58% | 94.71% | 96.08%   |
+| `S_GAMMA_05`    | EdgeNeXt | 3.65M  | 306    | ~14 MB | 99.78% | 95.55% | 92.48% | 95.74% | 97.03%  |
+| `BASE`          | EdgeNeXt | 18.2M  | 1399   | ~70 MB | 99.83% | 96.07% | 93.75% | 97.01% | 97.60%  |
+
+!!! info "Training Data & Reference"
+    **Paper**: [EdgeFace: Efficient Face Recognition Model for Edge Devices](https://arxiv.org/abs/2307.01838v2) (IEEE T-BIOM 2024)
+
+    **Source**: [github.com/otroshi/edgeface](https://github.com/otroshi/edgeface) | [github.com/yakhyo/edgeface-onnx](https://github.com/yakhyo/edgeface-onnx)
+
+---
+
 ## Facial Landmark Models

 ### 106-Point Landmark Detection
@@ -353,6 +371,36 @@ XSeg from DeepFaceLab outputs masks for face regions. Requires 5-point landmarks

 ---

+## Portrait Matting Models
+
+### MODNet
+
+MODNet (Real-Time Trimap-Free Portrait Matting) produces soft alpha mattes from full images without requiring a trimap. Uses MobileNetV2 backbone with low-resolution, high-resolution, and fusion branches.
+
+| Model Name | Variant | Size | Use Case |
+| ---------- | ------- | ---- | -------- |
+| `PHOTOGRAPHIC` :material-check-circle: | High-quality | 25 MB | Portrait photos |
+| `WEBCAM` | Real-time | 25 MB | Webcam feeds |
+
+!!! info "Model Details"
+    **Paper**: [MODNet: Real-Time Trimap-Free Portrait Matting via Objective Decomposition](https://arxiv.org/abs/2011.11961) (AAAI 2022)
+
+    **Source**: [yakhyo/modnet](https://github.com/yakhyo/modnet) — ported weights and clean inference codebase
+
+    **Output**: Alpha matte `(H, W)` in `[0, 1]`
+
+**Applications:**
+
+- Background removal / replacement
+- Green screen compositing
+- Video conferencing virtual backgrounds
+- Portrait editing
+
+!!! note "Input Requirements"
+    Operates on full images (not face crops). No trimap or face detection required.
+
+---
+
 ## Anti-Spoofing Models

 ### MiniFASNet Family
@@ -402,6 +450,7 @@ See [Model Cache & Offline Use](concepts/model-cache-offline.md) for full detail
 - **Head Pose Estimation**: [yakhyo/head-pose-estimation](https://github.com/yakhyo/head-pose-estimation) - 6D rotation head pose estimation training and ONNX models
 - **Face Parsing Training**: [yakhyo/face-parsing](https://github.com/yakhyo/face-parsing) - BiSeNet training code and pretrained weights
 - **Face Segmentation**: [yakhyo/face-segmentation](https://github.com/yakhyo/face-segmentation) - XSeg ONNX Inference
+- **Portrait Matting**: [yakhyo/modnet](https://github.com/yakhyo/modnet) - MODNet ported weights and inference (from [ZHKKKe/MODNet](https://github.com/ZHKKKe/MODNet))
 - **Face Anti-Spoofing**: [yakhyo/face-anti-spoofing](https://github.com/yakhyo/face-anti-spoofing) - MiniFASNet ONNX inference (weights from [minivision-ai/Silent-Face-Anti-Spoofing](https://github.com/minivision-ai/Silent-Face-Anti-Spoofing))
 - **FairFace**: [yakhyo/fairface-onnx](https://github.com/yakhyo/fairface-onnx) - FairFace ONNX inference for race, gender, age prediction
 - **InsightFace**: [deepinsight/insightface](https://github.com/deepinsight/insightface) - Model architectures and pretrained weights
@@ -414,4 +463,5 @@ See [Model Cache & Offline Use](concepts/model-cache-offline.md) for full detail
 - **AdaFace**: [AdaFace: Quality Adaptive Margin for Face Recognition](https://arxiv.org/abs/2204.00964)
 - **ArcFace**: [Additive Angular Margin Loss for Deep Face Recognition](https://arxiv.org/abs/1801.07698)
 - **SphereFace**: [Deep Hypersphere Embedding for Face Recognition](https://arxiv.org/abs/1704.08063)
+- **MODNet**: [Real-Time Trimap-Free Portrait Matting via Objective Decomposition](https://arxiv.org/abs/2011.11961)
 - **BiSeNet**: [Bilateral Segmentation Network for Real-time Semantic Segmentation](https://arxiv.org/abs/1808.00897)
--- a/docs/modules/attributes.md
+++ b/docs/modules/attributes.md
@@ -2,6 +2,11 @@

 Facial attribute analysis for age, gender, race, and emotion detection.

+<figure markdown="span">
+  ![Age & Gender Prediction](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/age_gender.jpg){ width="100%" }
+  <figcaption>Age and gender prediction with detection bounding boxes</figcaption>
+</figure>
+
 ---

 ## Available Models
--- a/docs/modules/detection.md
+++ b/docs/modules/detection.md
@@ -2,6 +2,11 @@

 Face detection is the first step in any face analysis pipeline. UniFace provides four detection models.

+<figure markdown="span">
+  ![Face Detection](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/detection.jpg){ width="100%" }
+  <figcaption>SCRFD detection with corner-style bounding boxes and 5-point landmarks</figcaption>
+</figure>
+
 ---

 ## Available Models
--- a/docs/modules/gaze.md
+++ b/docs/modules/gaze.md
@@ -2,6 +2,11 @@

 Gaze estimation predicts where a person is looking (pitch and yaw angles).

+<figure markdown="span">
+  ![Gaze Estimation](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/gaze.jpg){ width="100%" }
+  <figcaption>Gaze direction arrows with pitch/yaw angle labels</figcaption>
+</figure>
+
 ---

 ## Available Models
--- a/docs/modules/headpose.md
+++ b/docs/modules/headpose.md
@@ -2,6 +2,11 @@

 Head pose estimation predicts the 3D orientation of a person's head (pitch, yaw, and roll angles).

+<figure markdown="span">
+  ![Head Pose Estimation](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/headpose.jpg){ width="100%" }
+  <figcaption>3D head pose visualization with pitch, yaw, and roll angles</figcaption>
+</figure>
+
 ---

 ## Available Models
--- a/docs/modules/landmarks.md
+++ b/docs/modules/landmarks.md
@@ -2,6 +2,11 @@

 Facial landmark detection provides precise localization of facial features.

+<figure markdown="span">
+  ![106-Point Landmarks](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/landmarks.jpg){ width="50%" }
+  <figcaption>106-point facial landmark localization</figcaption>
+</figure>
+
 ---

 ## Available Models
--- a/docs/modules/matting.md
+++ b/docs/modules/matting.md
@@ -0,0 +1,157 @@
+# Portrait Matting
+
+Portrait matting produces a soft alpha matte separating the foreground (person) from the background — no trimap needed.
+
+<figure markdown="span">
+  ![Portrait Matting](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/matting.jpg){ width="100%" }
+  <figcaption>MODNet: Input → Matte → Green Screen</figcaption>
+</figure>
+
+---
+
+## Available Models
+
+| Model | Variant | Size | Use Case |
+|-------|---------|------|----------|
+| **MODNet Photographic** :material-check-circle: | PHOTOGRAPHIC | 25 MB | High-quality portrait photos |
+| MODNet Webcam | WEBCAM | 25 MB | Real-time webcam feeds |
+
+---
+
+## Basic Usage
+
+```python
+import cv2
+from uniface.matting import MODNet
+
+matting = MODNet()
+
+image = cv2.imread("photo.jpg")
+matte = matting.predict(image)
+
+print(f"Matte shape: {matte.shape}")   # (H, W)
+print(f"Matte dtype: {matte.dtype}")   # float32
+print(f"Matte range: [{matte.min():.2f}, {matte.max():.2f}]")  # [0, 1]
+```
+
+---
+
+## Model Variants
+
+```python
+from uniface.matting import MODNet
+from uniface.constants import MODNetWeights
+
+# Photographic (default) — best for photos
+matting = MODNet()
+
+# Webcam — optimized for real-time
+matting = MODNet(model_name=MODNetWeights.WEBCAM)
+
+# Custom input size
+matting = MODNet(input_size=256)
+```
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `model_name` | `PHOTOGRAPHIC` | Model variant to load |
+| `input_size` | `512` | Target shorter-side size for preprocessing |
+| `providers` | `None` | ONNX Runtime execution providers |
+
+---
+
+## Applications
+
+### Transparent Background (RGBA)
+
+```python
+import cv2
+import numpy as np
+
+matting = MODNet()
+image = cv2.imread("photo.jpg")
+matte = matting.predict(image)
+
+rgba = cv2.cvtColor(image, cv2.COLOR_BGR2BGRA)
+rgba[:, :, 3] = (matte * 255).astype(np.uint8)
+cv2.imwrite("transparent.png", rgba)
+```
+
+### Green Screen
+
+```python
+import numpy as np
+
+matte_3ch = matte[:, :, np.newaxis]
+bg = np.full_like(image, (0, 177, 64), dtype=np.uint8)
+green = (image * matte_3ch + bg * (1 - matte_3ch)).astype(np.uint8)
+cv2.imwrite("green_screen.jpg", green)
+```
+
+### Custom Background
+
+```python
+import cv2
+import numpy as np
+
+background = cv2.imread("beach.jpg")
+background = cv2.resize(background, (image.shape[1], image.shape[0]))
+
+matte_3ch = matte[:, :, np.newaxis]
+result = (image * matte_3ch + background * (1 - matte_3ch)).astype(np.uint8)
+cv2.imwrite("custom_bg.jpg", result)
+```
+
+### Webcam Matting
+
+```python
+import cv2
+import numpy as np
+from uniface.matting import MODNet
+
+matting = MODNet(model_name="modnet_webcam")
+cap = cv2.VideoCapture(0)
+
+while True:
+    ret, frame = cap.read()
+    if not ret:
+        break
+
+    matte = matting.predict(frame)
+    matte_3ch = matte[:, :, np.newaxis]
+    bg = np.full_like(frame, (0, 177, 64), dtype=np.uint8)
+    result = (frame * matte_3ch + bg * (1 - matte_3ch)).astype(np.uint8)
+
+    cv2.imshow("Matting", np.hstack([frame, result]))
+    if cv2.waitKey(1) & 0xFF == ord("q"):
+        break
+
+cap.release()
+cv2.destroyAllWindows()
+```
+
+---
+
+## Factory Function
+
+```python
+from uniface.matting import create_matting_model
+from uniface.constants import MODNetWeights
+
+# Default (Photographic)
+matting = create_matting_model()
+
+# With enum
+matting = create_matting_model(MODNetWeights.WEBCAM)
+
+# With string
+matting = create_matting_model("modnet_webcam")
+```
+
+---
+
+## Next Steps
+
+- [Parsing](parsing.md) - Face semantic segmentation
+- [Privacy](privacy.md) - Face anonymization
+- [Detection](detection.md) - Face detection
--- a/docs/modules/parsing.md
+++ b/docs/modules/parsing.md
@@ -2,6 +2,16 @@

 Face parsing segments faces into semantic components or face regions.

+<figure markdown="span">
+  ![Face Parsing](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/parsing.jpg){ width="80%" }
+  <figcaption>BiSeNet face parsing with 19 semantic component classes</figcaption>
+</figure>
+
+<figure markdown="span">
+  ![Face Segmentation](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/segmentation.jpg){ width="80%" }
+  <figcaption>XSeg face region segmentation mask</figcaption>
+</figure>
+
 ---

 ## Available Models
--- a/docs/modules/privacy.md
+++ b/docs/modules/privacy.md
@@ -2,6 +2,11 @@

 Face anonymization protects privacy by blurring or obscuring faces in images and videos.

+<figure markdown="span">
+  ![Face Anonymization](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/anonymization.jpg){ width="100%" }
+  <figcaption>Five anonymization methods: pixelate, gaussian, blackout, elliptical, and median</figcaption>
+</figure>
+
 ---

 ## Available Methods
--- a/docs/modules/recognition.md
+++ b/docs/modules/recognition.md
@@ -2,6 +2,11 @@

 Face recognition extracts embeddings for identity verification and face search.

+<figure markdown="span">
+  ![Face Verification](https://raw.githubusercontent.com/yakhyo/uniface/main/assets/demos/verification.jpg){ width="80%" }
+  <figcaption>Pairwise face verification with cosine similarity scores</figcaption>
+</figure>
+
 ---

 ## Available Models
@@ -10,6 +15,7 @@ Face recognition extracts embeddings for identity verification and face search.
 |-------|----------|------|---------------|
 | **AdaFace** | IR-18/IR-101 | 92-249 MB | 512 |
 | **ArcFace** | MobileNet/ResNet | 8-166 MB | 512 |
+| **EdgeFace** | EdgeNeXt/LoRA | 5-70 MB | 512 |
 | **MobileFace** | MobileNet V2/V3 | 1-10 MB | 512 |
 | **SphereFace** | Sphere20/36 | 50-92 MB | 512 |

@@ -113,6 +119,64 @@ recognizer = ArcFace(providers=['CPUExecutionProvider'])

 ---

+## EdgeFace
+
+Efficient face recognition designed for edge devices, using an EdgeNeXt backbone with optional LoRA low-rank compression. Competition-winning entry (compact track) at EFaR 2023, IJCB.
+
+### Basic Usage
+
+```python
+from uniface.detection import RetinaFace
+from uniface.recognition import EdgeFace
+
+detector = RetinaFace()
+recognizer = EdgeFace()
+
+# Detect face
+faces = detector.detect(image)
+
+# Extract embedding
+if faces:
+    embedding = recognizer.get_normalized_embedding(image, faces[0].landmarks)
+    print(f"Embedding shape: {embedding.shape}")  # (512,)
+```
+
+### Model Variants
+
+```python
+from uniface.recognition import EdgeFace
+from uniface.constants import EdgeFaceWeights
+
+# Ultra-compact (default)
+recognizer = EdgeFace(model_name=EdgeFaceWeights.XXS)
+
+# Compact with LoRA
+recognizer = EdgeFace(model_name=EdgeFaceWeights.XS_GAMMA_06)
+
+# Small with LoRA
+recognizer = EdgeFace(model_name=EdgeFaceWeights.S_GAMMA_05)
+
+# Full-size
+recognizer = EdgeFace(model_name=EdgeFaceWeights.BASE)
+
+# Force CPU execution
+recognizer = EdgeFace(providers=['CPUExecutionProvider'])
+```
+
+| Variant | Params | MFLOPs | Size | LFW | CALFW | CPLFW | CFP-FP | AgeDB-30 |
+|---------|--------|--------|------|-----|-------|-------|--------|----------|
+| **XXS** :material-check-circle: | 1.24M | 94 | ~5 MB | 99.57% | 94.83% | 90.27% | 93.63% | 94.92% |
+| XS_GAMMA_06 | 1.77M | 154 | ~7 MB | 99.73% | 95.28% | 91.58% | 94.71% | 96.08% |
+| S_GAMMA_05 | 3.65M | 306 | ~14 MB | 99.78% | 95.55% | 92.48% | 95.74% | 97.03% |
+| BASE | 18.2M | 1399 | ~70 MB | 99.83% | 96.07% | 93.75% | 97.01% | 97.60% |
+
+!!! info "Reference"
+    **Paper**: [EdgeFace: Efficient Face Recognition Model for Edge Devices](https://arxiv.org/abs/2307.01838v2) (IEEE T-BIOM 2024)
+
+    **Source**: [github.com/otroshi/edgeface](https://github.com/otroshi/edgeface)
+
+---
+
 ## MobileFace

 Lightweight face recognition models with MobileNet backbones.
@@ -287,9 +351,10 @@ else:
 ```python
 from uniface.recognition import create_recognizer

-# Available methods: 'arcface', 'adaface', 'mobileface', 'sphereface'
+# Available methods: 'arcface', 'adaface', 'edgeface', 'mobileface', 'sphereface'
 recognizer = create_recognizer('arcface')
 recognizer = create_recognizer('adaface')
+recognizer = create_recognizer('edgeface')
 ```

 ---
--- a/docs/notebooks.md
+++ b/docs/notebooks.md
@@ -20,6 +20,7 @@ Run UniFace examples directly in your browser with Google Colab, or download and
 | [Face Vector Store](https://github.com/yakhyo/uniface/blob/main/examples/10_face_vector_store.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/10_face_vector_store.ipynb) | FAISS-backed face database |
 | [Head Pose Estimation](https://github.com/yakhyo/uniface/blob/main/examples/11_head_pose_estimation.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/11_head_pose_estimation.ipynb) | 3D head orientation estimation |
 | [Face Recognition](https://github.com/yakhyo/uniface/blob/main/examples/12_face_recognition.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/12_face_recognition.ipynb) | Standalone face recognition pipeline |
+| [Portrait Matting](https://github.com/yakhyo/uniface/blob/main/examples/13_portrait_matting.ipynb) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yakhyo/uniface/blob/main/examples/13_portrait_matting.ipynb) | Portrait matting with MODNet |

 ---

--- a/docs/quickstart.md
+++ b/docs/quickstart.md
@@ -280,6 +280,34 @@ print(f"Detected {len(np.unique(mask))} facial components")

 ---

+## Portrait Matting
+
+Remove backgrounds without a trimap:
+
+```python
+import cv2
+import numpy as np
+from uniface.matting import MODNet
+
+matting = MODNet()
+
+image = cv2.imread("portrait.jpg")
+matte = matting.predict(image)  # (H, W) float32 in [0, 1]
+
+# Transparent PNG
+rgba = cv2.cvtColor(image, cv2.COLOR_BGR2BGRA)
+rgba[:, :, 3] = (matte * 255).astype(np.uint8)
+cv2.imwrite("transparent.png", rgba)
+
+# Green screen
+matte_3ch = matte[:, :, np.newaxis]
+bg = np.full_like(image, (0, 177, 64), dtype=np.uint8)
+result = (image * matte_3ch + bg * (1 - matte_3ch)).astype(np.uint8)
+cv2.imwrite("green_screen.jpg", result)
+```
+
+---
+
 ## Face Anonymization

 Blur faces for privacy protection:
--- a/examples/01_face_detection.ipynb
+++ b/examples/01_face_detection.ipynb
--- a/examples/02_face_alignment.ipynb
+++ b/examples/02_face_alignment.ipynb
--- a/examples/03_face_verification.ipynb
+++ b/examples/03_face_verification.ipynb
--- a/examples/04_face_search.ipynb
+++ b/examples/04_face_search.ipynb
--- a/examples/05_face_analyzer.ipynb
+++ b/examples/05_face_analyzer.ipynb
--- a/examples/06_face_parsing.ipynb
+++ b/examples/06_face_parsing.ipynb
--- a/examples/07_face_anonymization.ipynb
+++ b/examples/07_face_anonymization.ipynb
--- a/examples/08_gaze_estimation.ipynb
+++ b/examples/08_gaze_estimation.ipynb
--- a/examples/09_face_segmentation.ipynb
+++ b/examples/09_face_segmentation.ipynb
--- a/examples/10_face_vector_store.ipynb
+++ b/examples/10_face_vector_store.ipynb
--- a/examples/11_head_pose_estimation.ipynb
+++ b/examples/11_head_pose_estimation.ipynb
--- a/examples/12_face_recognition.ipynb
+++ b/examples/12_face_recognition.ipynb
--- a/examples/13_portrait_matting.ipynb
+++ b/examples/13_portrait_matting.ipynb
@@ -0,0 +1,265 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Portrait Matting with MODNet\n",
+    "\n",
+    "<div style=\"display:flex; flex-wrap:wrap; align-items:center;\">\n",
+    "  <a style=\"margin-right:10px; margin-bottom:6px;\" href=\"https://pepy.tech/projects/uniface\"><img alt=\"PyPI Downloads\" src=\"https://static.pepy.tech/personalized-badge/uniface?period=total&units=international_system&left_color=grey&right_color=blue&left_text=Downloads\"></a>\n",
+    "  <a style=\"margin-right:10px; margin-bottom:6px;\" href=\"https://pypi.org/project/uniface/\"><img alt=\"PyPI Version\" src=\"https://img.shields.io/pypi/v/uniface.svg\"></a>\n",
+    "  <a style=\"margin-right:10px; margin-bottom:6px;\" href=\"https://opensource.org/licenses/MIT\"><img alt=\"License\" src=\"https://img.shields.io/badge/License-MIT-blue.svg\"></a>\n",
+    "  <a style=\"margin-bottom:6px;\" href=\"https://github.com/yakhyo/uniface\"><img alt=\"GitHub Stars\" src=\"https://img.shields.io/github/stars/yakhyo/uniface.svg?style=social\"></a>\n",
+    "</div>\n",
+    "\n",
+    "**UniFace** is a lightweight, production-ready, all-in-one face analysis library built on ONNX Runtime.\n",
+    "\n",
+    "🔗 **GitHub**: [github.com/yakhyo/uniface](https://github.com/yakhyo/uniface) | 📚 **Docs**: [yakhyo.github.io/uniface](https://yakhyo.github.io/uniface)\n",
+    "\n",
+    "---\n",
+    "\n",
+    "This notebook demonstrates portrait matting using **MODNet** — a trimap-free model that produces soft alpha mattes from full images. No face detection or cropping required.\n",
+    "\n",
+    "## 1. Install UniFace"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%pip install -q uniface\n",
+    "\n",
+    "# Clone repo for assets (Colab only)\n",
+    "import os\n",
+    "if 'COLAB_GPU' in os.environ or 'COLAB_RELEASE_TAG' in os.environ:\n",
+    "    if not os.path.exists('uniface'):\n",
+    "        !git clone --depth 1 https://github.com/yakhyo/uniface.git\n",
+    "    os.chdir('uniface/examples')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 2. Import Libraries"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import cv2\n",
+    "import numpy as np\n",
+    "import matplotlib.pyplot as plt\n",
+    "\n",
+    "import uniface\n",
+    "from uniface.matting import MODNet\n",
+    "\n",
+    "print(f\"UniFace version: {uniface.__version__}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 3. Initialize Model\n",
+    "\n",
+    "MODNet has two variants:\n",
+    "- **PHOTOGRAPHIC** (default): optimized for high-quality portrait photos\n",
+    "- **WEBCAM**: optimized for real-time webcam feeds"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "matting = MODNet()\n",
+    "\n",
+    "print(f\"Input size: {matting.input_size}\")\n",
+    "print(f\"Input name: {matting.input_name}\")\n",
+    "print(f\"Output names: {matting.output_names}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 4. Helper Functions"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def compose(image, matte, background=None):\n",
+    "    \"\"\"Composite foreground over a background using the alpha matte.\"\"\"\n",
+    "    h, w = image.shape[:2]\n",
+    "    matte_3ch = matte[:, :, np.newaxis]\n",
+    "\n",
+    "    if background is None:\n",
+    "        bg = np.full_like(image, (0, 177, 64), dtype=np.uint8)\n",
+    "    else:\n",
+    "        bg = cv2.resize(background, (w, h), interpolation=cv2.INTER_AREA)\n",
+    "\n",
+    "    return (image * matte_3ch + bg * (1 - matte_3ch)).astype(np.uint8)\n",
+    "\n",
+    "\n",
+    "def show_results(image, matte):\n",
+    "    \"\"\"Display original, matte, and green screen as a single merged image.\"\"\"\n",
+    "    matte_vis = cv2.cvtColor((matte * 255).astype(np.uint8), cv2.COLOR_GRAY2BGR)\n",
+    "    green = compose(image, matte)\n",
+    "    merged = np.hstack([image, matte_vis, green])\n",
+    "\n",
+    "    plt.figure(figsize=(18, 6))\n",
+    "    plt.imshow(cv2.cvtColor(merged, cv2.COLOR_BGR2RGB))\n",
+    "    plt.axis(\"off\")\n",
+    "    plt.tight_layout()\n",
+    "    plt.show()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 5. Basic Matting"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "image = cv2.imread(\"../assets/demos/src_portrait1.jpg\")\n",
+    "print(f\"Image shape: {image.shape}\")\n",
+    "\n",
+    "matte = matting.predict(image)\n",
+    "print(f\"Matte shape: {matte.shape}\")\n",
+    "print(f\"Matte dtype: {matte.dtype}\")\n",
+    "print(f\"Matte range: [{matte.min():.3f}, {matte.max():.3f}]\")\n",
+    "\n",
+    "show_results(image, matte)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 6. Transparent Background (RGBA)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "alpha = (matte * 255).astype(np.uint8)\n",
+    "rgba = cv2.cvtColor(image, cv2.COLOR_BGR2BGRA)\n",
+    "rgba[:, :, 3] = alpha\n",
+    "\n",
+    "# Checkerboard background to visualize transparency\n",
+    "h, w = image.shape[:2]\n",
+    "checker = np.zeros((h, w, 3), dtype=np.uint8)\n",
+    "block = 20\n",
+    "for y in range(0, h, block):\n",
+    "    for x in range(0, w, block):\n",
+    "        if (y // block + x // block) % 2 == 0:\n",
+    "            checker[y:y+block, x:x+block] = 200\n",
+    "        else:\n",
+    "            checker[y:y+block, x:x+block] = 255\n",
+    "\n",
+    "matte_3ch = matte[:, :, np.newaxis]\n",
+    "rgba_vis = (image * matte_3ch + checker * (1 - matte_3ch)).astype(np.uint8)\n",
+    "\n",
+    "merged = np.hstack([image, rgba_vis])\n",
+    "\n",
+    "plt.figure(figsize=(16, 5))\n",
+    "plt.imshow(cv2.cvtColor(merged, cv2.COLOR_BGR2RGB))\n",
+    "plt.axis(\"off\")\n",
+    "plt.tight_layout()\n",
+    "plt.show()\n",
+    "\n",
+    "print(f\"RGBA shape: {rgba.shape}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 7. Custom Background"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Create a gradient background\n",
+    "h, w = image.shape[:2]\n",
+    "gradient = np.zeros((h, w, 3), dtype=np.uint8)\n",
+    "for y in range(h):\n",
+    "    ratio = y / h\n",
+    "    gradient[y, :] = [int(180 * (1 - ratio)), int(100 + 80 * ratio), int(220 * ratio)]\n",
+    "\n",
+    "custom_bg = compose(image, matte, gradient)\n",
+    "green_bg = compose(image, matte)\n",
+    "\n",
+    "merged = np.hstack([image, green_bg, custom_bg])\n",
+    "\n",
+    "plt.figure(figsize=(18, 6))\n",
+    "plt.imshow(cv2.cvtColor(merged, cv2.COLOR_BGR2RGB))\n",
+    "plt.axis(\"off\")\n",
+    "plt.tight_layout()\n",
+    "plt.show()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Summary\n",
+    "\n",
+    "MODNet provides trimap-free portrait matting:\n",
+    "\n",
+    "- **`predict(image)`** — returns `(H, W)` float32 alpha matte in `[0, 1]`\n",
+    "- **No face detection needed** — works on full images directly\n",
+    "- **Two variants** — `PHOTOGRAPHIC` for photos, `WEBCAM` for real-time\n",
+    "- **Compositing** — use the matte for transparent PNGs, green screen, or custom backgrounds\n",
+    "\n",
+    "For more details, see the [Matting docs](https://yakhyo.github.io/uniface/modules/matting/)."
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "base",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.13.5"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}
--- a/mkdocs.yml
+++ b/mkdocs.yml
@@ -150,6 +150,7 @@ nav:
      - Landmarks: modules/landmarks.md
      - Attributes: modules/attributes.md
      - Parsing: modules/parsing.md
+      - Matting: modules/matting.md
      - Gaze: modules/gaze.md
      - Head Pose: modules/headpose.md
      - Anti-Spoofing: modules/spoofing.md
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "uniface"
-version = "3.3.0"
+version = "3.5.0"
 description = "UniFace: A Comprehensive Library for Face Detection, Recognition, Tracking, Landmark Analysis, Face Parsing, Gaze Estimation, Age, and Gender Detection"
 readme = "README.md"
 license = "MIT"
--- a/tests/test_factory.py
+++ b/tests/test_factory.py
@@ -91,6 +91,12 @@ def test_create_recognizer_sphereface():
    assert recognizer is not None, 'Failed to create SphereFace recognizer'


+def test_create_recognizer_edgeface():
+    """Test creating an EdgeFace recognizer using factory function."""
+    recognizer = create_recognizer('edgeface')
+    assert recognizer is not None, 'Failed to create EdgeFace recognizer'
+
+
 def test_create_recognizer_invalid_method():
    """
    Test that invalid recognizer method raises an error.
--- a/tests/test_matting.py
+++ b/tests/test_matting.py
@@ -0,0 +1,158 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+
+from __future__ import annotations
+
+import numpy as np
+import pytest
+
+from uniface.constants import MODNetWeights
+from uniface.matting import MODNet, create_matting_model
+
+
+def test_modnet_initialization():
+    """Test MODNet initialization with default weights."""
+    matting = MODNet()
+    assert matting is not None
+    assert matting.input_size == 512
+
+
+def test_modnet_with_webcam_weights():
+    """Test MODNet initialization with webcam variant."""
+    matting = MODNet(model_name=MODNetWeights.WEBCAM)
+    assert matting is not None
+    assert matting.input_size == 512
+
+
+def test_modnet_custom_input_size():
+    """Test MODNet with custom input size."""
+    matting = MODNet(input_size=256)
+    assert matting.input_size == 256
+
+
+def test_modnet_preprocess():
+    """Test preprocessing produces correct tensor shape and dtype."""
+    matting = MODNet()
+
+    image = np.random.randint(0, 255, (480, 640, 3), dtype=np.uint8)
+    tensor, orig_h, orig_w = matting.preprocess(image)
+
+    assert tensor.dtype == np.float32
+    assert tensor.ndim == 4
+    assert tensor.shape[0] == 1
+    assert tensor.shape[1] == 3
+    assert tensor.shape[2] % 32 == 0
+    assert tensor.shape[3] % 32 == 0
+    assert orig_h == 480
+    assert orig_w == 640
+
+
+def test_modnet_preprocess_small_image():
+    """Test preprocessing with image smaller than input_size."""
+    matting = MODNet(input_size=512)
+
+    image = np.random.randint(0, 255, (128, 128, 3), dtype=np.uint8)
+    tensor, orig_h, orig_w = matting.preprocess(image)
+
+    assert tensor.shape[2] % 32 == 0
+    assert tensor.shape[3] % 32 == 0
+    assert orig_h == 128
+    assert orig_w == 128
+
+
+def test_modnet_preprocess_large_image():
+    """Test preprocessing with image larger than input_size."""
+    matting = MODNet(input_size=512)
+
+    image = np.random.randint(0, 255, (1080, 1920, 3), dtype=np.uint8)
+    tensor, orig_h, orig_w = matting.preprocess(image)
+
+    assert tensor.shape[2] % 32 == 0
+    assert tensor.shape[3] % 32 == 0
+    assert orig_h == 1080
+    assert orig_w == 1920
+
+
+def test_modnet_postprocess():
+    """Test postprocessing resizes matte to original dimensions."""
+    matting = MODNet()
+
+    dummy_output = np.random.rand(1, 1, 512, 672).astype(np.float32)
+    matte = matting.postprocess(dummy_output, original_size=(640, 480))
+
+    assert matte.shape == (480, 640)
+    assert matte.dtype == np.float32
+
+
+def test_modnet_predict():
+    """Test end-to-end prediction."""
+    matting = MODNet()
+
+    image = np.random.randint(0, 255, (480, 640, 3), dtype=np.uint8)
+    matte = matting.predict(image)
+
+    assert matte.shape == (480, 640)
+    assert matte.dtype == np.float32
+    assert matte.min() >= 0.0
+    assert matte.max() <= 1.0
+
+
+def test_modnet_callable():
+    """Test that MODNet is callable via __call__."""
+    matting = MODNet()
+    image = np.random.randint(0, 255, (256, 256, 3), dtype=np.uint8)
+
+    matte = matting(image)
+
+    assert matte.shape == (256, 256)
+    assert matte.dtype == np.float32
+
+
+def test_modnet_different_input_sizes():
+    """Test prediction with various image dimensions."""
+    matting = MODNet()
+
+    sizes = [(256, 256), (480, 640), (720, 1280), (300, 500)]
+
+    for h, w in sizes:
+        image = np.random.randint(0, 255, (h, w, 3), dtype=np.uint8)
+        matte = matting.predict(image)
+
+        assert matte.shape == (h, w), f'Failed for size {h}x{w}'
+        assert matte.dtype == np.float32
+
+
+# Factory tests
+
+
+def test_create_matting_model_default():
+    """Test factory with default parameters."""
+    matting = create_matting_model()
+    assert matting is not None
+    assert isinstance(matting, MODNet)
+
+
+def test_create_matting_model_with_enum():
+    """Test factory with enum."""
+    matting = create_matting_model(MODNetWeights.WEBCAM)
+    assert isinstance(matting, MODNet)
+
+
+def test_create_matting_model_with_string():
+    """Test factory with string model name."""
+    matting = create_matting_model('modnet_photographic')
+    assert isinstance(matting, MODNet)
+
+
+def test_create_matting_model_webcam_string():
+    """Test factory with webcam string model name."""
+    matting = create_matting_model('modnet_webcam')
+    assert isinstance(matting, MODNet)
+
+
+def test_create_matting_model_invalid():
+    """Test factory with invalid model name."""
+    with pytest.raises(ValueError, match='Unknown matting model'):
+        create_matting_model('invalid_model')
--- a/tests/test_recognition.py
+++ b/tests/test_recognition.py
@@ -8,7 +8,7 @@ from __future__ import annotations
 import numpy as np
 import pytest

-from uniface.recognition import ArcFace, MobileFace, SphereFace
+from uniface.recognition import ArcFace, EdgeFace, MobileFace, SphereFace


@pytest.fixture
@@ -35,6 +35,12 @@ def sphereface_model():
    return SphereFace()


+@pytest.fixture
+def edgeface_model():
+    """Fixture to initialize the EdgeFace model for testing."""
+    return EdgeFace()
+
+
@pytest.fixture
 def mock_aligned_face():
    """
@@ -176,6 +182,45 @@ def test_sphereface_normalized_embedding(sphereface_model, mock_landmarks):
    assert np.isclose(norm, 1.0, atol=1e-5), f'Normalized embedding should have norm 1.0, got {norm}'


+# EdgeFace Tests
+def test_edgeface_initialization(edgeface_model):
+    """Test that the EdgeFace model initializes correctly."""
+    assert edgeface_model is not None, 'EdgeFace model initialization failed.'
+
+
+def test_edgeface_embedding_shape(edgeface_model, mock_aligned_face):
+    """Test that EdgeFace produces embeddings with the correct shape."""
+    embedding = edgeface_model.get_embedding(mock_aligned_face)
+
+    assert embedding.shape[1] == 512, f'Expected 512-dim embedding, got {embedding.shape[1]}'
+    assert embedding.shape[0] == 1, 'Embedding should have batch dimension of 1'
+
+
+def test_edgeface_normalized_embedding(edgeface_model, mock_landmarks):
+    """Test that EdgeFace normalized embeddings have unit length."""
+    mock_image = np.random.randint(0, 255, (640, 640, 3), dtype=np.uint8)
+
+    embedding = edgeface_model.get_normalized_embedding(mock_image, mock_landmarks)
+
+    assert embedding.shape == (512,), f'Expected shape (512,), got {embedding.shape}'
+    norm = np.linalg.norm(embedding)
+    assert np.isclose(norm, 1.0, atol=1e-5), f'Normalized embedding should have norm 1.0, got {norm}'
+
+
+def test_edgeface_embedding_dtype(edgeface_model, mock_aligned_face):
+    """Test that EdgeFace embeddings have the correct data type."""
+    embedding = edgeface_model.get_embedding(mock_aligned_face)
+    assert embedding.dtype == np.float32, f'Expected float32, got {embedding.dtype}'
+
+
+def test_edgeface_consistency(edgeface_model, mock_aligned_face):
+    """Test that the same input produces the same EdgeFace embedding."""
+    embedding1 = edgeface_model.get_embedding(mock_aligned_face)
+    embedding2 = edgeface_model.get_embedding(mock_aligned_face)
+
+    assert np.allclose(embedding1, embedding2), 'Same input should produce same embedding'
+
+
 # Cross-model comparison tests
 def test_different_models_different_embeddings(arcface_model, mobileface_model, mock_aligned_face):
    """
--- a/tools/download_model.py
+++ b/tools/download_model.py
@@ -1,9 +1,11 @@
 import argparse

 from uniface.constants import (
+    AdaFaceWeights,
    AgeGenderWeights,
    ArcFaceWeights,
    DDAMFNWeights,
+    EdgeFaceWeights,
    HeadPoseWeights,
    LandmarkWeights,
    MobileFaceWeights,
@@ -15,9 +17,11 @@ from uniface.model_store import verify_model_weights

 MODEL_TYPES = {
    'retinaface': RetinaFaceWeights,
-    'sphereface': SphereFaceWeights,
-    'mobileface': MobileFaceWeights,
+    'adaface': AdaFaceWeights,
    'arcface': ArcFaceWeights,
+    'edgeface': EdgeFaceWeights,
+    'mobileface': MobileFaceWeights,
+    'sphereface': SphereFaceWeights,
    'scrfd': SCRFDWeights,
    'ddamfn': DDAMFNWeights,
    'agegender': AgeGenderWeights,
--- a/tools/recognize.py
+++ b/tools/recognize.py
@@ -16,16 +16,22 @@ import numpy as np

 from uniface.detection import SCRFD, RetinaFace
 from uniface.face_utils import compute_similarity
-from uniface.recognition import ArcFace, MobileFace, SphereFace
+from uniface.recognition import AdaFace, ArcFace, EdgeFace, MobileFace, SphereFace
+
+RECOGNIZERS = {
+    'arcface': ArcFace,
+    'adaface': AdaFace,
+    'edgeface': EdgeFace,
+    'mobileface': MobileFace,
+    'sphereface': SphereFace,
+}


 def get_recognizer(name: str):
-    if name == 'arcface':
-        return ArcFace()
-    elif name == 'mobileface':
-        return MobileFace()
-    else:
-        return SphereFace()
+    cls = RECOGNIZERS.get(name)
+    if cls is None:
+        raise ValueError(f"Unknown recognizer: '{name}'. Available: {list(RECOGNIZERS)}")
+    return cls()


 def run_inference(detector, recognizer, image_path: str):
@@ -91,7 +97,7 @@ def main():
        '--recognizer',
        type=str,
        default='arcface',
-        choices=['arcface', 'mobileface', 'sphereface'],
+        choices=list(RECOGNIZERS),
    )
    args = parser.parse_args()

--- a/uniface/init.py
+++ b/uniface/init.py
@@ -15,10 +15,11 @@

 This library provides unified APIs for:
 - Face detection (RetinaFace, SCRFD, YOLOv5Face, YOLOv8Face)
- Face recognition (AdaFace, ArcFace, MobileFace, SphereFace)
+- Face recognition (AdaFace, ArcFace, EdgeFace, MobileFace, SphereFace)
 - Face tracking (ByteTrack with Kalman filtering)
 - Facial landmarks (106-point detection)
 - Face parsing (semantic segmentation)
+- Portrait matting (trimap-free alpha matte)
 - Gaze estimation
 - Head pose estimation
 - Age, gender, and emotion prediction
@@ -30,7 +31,7 @@ from __future__ import annotations

 __license__ = 'MIT'
 __author__ = 'Yakhyokhuja Valikhujaev'
-__version__ = '3.3.0'
+__version__ = '3.5.0'

 import contextlib

@@ -51,9 +52,10 @@ from .detection import (
 from .gaze import MobileGaze, create_gaze_estimator
 from .headpose import HeadPose, create_head_pose_estimator
 from .landmark import Landmark106, create_landmarker
+from .matting import MODNet, create_matting_model
 from .parsing import BiSeNet, XSeg, create_face_parser
 from .privacy import BlurFace
-from .recognition import AdaFace, ArcFace, MobileFace, SphereFace, create_recognizer
+from .recognition import AdaFace, ArcFace, EdgeFace, MobileFace, SphereFace, create_recognizer
 from .spoofing import MiniFASNet, create_spoofer
 from .tracking import BYTETracker
 from .types import AttributeResult, EmotionResult, Face, GazeResult, HeadPoseResult, SpoofingResult
@@ -74,6 +76,7 @@ __all__ = [
    'create_detector',
    'create_face_parser',
    'create_gaze_estimator',
+    'create_matting_model',
    'create_head_pose_estimator',
    'create_landmarker',
    'create_recognizer',
@@ -87,6 +90,7 @@ __all__ = [
    # Recognition models
    'AdaFace',
    'ArcFace',
+    'EdgeFace',
    'MobileFace',
    'SphereFace',
    # Landmark models
@@ -97,6 +101,8 @@ __all__ = [
    # Head pose models
    'HeadPose',
    'HeadPoseResult',
+    # Matting models
+    'MODNet',
    # Parsing models
    'BiSeNet',
    'XSeg',
--- a/uniface/constants.py
+++ b/uniface/constants.py
@@ -57,6 +57,18 @@ class AdaFaceWeights(str, Enum):
    IR_18  = "adaface_ir_18"
    IR_101 = "adaface_ir_101"

+class EdgeFaceWeights(str, Enum):
+    """
+    EdgeFace: Efficient Face Recognition Model for Edge Devices.
+    Based on EdgeNeXt backbone with optional LoRA low-rank compression.
+    All models output 512-D embeddings from 112x112 aligned face crops.
+    https://github.com/yakhyo/edgeface-onnx
+    """
+    XXS         = "edgeface_xxs"
+    XS_GAMMA_06 = "edgeface_xs_gamma_06"
+    S_GAMMA_05  = "edgeface_s_gamma_05"
+    BASE        = "edgeface_base"
+
 class RetinaFaceWeights(str, Enum):
    """
    Trained on WIDER FACE dataset.
@@ -189,6 +201,15 @@ class XSegWeights(str, Enum):
    DEFAULT = "xseg"


+class MODNetWeights(str, Enum):
+    """
+    MODNet: Real-Time Trimap-Free Portrait Matting via Objective Decomposition.
+    https://github.com/yakhyo/modnet
+    """
+    PHOTOGRAPHIC = "modnet_photographic"
+    WEBCAM       = "modnet_webcam"
+
+
 class MiniFASNetWeights(str, Enum):
    """
    MiniFASNet: Lightweight Face Anti-Spoofing models.
@@ -278,6 +299,24 @@ MODEL_REGISTRY: dict[Enum, ModelInfo] = {
        sha256='f2eb07d03de0af560a82e1214df799fec5e09375d43521e2868f9dc387e5a43e'
    ),

+    # EdgeFace
+    EdgeFaceWeights.XXS: ModelInfo(
+        url='https://github.com/yakhyo/edgeface-onnx/releases/download/weights/edgeface_xxs.onnx',
+        sha256='dc674de4cbc77fa0bf9a82d5149558ab8581d82a2cd3bb60f28fd1a5d3ff8a2f'
+    ),
+    EdgeFaceWeights.XS_GAMMA_06: ModelInfo(
+        url='https://github.com/yakhyo/edgeface-onnx/releases/download/weights/edgeface_xs_gamma_06.onnx',
+        sha256='9206e2eb13a2761d7b5b76e13016d4b9acd3fa3535a9a09939f3adacd139a5ff'
+    ),
+    EdgeFaceWeights.S_GAMMA_05: ModelInfo(
+        url='https://github.com/yakhyo/edgeface-onnx/releases/download/weights/edgeface_s_gamma_05.onnx',
+        sha256='b850767cf791bda585600b5c4c7d7432b2f998ccd862caae34ef1afa967d2e54'
+    ),
+    EdgeFaceWeights.BASE: ModelInfo(
+        url='https://github.com/yakhyo/edgeface-onnx/releases/download/weights/edgeface_base.onnx',
+        sha256='b56942f072c67385f44734b9458b0ccc4a2226888a113f77e0c802ad0c77b4c3'
+    ),
+
    # SCRFD
    SCRFDWeights.SCRFD_10G_KPS: ModelInfo(
        url='https://github.com/yakhyo/uniface/releases/download/weights/scrfd_10g_kps.onnx',
@@ -413,6 +452,16 @@ MODEL_REGISTRY: dict[Enum, ModelInfo] = {
        url='https://github.com/yakhyo/face-segmentation/releases/download/weights/xseg.onnx',
        sha256='0b57328efcb839d85973164b617ceee9dfe6cfcb2c82e8a033bba9f4f09b27e5'
    ),
+
+    # MODNet (Portrait Matting)
+    MODNetWeights.PHOTOGRAPHIC: ModelInfo(
+        url='https://github.com/yakhyo/modnet/releases/download/weights/modnet_photographic.onnx',
+        sha256='5069a5e306b9f5e9f4f2b0360264c9f8ea13b257c7c39943c7cf6a2ec3a102ae'
+    ),
+    MODNetWeights.WEBCAM: ModelInfo(
+        url='https://github.com/yakhyo/modnet/releases/download/weights/modnet_webcam.onnx',
+        sha256='de03cc16f3c91f25b7c2f0b42ea1a8d34f40a752234f3887572655e744e55306'
+    ),
 }


--- a/uniface/matting/init.py
+++ b/uniface/matting/init.py
@@ -0,0 +1,53 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from __future__ import annotations
+
+from uniface.constants import MODNetWeights
+
+from .base import BaseMatting
+from .modnet import MODNet
+
+__all__ = ['BaseMatting', 'MODNet', 'create_matting_model']
+
+
+def create_matting_model(
+    model_name: str | MODNetWeights = MODNetWeights.PHOTOGRAPHIC,
+    **kwargs,
+) -> BaseMatting:
+    """Factory function to create a portrait matting model.
+
+    Args:
+        model_name: Model to create. Options: ``MODNetWeights.PHOTOGRAPHIC``
+            (high-quality photos), ``MODNetWeights.WEBCAM`` (real-time webcam).
+            Also accepts string values like ``"modnet_photographic"`` or
+            ``"modnet_webcam"``.
+        **kwargs: Additional arguments passed to the model constructor
+            (e.g. ``input_size``, ``providers``).
+
+    Returns:
+        An instance of the requested matting model.
+
+    Raises:
+        ValueError: If the model_name is not recognized.
+
+    Example:
+        >>> matting = create_matting_model()
+        >>> matte = matting.predict(image)
+    """
+    if isinstance(model_name, MODNetWeights):
+        return MODNet(model_name=model_name, **kwargs)
+
+    if isinstance(model_name, str):
+        try:
+            weights = MODNetWeights(model_name)
+            return MODNet(model_name=weights, **kwargs)
+        except ValueError:
+            pass
+
+        valid_models = [m.value for m in MODNetWeights]
+        raise ValueError(f"Unknown matting model: '{model_name}'. Valid options are: {', '.join(valid_models)}")
+
+    valid_models = [m.value for m in MODNetWeights]
+    raise ValueError(f"Unknown matting model: '{model_name}'. Valid options are: {', '.join(valid_models)}")
--- a/uniface/matting/base.py
+++ b/uniface/matting/base.py
@@ -0,0 +1,88 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+
+import numpy as np
+
+
+class BaseMatting(ABC):
+    """Abstract base class for portrait matting models.
+
+    Unlike face parsers that operate on face crops and produce class labels or
+    face-region masks, matting models operate on full images and produce a soft
+    alpha matte (float32 in [0, 1]) separating foreground from background.
+
+    Subclasses must implement the full pipeline: model initialization,
+    preprocessing, postprocessing, and the ``predict`` entry point.
+    """
+
+    @abstractmethod
+    def _initialize_model(self) -> None:
+        """Initialize the underlying model for inference.
+
+        This method should handle loading model weights, creating the
+        inference session (e.g., ONNX Runtime), and any necessary
+        setup procedures to prepare the model for prediction.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        raise NotImplementedError('Subclasses must implement the _initialize_model method.')
+
+    @abstractmethod
+    def preprocess(self, image: np.ndarray) -> tuple[np.ndarray, int, int]:
+        """Preprocess the input image for model inference.
+
+        Args:
+            image: An image in BGR format with shape ``(H, W, 3)``.
+
+        Returns:
+            A tuple of ``(tensor, orig_h, orig_w)`` where *tensor* is the
+            preprocessed image ready for inference.
+        """
+        raise NotImplementedError('Subclasses must implement the preprocess method.')
+
+    @abstractmethod
+    def postprocess(self, outputs: np.ndarray, original_size: tuple[int, int]) -> np.ndarray:
+        """Postprocess raw model outputs into an alpha matte.
+
+        Args:
+            outputs: Raw outputs from the model inference.
+            original_size: Original image size as ``(width, height)``.
+
+        Returns:
+            Alpha matte with shape ``(H, W)`` and values in ``[0, 1]``.
+        """
+        raise NotImplementedError('Subclasses must implement the postprocess method.')
+
+    @abstractmethod
+    def predict(self, image: np.ndarray) -> np.ndarray:
+        """Run end-to-end matting on an image.
+
+        Args:
+            image: An image in BGR format with shape ``(H, W, 3)``.
+
+        Returns:
+            Alpha matte with shape ``(H, W)``, float32 in ``[0, 1]``.
+
+        Example:
+            >>> matting = create_matting_model()
+            >>> matte = matting.predict(image)
+            >>> print(f'Matte shape: {matte.shape}, dtype: {matte.dtype}')
+        """
+        raise NotImplementedError('Subclasses must implement the predict method.')
+
+    def __call__(self, image: np.ndarray) -> np.ndarray:
+        """Callable shortcut for :meth:`predict`.
+
+        Args:
+            image: An image in BGR format with shape ``(H, W, 3)``.
+
+        Returns:
+            Alpha matte with shape ``(H, W)``, float32 in ``[0, 1]``.
+        """
+        return self.predict(image)
--- a/uniface/matting/modnet.py
+++ b/uniface/matting/modnet.py
@@ -0,0 +1,162 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from __future__ import annotations
+
+import cv2
+import numpy as np
+
+from uniface.constants import MODNetWeights
+from uniface.log import Logger
+from uniface.model_store import verify_model_weights
+from uniface.onnx_utils import create_onnx_session
+
+from .base import BaseMatting
+
+__all__ = ['MODNet']
+
+STRIDE = 32
+
+
+class MODNet(BaseMatting):
+    """MODNet: Real-Time Trimap-Free Portrait Matting with ONNX Runtime.
+
+    MODNet produces a soft alpha matte from a full image without requiring
+    a trimap. It uses a MobileNetV2 backbone with low-resolution, high-resolution,
+    and fusion branches to generate accurate mattes at real-time speed.
+
+    Two pretrained variants are available:
+
+    - ``PHOTOGRAPHIC``: optimized for high-quality portrait photos.
+    - ``WEBCAM``: optimized for real-time webcam feeds.
+
+    Reference:
+        Ke et al., "MODNet: Real-Time Trimap-Free Portrait Matting via
+        Objective Decomposition", AAAI 2022.
+        https://github.com/yakhyo/modnet
+
+    Args:
+        model_name: The enum specifying the MODNet variant to load.
+            Defaults to ``MODNetWeights.PHOTOGRAPHIC``.
+        input_size: Target size for the shorter side during preprocessing.
+            The image is resized so its shorter side equals this value
+            (aspect ratio preserved), then both dimensions are floored to
+            multiples of 32. Defaults to 512.
+        providers: ONNX Runtime execution providers. If ``None``, auto-detects
+            the best available provider.
+
+    Attributes:
+        input_size (int): Target shorter-side size for preprocessing.
+
+    Example:
+        >>> from uniface.matting import MODNet
+        >>>
+        >>> matting = MODNet()
+        >>> matte = matting.predict(image)  # (H, W) float32 in [0, 1]
+        >>>
+        >>> # Composite onto green background
+        >>> import numpy as np
+        >>> bg = np.full_like(image, (0, 177, 64), dtype=np.uint8)
+        >>> alpha = matte[..., np.newaxis]
+        >>> result = (image * alpha + bg * (1 - alpha)).astype(np.uint8)
+    """
+
+    def __init__(
+        self,
+        model_name: MODNetWeights = MODNetWeights.PHOTOGRAPHIC,
+        input_size: int = 512,
+        providers: list[str] | None = None,
+    ) -> None:
+        Logger.info(f'Initializing MODNet with model={model_name}, input_size={input_size}')
+
+        self.input_size = input_size
+        self.providers = providers
+
+        self.model_path = verify_model_weights(model_name)
+        self._initialize_model()
+
+    def _initialize_model(self) -> None:
+        """Initialize the ONNX model from the stored model path.
+
+        Raises:
+            RuntimeError: If the model fails to load or initialize.
+        """
+        try:
+            self.session = create_onnx_session(self.model_path, providers=self.providers)
+
+            input_cfg = self.session.get_inputs()[0]
+            self.input_name = input_cfg.name
+
+            outputs = self.session.get_outputs()
+            self.output_names = [output.name for output in outputs]
+
+            Logger.info(f'MODNet initialized with input_size={self.input_size}')
+
+        except Exception as e:
+            Logger.error(f"Failed to load MODNet model from '{self.model_path}'", exc_info=True)
+            raise RuntimeError(f'Failed to initialize MODNet model: {e}') from e
+
+    def preprocess(self, image: np.ndarray) -> tuple[np.ndarray, int, int]:
+        """Preprocess a BGR image for MODNet inference.
+
+        The image is converted to RGB, resized so its shorter side matches
+        ``input_size`` (aspect ratio preserved), floored to multiples of 32,
+        and normalized to ``[-1, 1]``.
+
+        Args:
+            image: Input image in BGR format with shape ``(H, W, 3)``.
+
+        Returns:
+            A tuple of ``(tensor, orig_h, orig_w)`` where *tensor* has shape
+            ``(1, 3, H', W')`` in float32.
+        """
+        orig_h, orig_w = image.shape[:2]
+        rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        if max(orig_h, orig_w) < self.input_size or min(orig_h, orig_w) > self.input_size:
+            if orig_w >= orig_h:
+                new_h = self.input_size
+                new_w = int(orig_w / orig_h * self.input_size)
+            else:
+                new_w = self.input_size
+                new_h = int(orig_h / orig_w * self.input_size)
+        else:
+            new_h, new_w = orig_h, orig_w
+
+        new_h = new_h - (new_h % STRIDE)
+        new_w = new_w - (new_w % STRIDE)
+        rgb = cv2.resize(rgb, (new_w, new_h), interpolation=cv2.INTER_AREA)
+
+        x = rgb.astype(np.float32) / 255.0
+        x = (x - 0.5) / 0.5
+        x = np.transpose(x, (2, 0, 1))
+
+        return np.expand_dims(x, axis=0), orig_h, orig_w
+
+    def postprocess(self, outputs: np.ndarray, original_size: tuple[int, int]) -> np.ndarray:
+        """Postprocess raw model output into an alpha matte.
+
+        Args:
+            outputs: Raw ONNX output with shape ``(1, 1, H', W')``.
+            original_size: Target size as ``(width, height)``.
+
+        Returns:
+            Alpha matte with shape ``(H, W)``, float32 in ``[0, 1]``.
+        """
+        matte = outputs[0, 0]
+        matte = cv2.resize(matte, original_size, interpolation=cv2.INTER_AREA)
+        return matte
+
+    def predict(self, image: np.ndarray) -> np.ndarray:
+        """Run portrait matting on a BGR image.
+
+        Args:
+            image: Input image in BGR format with shape ``(H, W, 3)``.
+
+        Returns:
+            Alpha matte with shape ``(H, W)``, float32 in ``[0, 1]``.
+        """
+        tensor, orig_h, orig_w = self.preprocess(image)
+        outputs = self.session.run(self.output_names, {self.input_name: tensor})
+        return self.postprocess(outputs[0], (orig_w, orig_h))
--- a/uniface/recognition/init.py
+++ b/uniface/recognition/init.py
@@ -4,8 +4,11 @@


 from .adaface import AdaFace
+from .arcface import ArcFace
 from .base import BaseRecognizer
-from .models import ArcFace, MobileFace, SphereFace
+from .edgeface import EdgeFace
+from .mobileface import MobileFace
+from .sphereface import SphereFace


 def create_recognizer(method: str = 'arcface', **kwargs) -> BaseRecognizer:
@@ -18,7 +21,7 @@ def create_recognizer(method: str = 'arcface', **kwargs) -> BaseRecognizer:

    Args:
        method (str): The recognition method to use.
-            Options: 'arcface' (default), 'adaface', 'mobileface', 'sphereface'.
+            Options: 'arcface' (default), 'adaface', 'edgeface', 'mobileface', 'sphereface'.
        **kwargs: Model-specific parameters passed to the recognizer's constructor.
            For example, `model_name` can be used to select a specific
            pre-trained weight from the available enums (e.g., `ArcFaceWeights.MNET`).
@@ -43,6 +46,10 @@ def create_recognizer(method: str = 'arcface', **kwargs) -> BaseRecognizer:

        >>> # Create a SphereFace recognizer
        >>> recognizer = create_recognizer('sphereface')
+
+        >>> # Create an EdgeFace recognizer
+        >>> from uniface.constants import EdgeFaceWeights
+        >>> recognizer = create_recognizer('edgeface', model_name=EdgeFaceWeights.XXS)
    """
    method = method.lower()

@@ -50,13 +57,15 @@ def create_recognizer(method: str = 'arcface', **kwargs) -> BaseRecognizer:
        return ArcFace(**kwargs)
    elif method == 'adaface':
        return AdaFace(**kwargs)
+    elif method == 'edgeface':
+        return EdgeFace(**kwargs)
    elif method == 'mobileface':
        return MobileFace(**kwargs)
    elif method == 'sphereface':
        return SphereFace(**kwargs)
    else:
-        available = ['arcface', 'adaface', 'mobileface', 'sphereface']
+        available = ['arcface', 'adaface', 'edgeface', 'mobileface', 'sphereface']
        raise ValueError(f"Unsupported method: '{method}'. Available: {available}")


-__all__ = ['AdaFace', 'ArcFace', 'BaseRecognizer', 'MobileFace', 'SphereFace', 'create_recognizer']
+__all__ = ['AdaFace', 'ArcFace', 'BaseRecognizer', 'EdgeFace', 'MobileFace', 'SphereFace', 'create_recognizer']
--- a/uniface/recognition/arcface.py
+++ b/uniface/recognition/arcface.py
@@ -0,0 +1,49 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from __future__ import annotations
+
+from uniface.constants import ArcFaceWeights
+from uniface.model_store import verify_model_weights
+
+from .base import BaseRecognizer, PreprocessConfig
+
+__all__ = ['ArcFace']
+
+
+class ArcFace(BaseRecognizer):
+    """ArcFace model for robust face recognition.
+
+    This class provides a concrete implementation of the BaseRecognizer,
+    pre-configured for ArcFace models. It handles the loading of specific
+    ArcFace weights and sets up the appropriate default preprocessing.
+
+    Args:
+        model_name (ArcFaceWeights): The specific ArcFace model variant to use.
+            Defaults to `ArcFaceWeights.MNET`.
+        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
+            configuration. If None, a default config for ArcFace is used.
+        providers (list[str] | None): ONNX Runtime execution providers. If None, auto-detects
+            the best available provider. Example: ['CPUExecutionProvider'] to force CPU.
+
+    Example:
+        >>> from uniface.recognition import ArcFace
+        >>> recognizer = ArcFace()
+        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
+
+    Reference:
+        https://arxiv.org/abs/1801.07698
+        https://github.com/yakhyo/face-reidentification
+    """
+
+    def __init__(
+        self,
+        model_name: ArcFaceWeights = ArcFaceWeights.MNET,
+        preprocessing: PreprocessConfig | None = None,
+        providers: list[str] | None = None,
+    ) -> None:
+        if preprocessing is None:
+            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
+        model_path = verify_model_weights(model_name)
+        super().__init__(model_path=model_path, preprocessing=preprocessing, providers=providers)
--- a/uniface/recognition/edgeface.py
+++ b/uniface/recognition/edgeface.py
@@ -0,0 +1,57 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from __future__ import annotations
+
+from uniface.constants import EdgeFaceWeights
+from uniface.model_store import verify_model_weights
+
+from .base import BaseRecognizer, PreprocessConfig
+
+__all__ = ['EdgeFace']
+
+
+class EdgeFace(BaseRecognizer):
+    """EdgeFace: Efficient Face Recognition Model for Edge Devices.
+
+    EdgeFace uses an EdgeNeXt backbone with optional LoRA low-rank
+    compression, offering a strong accuracy-efficiency trade-off for
+    deployment on resource-constrained hardware. Competition-winning
+    entry (compact track) at EFaR 2023, IJCB.
+
+    All variants output 512-D embeddings from 112x112 aligned face crops.
+
+    Args:
+        model_name (EdgeFaceWeights): The specific EdgeFace model variant to use.
+            - XXS: Ultra-compact (1.24M params, ~5 MB)
+            - XS_GAMMA_06: Compact with LoRA (1.77M params, ~7 MB)
+            - S_GAMMA_05: Small with LoRA (3.65M params, ~14 MB)
+            - BASE: Full-size model (18.2M params, ~70 MB)
+            Defaults to `EdgeFaceWeights.XXS`.
+        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
+            configuration. If None, a default config for EdgeFace is used.
+        providers (list[str] | None): ONNX Runtime execution providers. If None, auto-detects
+            the best available provider. Example: ['CPUExecutionProvider'] to force CPU.
+
+    Example:
+        >>> from uniface.recognition import EdgeFace
+        >>> recognizer = EdgeFace()
+        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
+
+    Reference:
+        https://arxiv.org/abs/2307.01838v2
+        https://github.com/otroshi/edgeface
+        https://github.com/yakhyo/edgeface-onnx
+    """
+
+    def __init__(
+        self,
+        model_name: EdgeFaceWeights = EdgeFaceWeights.XXS,
+        preprocessing: PreprocessConfig | None = None,
+        providers: list[str] | None = None,
+    ) -> None:
+        if preprocessing is None:
+            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
+        model_path = verify_model_weights(model_name)
+        super().__init__(model_path=model_path, preprocessing=preprocessing, providers=providers)
--- a/uniface/recognition/mobileface.py
+++ b/uniface/recognition/mobileface.py
@@ -0,0 +1,49 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from __future__ import annotations
+
+from uniface.constants import MobileFaceWeights
+from uniface.model_store import verify_model_weights
+
+from .base import BaseRecognizer, PreprocessConfig
+
+__all__ = ['MobileFace']
+
+
+class MobileFace(BaseRecognizer):
+    """Lightweight MobileFaceNet model for fast face recognition.
+
+    This class provides a concrete implementation of the BaseRecognizer,
+    pre-configured for MobileFaceNet models. It is optimized for speed,
+    making it suitable for edge devices.
+
+    Args:
+        model_name (MobileFaceWeights): The specific MobileFaceNet model variant to use.
+            Defaults to `MobileFaceWeights.MNET_V2`.
+        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
+            configuration. If None, a default config for MobileFaceNet is used.
+        providers (list[str] | None): ONNX Runtime execution providers. If None, auto-detects
+            the best available provider. Example: ['CPUExecutionProvider'] to force CPU.
+
+    Example:
+        >>> from uniface.recognition import MobileFace
+        >>> recognizer = MobileFace()
+        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
+
+    Reference:
+        https://arxiv.org/abs/1804.07573
+        https://github.com/yakhyo/face-recognition
+    """
+
+    def __init__(
+        self,
+        model_name: MobileFaceWeights = MobileFaceWeights.MNET_V2,
+        preprocessing: PreprocessConfig | None = None,
+        providers: list[str] | None = None,
+    ) -> None:
+        if preprocessing is None:
+            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
+        model_path = verify_model_weights(model_name)
+        super().__init__(model_path=model_path, preprocessing=preprocessing, providers=providers)
--- a/uniface/recognition/models.py
+++ b/uniface/recognition/models.py
@@ -1,112 +0,0 @@
-# Copyright 2025-2026 Yakhyokhuja Valikhujaev
-# Author: Yakhyokhuja Valikhujaev
-# GitHub: https://github.com/yakhyo
-
-from __future__ import annotations
-
-from uniface.constants import ArcFaceWeights, MobileFaceWeights, SphereFaceWeights
-from uniface.model_store import verify_model_weights
-
-from .base import BaseRecognizer, PreprocessConfig
-
-__all__ = ['ArcFace', 'MobileFace', 'SphereFace']
-
-
-class ArcFace(BaseRecognizer):
-    """ArcFace model for robust face recognition.
-
-    This class provides a concrete implementation of the BaseRecognizer,
-    pre-configured for ArcFace models. It handles the loading of specific
-    ArcFace weights and sets up the appropriate default preprocessing.
-
-    Args:
-        model_name (ArcFaceWeights): The specific ArcFace model variant to use.
-            Defaults to `ArcFaceWeights.MNET`.
-        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
-            configuration. If None, a default config for ArcFace is used.
-        providers (list[str] | None): ONNX Runtime execution providers. If None, auto-detects
-            the best available provider. Example: ['CPUExecutionProvider'] to force CPU.
-
-    Example:
-        >>> from uniface.recognition import ArcFace
-        >>> recognizer = ArcFace()
-        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
-    """
-
-    def __init__(
-        self,
-        model_name: ArcFaceWeights = ArcFaceWeights.MNET,
-        preprocessing: PreprocessConfig | None = None,
-        providers: list[str] | None = None,
-    ) -> None:
-        if preprocessing is None:
-            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
-        model_path = verify_model_weights(model_name)
-        super().__init__(model_path=model_path, preprocessing=preprocessing, providers=providers)
-
-
-class MobileFace(BaseRecognizer):
-    """Lightweight MobileFaceNet model for fast face recognition.
-
-    This class provides a concrete implementation of the BaseRecognizer,
-    pre-configured for MobileFaceNet models. It is optimized for speed,
-    making it suitable for edge devices.
-
-    Args:
-        model_name (MobileFaceWeights): The specific MobileFaceNet model variant to use.
-            Defaults to `MobileFaceWeights.MNET_V2`.
-        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
-            configuration. If None, a default config for MobileFaceNet is used.
-        providers (list[str] | None): ONNX Runtime execution providers. If None, auto-detects
-            the best available provider. Example: ['CPUExecutionProvider'] to force CPU.
-
-    Example:
-        >>> from uniface.recognition import MobileFace
-        >>> recognizer = MobileFace()
-        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
-    """
-
-    def __init__(
-        self,
-        model_name: MobileFaceWeights = MobileFaceWeights.MNET_V2,
-        preprocessing: PreprocessConfig | None = None,
-        providers: list[str] | None = None,
-    ) -> None:
-        if preprocessing is None:
-            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
-        model_path = verify_model_weights(model_name)
-        super().__init__(model_path=model_path, preprocessing=preprocessing, providers=providers)
-
-
-class SphereFace(BaseRecognizer):
-    """SphereFace model using angular margin for face recognition.
-
-    This class provides a concrete implementation of the BaseRecognizer,
-    pre-configured for SphereFace models, which were among the first to
-    introduce angular margin loss functions.
-
-    Args:
-        model_name (SphereFaceWeights): The specific SphereFace model variant to use.
-            Defaults to `SphereFaceWeights.SPHERE20`.
-        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
-            configuration. If None, a default config for SphereFace is used.
-        providers (list[str] | None): ONNX Runtime execution providers. If None, auto-detects
-            the best available provider. Example: ['CPUExecutionProvider'] to force CPU.
-
-    Example:
-        >>> from uniface.recognition import SphereFace
-        >>> recognizer = SphereFace()
-        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
-    """
-
-    def __init__(
-        self,
-        model_name: SphereFaceWeights = SphereFaceWeights.SPHERE20,
-        preprocessing: PreprocessConfig | None = None,
-        providers: list[str] | None = None,
-    ) -> None:
-        if preprocessing is None:
-            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
-
-        model_path = verify_model_weights(model_name)
-        super().__init__(model_path=model_path, preprocessing=preprocessing, providers=providers)
--- a/uniface/recognition/sphereface.py
+++ b/uniface/recognition/sphereface.py
@@ -0,0 +1,50 @@
+# Copyright 2025-2026 Yakhyokhuja Valikhujaev
+# Author: Yakhyokhuja Valikhujaev
+# GitHub: https://github.com/yakhyo
+
+from __future__ import annotations
+
+from uniface.constants import SphereFaceWeights
+from uniface.model_store import verify_model_weights
+
+from .base import BaseRecognizer, PreprocessConfig
+
+__all__ = ['SphereFace']
+
+
+class SphereFace(BaseRecognizer):
+    """SphereFace model using angular margin for face recognition.
+
+    This class provides a concrete implementation of the BaseRecognizer,
+    pre-configured for SphereFace models, which were among the first to
+    introduce angular margin loss functions.
+
+    Args:
+        model_name (SphereFaceWeights): The specific SphereFace model variant to use.
+            Defaults to `SphereFaceWeights.SPHERE20`.
+        preprocessing (Optional[PreprocessConfig]): An optional custom preprocessing
+            configuration. If None, a default config for SphereFace is used.
+        providers (list[str] | None): ONNX Runtime execution providers. If None, auto-detects
+            the best available provider. Example: ['CPUExecutionProvider'] to force CPU.
+
+    Example:
+        >>> from uniface.recognition import SphereFace
+        >>> recognizer = SphereFace()
+        >>> # embedding = recognizer.get_normalized_embedding(image, landmarks)
+
+    Reference:
+        https://arxiv.org/abs/1704.08063
+        https://github.com/yakhyo/face-recognition
+    """
+
+    def __init__(
+        self,
+        model_name: SphereFaceWeights = SphereFaceWeights.SPHERE20,
+        preprocessing: PreprocessConfig | None = None,
+        providers: list[str] | None = None,
+    ) -> None:
+        if preprocessing is None:
+            preprocessing = PreprocessConfig(input_mean=127.5, input_std=127.5, input_size=(112, 112))
+
+        model_path = verify_model_weights(model_name)
+        super().__init__(model_path=model_path, preprocessing=preprocessing, providers=providers)
Author	SHA1	Message	Date
Yakhyokhuja Valikhujaev	6ce397b811	feat: Add MODNet portrait matting (#108 ) * feat: Add MODNet portrait matting * docs: Update docs and example of portrait matting * fix: Fix linting issue	2026-04-11 23:30:32 +09:00
Yakhyokhuja Valikhujaev	9bf54f5f78	feat: Add EdgeFace recognition model (#105 ) * refactor: Split recognition models into separate files * feat: Add EdgeFace recognition model * release: Bump version to v3.4.0	2026-04-04 20:11:28 +09:00
Yakhyokhuja Valikhujaev	c87ec1ad0f	docs: Add example images and update MkDocs files (#104 ) * chore: Add example inference results * docs: Update MkDocs and README files	2026-04-04 18:28:27 +09:00
Yakhyokhuja Valikhujaev	9e56a86963	chore: Update docs and clean up notebook outputs before before commit (#102 ) * chore: Add links for repo and docs on example notebooks * ref: Compress jupyter notebook sizes * ci: Add nbstripout pre-commit hook for notebook output stripping * docs: Add coding agent docs and commit message tag	2026-04-03 10:10:51 +09:00