| --- |
| license: apache-2.0 |
| library_name: onnx |
| tags: |
| - image-segmentation |
| - face-parsing |
| - portrait-segmentation |
| - onnx |
| - comfyui |
| --- |
| |
| # EasyPortrait ONNX |
|
|
| ONNX exports of the pretrained EasyPortrait portrait segmentation and face parsing checkpoints. |
|
|
| The original checkpoints and training code were published by the EasyPortrait authors in [hukenovs/easyportrait](https://github.com/hukenovs/easyportrait). This repository exists to make the models easier to use in modern runtimes, especially ComfyUI environments where installing the old `mmsegmentation` / `mmcv-full` stack can be slow or fragile. |
|
|
| ## Files |
|
|
| This repository contains 26 validated ONNX models and a `manifest.json`. |
|
|
| Model families: |
|
|
| - Portrait segmentation: |
| - BiSeNet-V2 |
| - DANet |
| - DeepLabV3 |
| - FastSCNN |
| - FCN + MobileNetV2 |
| - FPN + ResNet50 at 224, 384, 512, and 1024 |
| - SegFormer-B0 at 224, 384, 512, and 1024 |
| - Face parsing: |
| - BiSeNet-V2 |
| - DANet |
| - DeepLabV3 |
| - FastSCNN |
| - FCN + MobileNetV2 |
| - FPN + ResNet50 at 224, 384, 512, and 1024 |
| - SegFormer-B0 at 224, 384, 512, and 1024 |
|
|
| Three upstream README-only checkpoints are not included because they do not include enough model config metadata for reliable conversion: `extremec3net_ps`, `sinet_ps`, and `ehanet_fp`. |
|
|
| ## Input And Output |
|
|
| Each model expects a normalized tensor: |
|
|
| - Input name: `image` |
| - Input shape: `[batch, 3, model_size, model_size]` |
| - Input dtype: `float32` |
| - Channel order before normalization: BGR |
| - Mean: `[143.55267075, 132.96705975, 126.94924335]` |
| - Std: `[60.2625333, 60.32740275, 59.30988645]` |
|
|
| Each model returns logits: |
|
|
| - Output name: `logits` |
| - Output shape: `[batch, classes, model_size, model_size]` |
| - Output dtype: `float32` |
|
|
| Use `argmax(axis=1)` to get the segmentation class map. |
|
|
| Labels: |
|
|
| - Portrait segmentation: `background`, `person` |
| - Face parsing: `background`, `skin`, `left brow`, `right brow`, `left eye`, `right eye`, `lips`, `teeth` |
|
|
| ## Validation |
|
|
| The ONNX models were exported from the original PyTorch/mmseg checkpoints and validated with ONNX Runtime against the PyTorch/mmseg output. The validation compares logits and verifies that the final `argmax` segmentation agrees. |
|
|
| See `manifest.json` for per-model metadata and validation values. |
|
|
| ## ComfyUI |
|
|
| These files are used by the ComfyUI custom node: |
|
|
| [ComfyUI-EasyPortrait](https://github.com/sadzip/ComfyUI-EasyPortrait) |
|
|
| The node downloads models from this repository on first use. |
|
|
| ## Credits |
|
|
| Thanks to the EasyPortrait authors for training and releasing the original models: |
|
|
| - Original repository: [hukenovs/easyportrait](https://github.com/hukenovs/easyportrait) |
|
|
| Co-authors of this ONNX conversion and ComfyUI integration: |
|
|
| - sadzip |
| - OpenAI Codex |
|
|
