Split	IoU	Dice	Precision	Recall
Fine-tuned test	0.9368	0.9674	0.9651	0.9697
Fine-tuned test_hard	0.9393	0.9687	0.9663	0.9711
Baseline SegFace test	0.1360	0.2394	0.1360	1.0000
Baseline SegFace test_hard	0.1803	0.3055	0.1803	1.0000

Comparison Notes

Test IoU improved from 0.1360 to 0.9368 (+0.8008).
Test Dice improved from 0.2394 to 0.9674 (+0.7280).
Hard test IoU improved from 0.1803 to 0.9393 (+0.7590).
Hard test Dice improved from 0.3055 to 0.9687 (+0.6632).
The pretrained baseline tended to over-predict hair regions, showing high recall but very low precision.

Test Split Summary

train / val / test / test_hard: 50,000 / 5,000 / 921 / 150
source overlap train-test: 0
source overlap val-test: 0

Qualitative Comparison

Test

Selection: top-improved, previews: 12
Representative sample ad5c2098-485e-48d7-979b-6190bb01e7a2: baseline IoU 0.0481 -> fine-tuned IoU 0.9751 (+0.9270).

Test Hard

Selection: top-improved, previews: 4
Representative sample ccfa01f0-1b0a-4a39-a1c9-9d5c883f3fb2: baseline IoU 0.0707 -> fine-tuned IoU 0.9546 (+0.8839).

Bundle Contents

best.pt: inference checkpoint
config.json: training-time model config
training_run_summary.json: run summary and validation metrics
reports/*.json: hold-out test and baseline comparison reports when available
reports/mask_comparison/**: baseline vs fine-tuned qualitative comparison sheets and sample previews when available
docs/test_split_summary.json: hold-out split summary when available
inference.py: local / Hub inference example
requirements.txt: minimal runtime dependencies
hair_mask_dataset/, models/: custom model code required to load the checkpoint

Inference

Run locally from the root of this model bundle:

bash

python -m pip install -r requirements.txt
python inference.py --checkpoint best.pt --input path/to/input.jpg --output-mask output_mask.png --output-overlay output_overlay.png

You can also load directly from the Hugging Face Hub after uploading:

bash

python -m pip install -r requirements.txt
python inference.py --repo-id siik/segface_hair_khairstyle --input path/to/input.jpg --output-mask output_mask.png --output-overlay output_overlay.png

Inputs and Outputs

Input: RGB portrait image file such as .jpg or .png
Output mask: single-channel binary hair mask PNG
Output overlay: optional RGB overlay image for quick visual inspection

Notes

This repo contains custom code and a raw PyTorch checkpoint, not a Transformers-format model.
Preprocessing expects RGB input, resize to 512, ImageNet normalization, and sigmoid threshold 0.5.
Before making the repository public, verify whether your AIHub / K-Hairstyle data usage terms allow public redistribution of derived model weights.

Training Artifacts