StyleID : A Perception-Aware Dataset and Metric for Stylization-Agnostic Facial Identity Recognition

SIGGRAPH 2026 / ACM TOG Journal Track

StyleID is a CLIP-based image encoder designed to produce identity embeddings that remain robust under stylization.
It can be used for identity similarity measurement, retrieval, evaluation, and identity-aware conditioning in generative pipelines.

---

Project Status

• (23.April) Quick start recognition model release
• (24.April) ArXiv released
• Dataset release
• Data generation code release

---

Quick Start

import torch
from transformers import CLIPModel, CLIPProcessor
from PIL import Image

device = "cuda" if torch.cuda.is_available() else "cpu"

model = CLIPModel.from_pretrained("kwanY/styleid").to(device)
processor = CLIPProcessor.from_pretrained("kwanY/styleid")

img_path = "example.jpg"
img = Image.open(img_path).convert("RGB")

inputs = processor(images=img, return_tensors="pt").to(device)

with torch.no_grad():
    emb = model.get_image_features(**inputs)
    emb = emb / emb.norm(dim=-1, keepdim=True)  # optional but recommended

Notes

• Not suitable for images with multiple faces
• Rough center crop near the face is recommended for better performance

---

Intended Uses

StyleID embeddings can be used for:

• Identity similarity comparison
• Image retrieval
• Stylized identity evaluation
• Identity-aware conditioning for generative models
• Research on face recognition under domain shift and stylization

---

License and Usage Notice

• StyleID is released for non-commercial research use.
• Do not use FFHQ-derived data for biometric human recognition

Citation

If you find this work useful, please cite the paper:

---