ID-Flow

Pretrained ID-Flow weights from SynthFaces: Balanced Large Scale Human Dataset (accepted to BMVC 2026).

ID-Flow is a label-conditional geodesic flow-matching model on the unit hypersphere of ArcFace identity embeddings (~11.8M parameters). Given an (age, race, gender) combination it samples a new identity embedding, which Arc2Face turns into face images. Sampling all 126 label combinations uniformly gives a demographically balanced set of synthetic identities.

• Code: github.com/ControlGenAI/SynthFaces (to be released)
• Paper: arXiv:2610.40119

Files

file	description
idflow.pt	weights-only, self-describing checkpoint (model weights + hyper-parameters), 47 MB

Label order is [age, race, gender]: age 0-2, 3-9, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+; race White, Black, Latino_Hispanic, East Asian, Southeast Asian, Indian, Middle Eastern; gender Male, Female.

Usage: ID-Flow → Arc2Face

Setup (from the code repository):

git clone https://github.com/ControlGenAI/SynthFaces && cd SynthFaces
pip install -r requirements.txt
huggingface-cli download shredder67/id-flow idflow.pt --local-dir .
python scripts/download_arc2face.py --root Arc2Face      # Arc2Face + InsightFace models

Sample a balanced set of identities (all 126 label groups), then generate images with Arc2Face:

python scripts/sample_identities.py --ckpt idflow.pt --n_total 1000 --output_dir outputs/identities

python scripts/sample_images.py sample \
    --embeddings_path outputs/identities/id_embeddings.pt \
    --output_dir outputs/images --images_per_identity 4
python scripts/sample_images.py merge --output_dir outputs/images --n_parts 1

Or sample identity embeddings for chosen labels directly in Python:

import torch
from idflow import load_idflow

model = load_idflow("idflow.pt", device="cuda:0")
y = torch.tensor([[3, 0, 1]] * 16, device="cuda:0")   # age 20-29, White, Female
ids = model.sample(bs=16, y=y, guidance_scale=2.0, device="cuda:0")  # (16, 512) unit-norm ArcFace embeddings

The embeddings can be passed to Arc2Face as id_embeddings.pt (see scripts/sample_images.py). A guidance scale of 2.0 is used in the paper; higher values follow the labels more closely at the cost of diversity.

Notes

The Arc2Face, InsightFace and Stable Diffusion 1.5 weights used for image generation are distributed under their own licenses (the Arc2Face and InsightFace weights are non-commercial); please check them before use.

Citation

@article{karpukhin2026synthfaces,
  title   = {SynthFaces: Balanced Large Scale Human Dataset},
  author  = {Karpukhin, Sergey and Bobkov, Denis and Morozov, Kirill and Alanov, Aibek},
  journal = {arXiv preprint arXiv:2610.40119},
  year    = {2026},
  note    = {Accepted to BMVC 2026}
}