SynthFaces dataset enables balanced human generation for fair AI models
Original title: SynthFaces: Balanced Large Scale Human Dataset
SynthFaces expands Arc2Face diffusion model with flow matching to create diverse synthetic faces, reducing privacy risks and biases in real datasets.
The work proposes a pipeline for controllable generation of a balanced large-scale human dataset named SynthFaces. It expands foundational Arc2Face diffusion model with a flow matching generative model in the manifold of identity embeddings to achieve generative fidelity and label diversity. The synthetic images and identities reduce direct reliance on real personal information while empirically verifying identity separation with a nearest-neighbor non-leakage metric. SynthFaces is released for training and tuning of fair models for downstream tasks including finetuning adapter-based personalized generation.
- SynthFaces expands Arc2Face diffusion model with flow matching in identity embeddings manifold
- Collects highly balanced identity-centric dataset of humans
- Synthetic images reduce direct reliance on real personal images
- Identity separation verified with nearest-neighbor non-leakage metric
- Products
- SynthFaces
- Organizations
- Hugging Face
- AI models
- Arc2Face
From the original article
1 min read at huggingface.co
The pursuit of fair and equitable AI in human-centric tasks is critically hindered by the inherent biases and privacy risks of real-world datasets, which offer imbalanced representations of identity and often contain sensitive personal information. Consequently, state-of-the-art generative models trained on this skewed data exhibit unfair performance, particularly in identity-preserving generation.
By Sergey Karpukhin, Denis Bobkov, Kirill Morozov · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.