Sergey Karpukhin, Denis Bobkov, Kirill Morozov, Aibek Alanov
8 min
Abstract
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. In this work we propose a pipeline for controllable generation of dataset with diverse faces for generative models training and finetuning. We achieve this by expanding foundational Arc2Face diffusion model with a flow matching generative model in the manifold of identity embeddings. We demonstrate that such approach achieves both goals of generative fidelity and label diversity not present in real human-centric datasets. As a result we collect and further contextualize a highly balanced identity-centric dataset of humans named SynthFaces. By design, our released images and identities are synthetic, reducing direct reliance on real personal images; we empirically check identity separation with a nearest-neighbor non-leakage metric. We purpose it for training and tuning of fair models for various downstream tasks, and showcase its usability in finetuning an existing adapter-based method for personalized generation.
Sam: What distinguishes that from earlier synthetic identity methods? They also work with face representations.
Alex: Some earlier approaches interpolate or perturb representations of real people. ID-Flow learns from real representations during training, but does not require a reference person's representation when generating a new identity. Compared with HyperFace, another synthetic identity approach, it adds explicit demographic control and checks separation from training identities.
Sam: That gives them a numerical identity, not a usable training image. How does it become the dataset?
Alex: Arc2Face, a face generator, turns the sampled representation into a face image. The pipeline removes problematic images using image statistics, automated visual assessments, and identity-similarity checks. After filtering, roughly fifteen thousand identities remain. Personalization models then place those people into varied scenes with captions.
Sam: The same identity needs different surroundings without the caption changing who that person is. How do they manage that?
Alex: They strip captions of facial descriptions and other attributes already controlled during identity sampling. Then they match captions to suitable age groups. For example, a child's identity receives a caption suitable for a child. That separates choosing the person from choosing the scene.
Sam: Filtering often changes the distribution you carefully constructed. Did it undo some of the balance?
Alex: It did weaken it. Filtering improves their image-quality measures, but worsens age balance and reduces some measures of distribution coverage. Rare demographic combinations also produced clusters of similar identities, requiring extra filtering. The authors report that trade-off rather than treating curation as a free improvement.
Sam: After those trade-offs, how much more balanced is the data? Give us one comparison we can hold onto.
Alex: Their reported label-diversity comparison gives SynthFaces a distance from uniform of about point zero five, versus about point four five for FFHQ. Lower means closer to equal representation. That score averages across the selected demographic label categories; it does not establish equal diversity within every group.
Sam: Balanced labels are useful, but researchers ultimately need better models. Where is the strongest evidence that this changes downstream performance?
Alex: Face recognition provides a useful test. They train a recognition model on a separate synthetic collection containing about half a million images. They then evaluate both conventional recognition benchmarks and Balanced Faces in the Wild, which has human-assigned demographic labels.
Sam: What counts as fairer recognition here? Equal average accuracy could still hide different kinds of mistakes.
Alex: They set a global false match rate: how often different people are mistakenly accepted as matching. Then they examine missed genuine matches within each demographic group. Their disparity measure compares each group's miss rate with the overall miss rate, rather than simply comparing raw accuracy.
Sam: So the question is whether one group bears more of the rejection burden at the same operating threshold. What happens?
Alex: On Balanced Faces in the Wild, their model has the lowest reported disparity among the compared methods. At a false match rate of a tenth of a percent, the worst-group miss rate is nearly equal to the overall rate. Across the conventional benchmarks, it also has the best average racial disparity score.
Sam: Does that establish that balancing alone caused the improvement? Or are the compared models different in other ways?
Alex: They differ in architecture and training collections, so this is not an isolated test of balancing alone. The paper also acknowledges lower overall recognition accuracy on established benchmarks. Better relative parity does not mean the model makes fewer mistakes overall.
Sam: Do the other applications support the same story, or is recognition carrying the argument?
Alex: They provide supporting, but qualified, evidence. Fine-tuning a personalized image generator improves identity reconstruction on their balanced face benchmark. On richer contextual prompts, performance is mixed rather than clearly better. That distinction matters if your application puts people in scenes rather than reconstructing close-up faces.
Sam: They also test answering questions about images. What does that add?
Alex: They fine-tune a model for visual question answering, here predicting demographic attributes. On their internal face-image evaluation, average accuracy rises from roughly sixty-five percent to eighty percent, while cross-group accuracy gaps shrink. But evaluation images match each model's training domain. It does not establish transfer to arbitrary real photographs.
Sam: Before we reach a verdict, what should listeners make of the privacy claim? The released people are synthetic, but the training representations came from real people.
Alex: The authors explicitly say there is no formal privacy guarantee. They measure how far generated identities sit from their nearest training identities. That is evidence of separation under a face representation, not proof against every kind of leakage. And upstream generators still constrain which faces can be produced faithfully.
Sam: Given those limits, who should read the full document, and where should they start?
Alex: Researchers building synthetic human datasets should start with Dataset construction, especially the filtering trade-offs and the comparison with balanced text prompts. Recognition researchers should read Face Recognition Fairness, focusing on threshold calibration alongside the accuracy caveat. For demographic question answering, inspect the domain-matched evaluation before assuming broader transfer.
Sam: And what should everyone else remember when they next hear that synthetic data solves bias?
Alex: Choosing which people a model learns from matters, even when those people are invented. SynthFaces offers a practical way to control that choice, with encouraging task-specific results. But balanced labels are not a privacy guarantee, or a promise of equal performance everywhere.
Sam: Keep the composition of the dataset in view.