A preprint on synthetic face generation reports demographic mixes closer to an even target for gender, race, age and joint gender-and-race tests when a single latent-space guidance step is used. The approach was designed to ask whether a simple semantic boundary predictor could improve balance across individual and combined attributes while preserving perceptual quality and keeping generation practical. Its evidence comes from generated images and reported proxy scores.
The method, called the Semantic Boundary Predictor, or SBP, treats a latent representation—the model's internal numerical state—as a place where demographic boundaries can be drawn. It estimates each boundary with a lightweight linear classifier from final-stage representations at t=0. During generation, SBP applies one shift in that direction to the initial noisy latent before reverse denoising, while leaving the pre-trained diffusion model unchanged. In practical terms, the extra guidance is inserted once; the underlying model is not retrained or fine-tuned.
The Unusual Timing Is the Point
The choice of timing follows a pattern measured across denoising. A linear classifier's gender accuracy rose from 59% at t=50 to 97% at t=0; for four-class race, it rose from 71% to 91%. In the reported experiments, demographic attributes were therefore more linearly separable later in denoising. A separate compactness test found that about 92% of variance could be retained with 5,000 of 12,288 principal components at t=0, while about 8,000 were needed at t=50 for comparable variance.
Boundary learning used 15,000 generated samples, split evenly between training and validation. Evaluation used 10,000 generated images for each demographic group and covered gender, binary race, four-class race, age and joint Gender + Race. The main fairness measure, Fairness Discrepancy or FD, is lower when the predicted demographic distribution is closer to a uniform target; FID served as the perceptual-quality measure.
The Clearest Changes Were in Balance Scores
In the reported CelebA-HQ gender comparison, FD was 0.051 before SBP and 0.001 with SBP, which the paper describes as approximately 98% lower; FID was 34.66 and 31.54, respectively. On FFHQ, FD was reported as approximately 91% lower with SBP, while FID was marginally higher. These are reported score comparisons within the tested settings, not evidence that the same result will hold for every generator or dataset.
Race comparisons varied by the number of categories. For binary race, SBP's FD/FID were 0.018/37.11, compared with 0.387/39.46 for the listed pre-trained CelebA model; the paper summarizes the FD difference as approximately 95%. For four-class race, the corresponding figures were 0.247/34.80 versus 0.289/34.48, a reported 15% FD difference, with FID values close.
On CelebA, the reported age FD/FID was 0.135/36.98 with SBP, versus 0.450/46.01 for the pre-trained model. The age comparison therefore showed lower reported values for both Fairness Discrepancy and FID in the SBP result.
The approach was also evaluated on a combined attribute rather than one label at a time. For joint Gender + Race on CelebA, SBP's FD/FID were 0.055/45.16, versus 0.256/60.68 for random sampling and 0.075/49.91 for Balancing Act. The paper reports FD differences of 78% and 26% against those comparators and lists SBP as the lowest-FD and lowest-FID method in that comparison.
Looking within groups, average race FID was 59.76 versus 45.45, and average age FID was 57.43 versus 42.52. The paper describes those as approximately 24% and 26% differences. Its largest stated gains involved Indian, Asian, young and old faces, while changes for White and adult faces were described as small.
Quality and Speed Stayed Competitive
The reported speed stayed close to the simplest comparators. SBP generated 1.41 samples per second, compared with 1.56 for random sampling and 1.45 for Balancing Act. It was faster than Diff-AE at 0.30 and Unbiased-Diff at 0.26. That result addresses the study's question of whether one-shot guidance can remain computationally practical.
The guidance-strength test pointed to δ=4 as the reported operating point. At δ=0, gender FD/FID were 0.051/34.66; at δ=4 they were 0.001/31.54; and at δ=5 they were 0.001/33.20. The last setting added no FD improvement beyond δ=4 and had the higher FID of the two guided settings.
The reported training-size comparison showed a plateau in the fairness score. With 5,000, 15,000 and 45,000 training samples, FD/FID were 0.002/36.27, 0.001/31.54 and 0.001/31.46. FD remained 0.001 after 15,000 samples, while moving from 15,000 to 45,000 changed FID only marginally.
What the Tests Leave Open
The results have a narrower reach than a general claim about fairness in face technology. They describe synthetic images evaluated with classifier-derived demographic labels, FD, FID and throughput. The supplied analysis reports no human judgments or downstream face-recognition performance, no confidence intervals or formal significance tests, and no repeated-run variability estimates. It also notes that boundary learning depends on external attribute classifiers and that the baseline methods do not all use the same pretrained models and training datasets.
The next tests are whether intermediate denoising timesteps change the fairness-quality trade-off, whether SBP transfers to other datasets, generators or demographic definitions, and whether better synthetic balance translates into face-recognition accuracy or subgroup fairness. The authors do not claim global linearity of demographic attributes or perfect semantic disentanglement. For now, the document remains a preprint, so these reported results are best read as findings from the tested computational settings.
Paper data and sources
Original title: Learning Late, Guiding Early: Timestep-Decoupled Semantic Guidance for Fair Face Generation
Authors: Subir Kumar Parida, Rajbabu Velmurugan, Ketan Kotwal et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text