An arXiv version-1 preprint dated 26 August 2026 reports that a one-step adjustment to the starting noise of a face-generation model was accompanied by substantially lower demographic imbalance in tests covering gender, race, age and joint gender-race categories. The method, called Semantic Boundary Predictor, or SBP, leaves the model’s denoising process unchanged after that initial intervention.
The comparison focused on the generator’s output.
The main fairness measure, Fairness Discrepancy (FD), is the distance between the demographic mix predicted in generated images and an even target mix; a lower FD means the output is closer to uniform. The paper also reports Fréchet Inception Distance, or FID, as a perceptual-quality score, with lower values treated as better under its evaluation protocol.
The reference data included CelebA-HQ, with 30,000 256-by-256 facial images, and FFHQ, with 70,000 high-quality facial images. SBP boundary learning used 15,000 generated samples, while evaluation generated 10,000 images for each demographic group.
The boundary is learned late and used early.
SBP learns a linear semantic boundary, essentially a direction in the model’s latent representation associated with a demographic attribute, from late-stage representations. It applies that direction once to the initial noisy latent by adding or subtracting it with a chosen guidance strength; the remaining reverse-denoising steps and the underlying model weights are unchanged.
The paper’s classifier measurements showed stronger separation later in denoising: reported gender accuracy rose from 59% at timestep 50 to 97% at timestep 0, while four-class race accuracy rose from 71% to 91%.
The biggest changes appeared in gender and binary-race tests.
In the reported CelebA-HQ gender comparison, FD was 0.051 for the pre-trained model and 0.001 with SBP, described as an approximately 98% reduction. FID also fell, from 34.66 to 31.54. On FFHQ, the paper reports an approximately 91% reduction in gender FD, alongside a marginal increase in FID.
Race results varied with the category scheme. In the binary-race comparison, SBP’s FD/FID were 0.018/37.11, compared with 0.387/39.46 for the pre-trained model. In the four-class comparison, SBP’s FD/FID were 0.247/34.80, compared with 0.289/34.48. The paper reports these as approximately 95% and 15% lower FD, respectively.
Age generation also showed lower reported scores in the CelebA-HQ tests: SBP’s FD/FID were 0.135/36.98, compared with 0.450/46.01 for the pre-trained model. In joint gender-race generation, the values were 0.055/45.16 with SBP versus 0.256/60.68 with random sampling.
The extra step kept sampling close to the comparison systems.
On a single NVIDIA A100 GPU, the reported throughput was 1.41 samples per second for SBP, versus 1.56 for the original latent diffusion model and 1.45 for Balancing Act. The numbers put the added method close to its comparison systems in this hardware test.
Guidance strength mattered in the gender ablation. FD was 0.051 at delta=0 and 0.001 at both delta=4 and delta=5; FID was 34.66, 31.54 and 33.2 at those same settings. Increasing delta beyond 4 therefore did not lower FD further and raised FID.
The scores do not settle downstream fairness.
The evidence is limited to controlled computational tests on synthetic face outputs and proxy measures. It does not establish fairness in a downstream face-recognition system or show that the images represent real populations beyond the paper’s predefined categories and classifier labels.
SBP relies on external attribute classifiers, so errors or demographic bias in predicted labels may affect the boundaries it learns. The method also requires a separate semantic boundary for each demographic attribute.
The paper reports no confidence intervals, hypothesis tests or formal uncertainty estimates for the reported differences. Taken together, the preprint presents SBP as an inference-time way to shift demographic balance without retraining the diffusion model, but whether it transfers across other architectures and datasets or improves downstream recognition fairness remains unresolved.
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