MUMBAI, India, June 26 -- Intellectual Property India has published a patent application (202621054342 A) filed by Symbiosis International Deemed University on April 28, 2026, for Style-Based Generative Adversarial Network System For High-Fidelity Synthetic Human Face Generation.
Inventors include Dr. Latika Pinjarkar; Dr. Gagandeep Kaur; Vanshika Gadhwal; Arya Rehpade; and Harshal Meshram.
The application for the patent was published on June 19, 2026, under issue no. 25/2026.
Abstract: ABSTRACT STYLE-BASED GENERATIVE ADVERSARIAL NETWORK SYSTEM FOR HIGH-FIDELITY SYNTHETIC HUMAN FACE GENERATION The present invention discloses a style-based generative adversarial network system (100) for generating high-fidelity synthetic human faces. The system comprises a mapping network module (110) that transforms random latent vectors into a disentangled intermediate latent space enabling independent manipulation of facial attributes including age, gender, expression, and pose. An adaptive instance normalization module (120) applies scale-specific style transformations at multiple synthesis layers controlling coarse structure through fine textures. A progressive growing training module (130) incrementally increases resolution from 4x4 to 1024x1024 pixels ensuring stable training convergence. A quality evaluation module (140) implements Frechet Inception Distance and Inception Score metrics achieving FID below 7.0 and IS above 5.0 for optimized configurations. The generator-discriminator framework (150) employs adversarial training with R1 gradient penalty and path length regularization. The system achieves 97.3% artifact-free image generation with photorealistic quality indistinguishable from real images in human perception studies. [
Disclaimer: Curated by HT Syndication.
