MUMBAI, India, Aug. 17 -- Intellectual Property India has published a patent application (202621078019 A) filed by Indian Institute Of Technology Gandhinagar on June 24, 2026, for System And Method For Efficient Intrinsic Image Decomposition Using Latent Flow Matching.
Inventors include Singla, Mithlesh; Kumari, Seema; and Raman, Shanmuganathan.
The application for the patent was published on August 14, 2026, under issue no. 33/2026.
Abstract: ABSTRACT SYSTEM AND METHOD FOR EFFICIENT INTRINSIC IMAGE DECOMPOSITION USING LATENT FLOW MATCHING A system (100) for efficient intrinsic image decomposition using latent flow matching is disclosed. The system comprises a feature encoding module (102) configured to extract feature representations from an input image, a latent representation module (104) configured to generate a latent input, and a latent transport module (106) operatively coupled thereto to transform the latent input into a shading latent representation based on the feature representations. The latent transport module (106) performs a learned transformation in a reduced-dimensional latent space using latent flow matching in a single forward pass without iterative sampling. A decoder module (108) generates a shading component from the shading latent representation, and a reflectance determination module (110) derives a reflectance component based on the input image and the shading component. The system enables intrinsic decomposition without requiring auxiliary inputs, while improving computational efficiency and consistency.
Disclaimer: Curated by HT Syndication.
