MUMBAI, India, Oct. 26 -- Intellectual Property India has published a patent application (202641123781 A) filed by Vellore Institute Of Technology on October 15, 2026, for Method For Resolution-Adaptive Spatial-Semantic Attention Conditioning In Diffusion Image Generation.
Inventors include Deepika J; Keshav Dadhich; and Khanak Shah.
The application for the patent was published on October 23, 2026, under issue no. 43/2026.
Abstract: ABSTRACT METHOD FOR RESOLUTION-ADAPTIVE SPATIAL-SEMANTIC ATTENTION CONDITIONING IN DIFFUSION IMAGE GENERATION A method for generating an image using a structural-guided diffusion model receives a structural input (102) and textual information (104), encodes the textual information into text tokens (108) using a text encoder (106), and processes the structural input using a structural-conditioning network (110) to obtain an intermediate structural feature representation (112). A region extractor (114) converts the intermediate structural feature representation into localized region tokens (116) projected into a semantic embedding space of the text tokens. A spatial-semantic attention module (118) determines a learned spatial-semantic correspondence (120) between the localized region tokens and the text tokens. A resolution-adaptive attention processor (122) transforms the correspondence according to the resolution of each attention layer (128) of a diffusion denoising network (126) into a resolution-adapted correspondence (124) incorporated into attention scores during iterative denoising of a latent representation (130) to produce a generated image (132). (Fig. 1)
Disclaimer: Curated by HT Syndication.
