16:57 pm

  • LLM-powered natural language search recommendation engine for images and collections: The user can type out the kinds of images or collections they are looking for, by describing in natural language, including keywords, labels, etc, and we perform a vector based search to find the best-fit images or collections

  • We provide two options - individual image search or collection search, allowing the user to toggle for either

  • We show a sequence of 20 images for image search, and 5 collections for collection search, in a decreasing order of accuracy or best-fit metric

  • This painting/visual/image recommender won’t use embeddings; instead, we will search through the image descriptions and labels in the image metabucket to return the images with the highest convergence [defined loosely as most number of overlapping labels and words in description]

Eg of flow

The user wants to find a painting which depicts someone on a mountain, with the backdrop of the milky way, gazing ahead solemnly. They describe this in a natural language form, and the system returns images with the highest convergence by searching through image labels and descriptions.

  • Instead, we will be using the crude but pretty effective tool of natural language to find images using labels and descriptions.

Links : TODO

Tags :

Date : 11th April, Friday, 2025, (Wikilinks: 11th April, April 25, April, 2025. Friday)

Category : Others