whathelll / vectordb

plan

• explore and understand the CLIP interaction notebook.
• load images into vector db
• explore the space of duplicate image search
• explore similar image search
• explore search of images based on text
• negative prompt
• reverse caption
• portfolio search (multiple images at the same time)

Scenarios we need to test

• I like this image, find me others like it
• recommender engine for similar items that I've looked at recently
• semantic search instead of keyword search based on predefined tags
• semantic search of images
• reverse image search
• define likes and we recommend similar items for you

Analysis

• It seems using OpenAI clip on MNIST is shit. This was noted in their paper.
• However it looks really interesting on a dataset such as Modern Architecture. See Modern Architecture

Interesting Links

https://colab.research.google.com/github/robgon-art/open-clip/blob/main/Create_Captions_with_OpenCLIP.ipynb