Live data from Hacker News

Show HN: Search photos on Unsplash using natural language queries

github.com

1–5 of 5 posts

Re: Show HN: Search photos on Unsplash using natural language queries

#2
This is one of my projects using OpenAI's CLIP model to do interesting things. So the big credit goes to them!

I processed 2M images from the Unsplash dataset with CLIP and stored the feature vector representation of each photo (a 512 element vector). You can now encode a text query with CLIP in the same latent space and search the database.

I also tried doing some simple arithmetic with the feature vectors to combine the result of a search query and a photo. You can for example search for "Sydney Opera house" and give it a bight photo and you will get a photo of the Sydney Opera at night.

You can directly jump to the Google Colab notebook if you want to give it a try: https://colab.research.google.com/github/haltakov/natural-la...

Re: Show HN: Search photos on Unsplash using natural language queries

#4
post #2

This is one of my projects using OpenAI's CLIP model to do interesting things. So the big credit goes to them! I processed 2M images from the Unsplash dataset with CLIP and stored the feature vector representation of each photo (a 512 element vector). You can now encode a text query with CLIP in the same latent space and search the database. I also tried doing some simple arithmetic with the feature vectors to combin…

Quite an interesting project!

Wanted to give it a shot but it seems to fail when attempting to fetch the precomputed features - 504 Gateway Time-outs. Quota limits might have been reached?

Re: Show HN: Search photos on Unsplash using natural language queries

#5
post #2

This is one of my projects using OpenAI's CLIP model to do interesting things. So the big credit goes to them! I processed 2M images from the Unsplash dataset with CLIP and stored the feature vector representation of each photo (a 512 element vector). You can now encode a text query with CLIP in the same latent space and search the database. I also tried doing some simple arithmetic with the feature vectors to combin…

Quite an interesting project! Wanted to give it a shot but it seems to fail when attempting to fetch the precomputed features - 504 Gateway Time-outs. Quota limits might have been reached?

Yeah, it's a 2 GB file, so it hits the quota limit on the Google Drive pretty fast. There is a problem with the mirror at transfer.army that I'll fix today.