Show HN: I made a Pinterest clone using SigLIP image embeddings | Hacker News Reader

Show HN: I made a Pinterest clone using SigLIP image embeddings

mood-amber.vercel.app

98 pointsverse21 comments

Click an image to get similar images.

I crawled Tumblr and used SigLIP to get vector embeddings for many images.

When you click an image, it finds the most similar vector embeddings in the database, and returns the corresponding images.

Open articleView on HN

Sometimes there are duplicate results, e.g. https://mood-amber.vercel.app/images/0b733fc2-7093-4443-8872... has two copies of https://mood-amber.vercel.app/images/f920a599-bbd7-4805-3317... right next to each other. (The link UUID is the same, so I assume this is an issue with the search algorithm, not simply duplicate data that got scraped.)

ah! thank you for pointing this out. will fix

Also, check https://same.energy/

Very cool! How did you get the collage layout? I noticed images in each column don’t have the same size. I assume images get Centre cropped?

It’s called a masonry grid. Images retain their aspect ratio so they don’t need to be cropped. You can kind of simulate it with css but there’s proposals to add a proper masonry layout to css

yeah. I actually wrote the logic for the layout myself (wasn't really happy with the available libraries). may open source this if people are interested!

Can you share your GH to follow updates? Also, take a look at this, they have a layout too https://github.com/lit/lit/tree/main/packages/labs/virtualiz...

will post on twitter:

https://x.com/verse_

I loved your text effects! You did some cool side-projects

Isn't the time for some big movements? Get in touch

Cool! I haven’t tried SigLIP out yet but it seems to be the new hotness over CLIP… I just dont have a good project idea yet

Is there a repo, especially for training? I'd like to see how SigLIP performs on a dataset of only anime images.

The the vision training models are available here: https://github.com/google-research/big_vision/tree/main which I am assuming, based on the research paper is what was used for the project.

There are some interesting images there. Why are you not including the source of the images?

Good stuff! Do you have any intuitive sense of whether SigLIP is particularly stronger than CLIP here? Also vector DB over Faiss index?

I haven't done much testing or anything, but it seems to me that siglip "understands" what it's looking at more than CLIP

also no, I just put everything on Supabase and added pgvector. super easy:

https://supabase.com/docs/guides/database/extensions/pgvecto...

qdrant doesn't support vector DB over Faiss index?

Also, pgvector or qdrant? which is better?

Cool project! Thanks for sharing

Neat! How many images are in the dataset out of curiosity?

how far we've come since https://www.karlsims.com/genetic-images.html

quite a bit, but surprisingly not

nice, we always need more clones and improvements.... hope you get traction.

I never click Pinterest links because the experience is too bad.

I use unpinterested extension in Chrome to remove pinterest from search results, I was annoyed so much at some point. Maybe they're SEO spam is more under control now, not sure.

Reply on news.ycombinator.com