Show HN: I made a Pinterest clone using SigLIP image embeddings (opens in new tab)

(mood-amber.vercel.app)

98 pointsverse2y ago21 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.

21 comments

yorwba2y ago

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.)

verseOP2y ago

ah! thank you for pointing this out. will fix

lulzx2y ago

Also, check https://same.energy/

wucaworld2y ago

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?

jkcxn2y ago

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

verseOP2y ago

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!

1 more reply

omeze2y ago

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

Tiberium2y ago

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

jarebear6expepj2y ago

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.

gammalost2y ago

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

GamerAlias2y ago

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

verseOP2y ago

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...

ReD_CoDE2y ago

qdrant doesn't support vector DB over Faiss index?

Also, pgvector or qdrant? which is better?

squam2y ago

Cool project! Thanks for sharing

Yenrabbit2y ago

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

convolvatron2y ago

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

quite a bit, but surprisingly not

ijhuygft7762y ago

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

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

karolist2y ago

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.

j / k navigate · click thread line to collapse

21 comments

yorwba2y ago

verseOP2y ago

ah! thank you for pointing this out. will fix

lulzx2y ago

Also, check https://same.energy/

wucaworld2y ago

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?

jkcxn2y ago

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

verseOP2y ago

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!

1 more reply

omeze2y ago

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

Tiberium2y ago

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

jarebear6expepj2y ago

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.

gammalost2y ago

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

GamerAlias2y ago

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

verseOP2y ago

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...

ReD_CoDE2y ago

qdrant doesn't support vector DB over Faiss index?

Also, pgvector or qdrant? which is better?

squam2y ago

Cool project! Thanks for sharing

Yenrabbit2y ago

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

convolvatron2y ago

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

quite a bit, but surprisingly not

ijhuygft7762y ago

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

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

karolist2y ago

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.

j / k navigate · click thread line to collapse