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

Show HN: I made a Pinterest clone using SigLIP image embeddings(mood-amber.vercel.app)

98 points by verse 2 years ago | 21 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.

yorwba 2 years 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.)

verse 2 years ago | |

ah! thank you for pointing this out. will fix

lulzx 2 years ago |

Also, check https://same.energy/

wucaworld 2 years 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?

jkcxn 2 years 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

verse 2 years 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!

omeze 2 years 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

Tiberium 2 years ago |

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

jarebear6expepj 2 years 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.

gammalost 2 years ago |

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

GamerAlias 2 years ago |

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

verse 2 years 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_CoDE 2 years ago | |

qdrant doesn't support vector DB over Faiss index?

Also, pgvector or qdrant? which is better?

squam 2 years ago |

Cool project! Thanks for sharing

Yenrabbit 2 years ago |

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

convolvatron 2 years ago |

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

quite a bit, but surprisingly not

ijhuygft776 2 years ago |

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

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

karolist 2 years 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.