Same.energy: Image Search by Similarity | Hacker News Reader

Same.energy: Image Search by Similarity

same.energy

185 pointsmprime137 comments

Unfortunately, this has been abandoned :( Since more than a year, uploading an image or signing up just leads to a server error. Was really amazing when it worked. The creator hasn't been active on Twitter, GitHub, LinkedIn, HN, etc. either.

Yeah. It was a very promising tool. Creator is unresponsive.

Maybe he passed away?

No way he passed away. Running a service like this costs quite a bit of money to host all these images and pay for all their bandwidth. He's still probably shelling out $200+/month on this.

* Just saw that they are served by Cloudfront. Probably not that expensive, but still.

Oh no. I never understood why the upload didn't work but had no idea it was abandoned. That really sucks.

For a while now I've been using it a few times a week to find reference images for drawing. I find it beats Pinterest in terms of speed and relevance for most of my searches.

I really really don't want it to go down.

---

I just browsed the creator's inactive social media pages. I wonder if they've passed away..?

Either that, or got hired by top-secret image-processing government/private agencies requiring complete silence.

I've done as best of a search as I could and can't find anything which indicates they've passed — but they have been inactive for a few years now.

Maybe someone at Tabnine knows...

his instagram is private. maybe there? i hope he's ok.

Still works though (mostly) and still a pretty cool service.

Below is an OSS alternative, originating from same engineering as stable diffusion. There is other awesome work in rom1504's Github I recommend exploring as well.

https://rom1504.github.io/clip-retrieval/ (usually performant)

https://github.com/rom1504/clip-retrieval

The images are very different, and the UX is totally missing.

I searched "boobs" in both (my litmus test). The posted one delivers nice boobs, as expected, very fast. This one delivers medical images of breast surgery, some bizarre pictures, and very slowly (tight budget maybe?). When did we lose our way?

It's great but has all the UX of early DOS programs.

Hehe, well you know, PR welcome, the front end is 500 lines https://github.com/rom1504/clip-retrieval/blob/main/front/sr...

Other people have done a few alternate front ends already

This one is meant to be functional, but could sure be made prettier

The website is long-gone, but I built something similar-ish ~10 years ago[1] using simple histogram analysis. I was surprised at how far you got with just comparing buckets with certain tolerances. Was a super fun project, but got very little traction. Open sourced it here[2].

[1] https://www.youtube.com/watch?v=3WswSywx6TI

[2] https://github.com/dvx/skrch

Sick demo, how would you compare histogram similarity with modern approaches?

Extremely rudimentary :) back then, SURF and SIFT were just starting to get big (they are both obsolete now in favor of more robust machine learning techniques) and I wasn't even doing that.

Just click on an image that you like. Then keep clicking.

Hit the lightning at the top of the page to restart from scratch.

I lost a very enjoyable hour going down random rabbit holes of beautiful images.

Previously on HN ~2 years ago: https://news.ycombinator.com/item?id=24434165

https://hn.algolia.com/?q=https%3A%2F%2Fsame.energy

I'm not associated with this project in any way.

I keep ending up in rabbit holes of ever-increasing cleavages. Must be a bug.

I think that's more like a valley than a hole.

It's not that hard to build nowadays. A vision transformer + vector database will get you 90% of the way there.

[I'm a part of the Milvus community.]

Call me spoiled, but I'm borderline unwilling to trial any service that requires registration in order to.

Fantastic for NSFW content.

I wasn't the only one to discover the same, lol.

ask HN: what is state-of-the-art perceptual hashing tech for images?

perceptual hashing is one approach to image similarity, there are many others. The au courant way of doing ANN search is to use embeddings from the powerful pre-trained models. In this tool the creator uses OpenAI's CLIP.

TIL of "au courant".

the author mentions CLIP, but it looks like he made his own variant. i wish i knew what it was since it seems he is now AWOL

Look at the papers of Andrew Zhai (Pinterest) and how they do it.

Check out SSCD from Meta https://github.com/facebookresearch/sscd-copy-detection

(full disclosure: I am an employee who has worked with this team before internally, but I was not associated with this paper)

We've built a very similar feature across https://icons8.com content.

What keeps exciting me when you upload a photo of a building and find tiny icons with a somewhat similar architecture, or illustrations, or realistic renders/other photos.

What's the math behind this? There's plenty of possible image embeddings/spaces.

The about page mentions, https://same.energy/about

    Same Energy's core search uses deep learning. The most similar published work is CLIP by OpenAI.

He worked at OpenAI, and has written a few blog posts on ML topics. https://jacobjackson.com/

I have been using this site for a while, however it seems the creator has gone AWOL.

You are not able to sign up nor actually upload images as starting points, which is a shame. The creator does not respond thruogh the site nor twitter. Hasn’t been active either.

There used to be a porn website like this that scraped all of the porn subs.

You should try https://github.com/marqo-ai/marqo instead. Does all this and more and still very active development.

so awesome

Reply on news.ycombinator.com