I was planning to do this myself lol. I was going to use SQLite as the index, and use `sqlite-vec` or something similar to query for similar files directly. I think the only other thing I was planning were more filters, `"positive term" -"negative term"` to be able to negate results, `>90"search"` to find images that match by >90% and some generic filters like `--size >1mb` to help narrow it down when you are looking for a specific image. Quantizing embeddings to make them smaller/faster also seemed interesting but I haven't tried doing it yet.
This script doesn't do search, but it generates keywords for images and places them in the image metadata. You can then search for keywords using something like Diffractor. I will warn though that any AI solution not geared towards NSFW will not give good information on NSFW images, though it may give a keyword such as 'intimate' or 'adult content' which is all you need.
How does CLIP compare to YOLO[1]? I haven't looked into image classification/object recognition for a while, but I remember that YOLO was quite good was working on realtime video too.
CLIP and YOLO work completely differently and have different purposes. CLIP uses transformers and embeddings and can compare text with images for classification. YOLO using a CNN and is trained with bounding boxes on images and is used for image recognition.
Give an image to CLIP and you can compare the similarity between the image and a sentence like 'a vase with roses in it'. Whereas with YOLO you give it an image and get the coordinates of bounding boxes around a vase, and around roses.
We have not decided what to do with it yet. It could be free, paid, or open source. However, the logic code for using semantic search with CLIP-compatible models on Android will be available on GitHub.
Yeah I noticed it was pretty shaky, change the prompt a bit and the result changes a lot. Not very reliable after all by itself, but used in conjunction with other methods.
It’s not that straightforward due to spelling. Does that catch køk? Tihts? P. Nus? For a non English swear word, I had to ask 3 times and about a specific language to finally make that connection.
I read about a company in the 1990s that did that. They went one step further - picking culturally appropriate colors, shapes, numbers, and then permuting the brand names to favorable variations for a country. My (probably wrong) 25 year old recollection was when they introduced subway in China they basically found a way to pronounce it that translated to "this place is delicious". I bet it was in Wired. If not that, probably New York Magazine.
In this program CLIP is being used to create embeddings. A multimodal LLM does something very similar. In this case the language model is not needed because the embeddings are being used to search directly.
I was planning to do this myself lol. I was going to use SQLite as the index, and use `sqlite-vec` or something similar to query for similar files directly. I think the only other thing I was planning were more filters, `"positive term" -"negative term"` to be able to negate results, `>90"search"` to find images that match by >90% and some generic filters like `--size >1mb` to help narrow it down when you are looking for a specific image. Quantizing embeddings to make them smaller/faster also seemed interesting but I haven't tried doing it yet.
Uses only 1 core 100% under linux, can this be changed?
10 images, each ~20 kb size, took more than 10 minutes to index, is that normal without GPU-acceleration?
No it is not normal, I only tested x64/arm64 macs, I will try on linux.
What’s normal? On your Apple silicon.
Wow that’s atrocious performance. So there’s no chance to use this on real photos
Very cool! Here is a similar python version.
https://github.com/spullara/photoindex
Oh and if you want to run something locally on your iphone you can use my app I am still testing:
https://x.com/getrememberwhen
This is cool. Is there also a way to show contents of the image as indexed? i.e. image 1 has cat and dog
There are a lot of tool/apps that let you “search images” but not much that lets you just as easily “read images”
I have wanted to clean up my photo collection for ages and remove any nsfw picture that might hide somewhere.
Would this be able to do that and how likely is it It will see a pc release.
This script doesn't do search, but it generates keywords for images and places them in the image metadata. You can then search for keywords using something like Diffractor. I will warn though that any AI solution not geared towards NSFW will not give good information on NSFW images, though it may give a keyword such as 'intimate' or 'adult content' which is all you need.
* https://github.com/jabberjabberjabber/LLavaImageTagger/
I've been enjoying https://github.com/mazzzystar/Queryable on iPhone
How does CLIP compare to YOLO[1]? I haven't looked into image classification/object recognition for a while, but I remember that YOLO was quite good was working on realtime video too.
[1]: https://pjreddie.com/darknet/yolo/
CLIP and YOLO work completely differently and have different purposes. CLIP uses transformers and embeddings and can compare text with images for classification. YOLO using a CNN and is trained with bounding boxes on images and is used for image recognition.
Give an image to CLIP and you can compare the similarity between the image and a sentence like 'a vase with roses in it'. Whereas with YOLO you give it an image and get the coordinates of bounding boxes around a vase, and around roses.
using same app rclip: https://github.com/yurijmikhalevich/rclip
I have made similar android app for semantic image search, works offline too, still gathering feedback and polishing UI, but it works, if you are brave enough here is it https://drive.google.com/file/d/1tE0cY6umj5h5zCY_Jvaou1M8sCf...
Why yes, I'll download a 695MB APK file from an internet stranger.
Yes, the size is 99% 2 models weights required to run inference offline, there no way around it.
Is there a github link?
We have not decided what to do with it yet. It could be free, paid, or open source. However, the logic code for using semantic search with CLIP-compatible models on Android will be available on GitHub.
In russian, "sisi" is a variation of "tits".
Is there a job/services that confirm that branding is appropriate across different languages? Seems like a non trivial problem to solve.
> Seems like a non trivial problem to solve.
Took me 5 minutes to land this GPT prompt.
https://chatgpt.com/share/66e84c0c-a92c-800a-b452-255d6fe942...
Results:
- Chinese (Simplified) 四四 (sì sì) – sounds like "four-four", which can be associated with bad luck due to the number four in Chinese culture
- Arabic "Sisi" is a common nickname, also associated with Egypt's President Abdel Fattah el-Sisi
- Russian Сиси (sisi) – slang for breasts
- Bulgarian Сиси (sisi) – slang for breasts
- Serbian Сиси (sisi) – slang for breasts
- Croatian Sisi – slang for breasts
You should probably complement with a web search and a wiktionary search because they have all languages on a single page.
Does ChatGPT get anything right, ever?
In Bulgarian the slang is Цици (tsi tsi). I imagine it's near-identical for many other Slavic languages.
Yeah I noticed it was pretty shaky, change the prompt a bit and the result changes a lot. Not very reliable after all by itself, but used in conjunction with other methods.
It’s not that straightforward due to spelling. Does that catch køk? Tihts? P. Nus? For a non English swear word, I had to ask 3 times and about a specific language to finally make that connection.
Tried with "hui" - for ChatGPT this word "has no specific meaning and is safe for use in any language".
Cool LLM application! Might not be enough though.
that's a nice start, maybe does 99% of the job, but to be 100% sure, you still need additional (manual?) checks.
That is sad, the name sisi comes from the sisi empress: https://en.m.wikipedia.org/wiki/Empress_Elisabeth_of_Austria
Not as bad as the Pidora project
https://en.m.wiktionary.org/wiki/%D0%BF%D0%B8%D0%B4%D0%BE%D1...
I read about a company in the 1990s that did that. They went one step further - picking culturally appropriate colors, shapes, numbers, and then permuting the brand names to favorable variations for a country. My (probably wrong) 25 year old recollection was when they introduced subway in China they basically found a way to pronounce it that translated to "this place is delicious". I bet it was in Wired. If not that, probably New York Magazine.
it's also the name of egypt's authoritarian leader
https://en.wikipedia.org/wiki/Abdel_Fattah_el-Sisi
Yes this is the first thing that came to mind for me, strange name choice
Even if that was the intent, which it almost certainly isn't, why would it be strange enough to warrant discussion?
As an American who monitors world affairs, the choice of a quasi-authoritarian junta leader as a name would be quite novel.
It's definitely not a good name in English either
I assume you're reading it as "sissy", but I read it as "seesee", which is fine in English.
I read it as sisi, but which means “thank you” in viet.
Sounds like something one might try to train an AI to do :)
In cantonese it’s what a toddler might call poop
A lot of the prodemently used programming languages and libraries have references to feces if you speak Farsi.
Isn‘t clip superseeded by multimodal llms?
In this program CLIP is being used to create embeddings. A multimodal LLM does something very similar. In this case the language model is not needed because the embeddings are being used to search directly.