clip-interrogator VS laion-datasets

It's image -> words, the inverse of stable diffusion.

see: https://github.com/pharmapsychotic/clip-interrogator

I have been using this on a CPU https://github.com/pharmapsychotic/clip-interrogator, I tried a lot of pre-trained models combinations, all are slow.

I'm addition to controlnet, try the clip interrogator to see how clip would describe the image and then use that language in your prompt. You can try the whole image or cropped portions. There is a colab available if you don't want to run it locally.

In my current process, I use CLIP Interrogator to produce a high level caption and wd14 tagger for more granular booru tags. Typically in that order, because you can append the results from the latter to the former. Both tools perform with greater accuracy than the standard interrogators in img2img and give you more flexibility and features as well. You still have to do some manual adjustments, but I generally prefer this process over starting from scratch.

If you are interested in this, definitely see if you can get some of the OSS models running and get a feel for how to interrogate them. Maybe see if you can get some mileage out of the CLIP-Interrogator

Next, I used the lists of artists, flavors, mediums, movements, and negatives that are used for the clip-interrogator and pasted these in the chat and told the bot to categorize them accordingly. As you can only paste up to certain characters in single message (4-5K in 3.5 and 6-8K in 4).

Here’s the specific one I’m using (runs in browser)

I really enjoy using the CLIP Interrogator on huggingspaces, but it is often super slow and sometimes straight up breaks. Now it is possible to locally install it, https://github.com/pharmapsychotic/clip-interrogator but I don't know if its viable to run on a laptop with 6gb videocard anyway.

Not true, it uses the MIT license, which allows for any use including commercial. According to the license you could even sell the Laion datasets yourself if you wanted.

Datasets can be manually curated to produce more aesthetic results if this becomes a real issue. For example, classifiers can predict whether an image is generated or not. You could adapt the process used to create laion-aesthetic[0] to remove generated images.

[0]: https://github.com/LAION-AI/laion-datasets/blob/main/laion-a...

source: https://github.com/LAION-AI/laion-datasets/blob/main/laion-aesthetic.md

My hunch is that is the result of this: https://github.com/CompVis/stable-diffusion#weights

> 515k steps at resolution 512x512 on "laion-improved-aesthetics" (a subset of laion2B-en, filtered to images with an original size >= 512x512, estimated aesthetics score > 5.0

https://github.com/LAION-AI/laion-datasets/blob/main/laion-a... for more details.

What's remarkable is this: https://github.com/LAION-AI/laion-datasets/blob/main/laion-a...

That aesthetic predictor was apparently trained on only 4000 images. If my thinking is correct, imagine the impact those 4000 ratings have had on all of the output of this model.

You can see samples (some NSFW) of different images from the original training set in different rating buckets here, to get an idea of what was included or not in those training steps. http://3080.rom1504.fr/aesthetic/aesthetic_viz.html