Versions are completely different files with new prompt lists, not improvements over previous lists. Check version notes for more details.
I make a lot of models, mixes, and LoRAs and often want to explore their limits, generate showcase images, and generally generate a lot of images without babysitting prompts. But I've had issues finding really great wildcard files that will generate a huge range of random images. So, I decided to create my own.
For version 1, I started by pulling just over 36,000 images from various datasets. Quite a few were Danbooru, but they also came from other sources. I got a mix of anime, realistic, and artistic styles. Then, I used a WD14 captioning script to create captions for all 36k images. I ran another script to combine all the tag files into one massive file with 36k. I also either put a few tags on a blacklist or edited out a few more tags -- so these prompts shouldn't contain common Danbooru tags associated with loli, children, rape, etc. Finally, I used NSFW sorting tags to create individual wildcard files that contain only general, sensitive, questionable, or explicit prompts.
There are individual files for each category and one combined prompt file. Keep in mind the files contain entire prompts, these wildcards aren't really made for adding randomness to existing prompts. You can either use them with the dynamic prompts plug-in or load them as a prompt list file.
A word of warning — there were 36,000 prompts to screen and I used AI to create the prompts. Though I used precautions, I can’t guarantee every prompt has had every keyword removed that could generate unfavorable images. Use some personal discretion before posting generated images.
Description
A collection of almost 200,000 NSFW random prompts for use in testing new models (or just generating random NSFW images.
FAQ
Comments (11)
The 2.0 version is such a big files size for a text document that my computer freezes, as well as A1111 when trying to load a wildcards file. Are you able to uploaded a separated version?
Sure, I can. The prompts are much more details and there are many more of them. What size would be better for you?
I uploaded an archive of 6 smaller text files instead of the one giant one. Sorry about that. I have 80G of RAM and forgot that opening files that large isn't possible on every machine.
My advice on this: distributing such a large corpus of text(s) in .txt format is not a great approach. A serialized format like .json would be more appropriate. If you were to draw from that corpus programmatically, you'd probably serialize it anyway before using it, so providing it in an already serialized format is the way to go. It would also prevent users from accessing the data through volatile software like MS Notepad, which I assume BinaryBottleBake tried. As described by them, that program is not intended for accessing large amounts of data and will result in a bad user experience. (Side note: use Notepad++ as a general purpose text editor instead!)
@xnapx I was trying to be more accessible by putting it in .txt, but I neglected to think about how crappy it would load in MS Notepad with 8 or 16G of RAM. I use Notepad ++ for everything. I have no issues loading 200k lines in as wildcards, but it's not the standard.
@Dark_infinity 80G?! Bro share your setup details, please.
@haasetomoko203 Not all that impressive of a system, older gen 11700 CPU, but its the system RAM that does it -- 128 gb goes a long way for opening and using huge files. And with DDR4 less than $100 USD for 64gb kits, seemed like an easy and cheap upgrade. Its also great for training, I can cache a huge number of latents to system cache and its great for pre-processing images, as I can load thousands into system memory all at once.
Can you make it a nested format?
A nested format would require some programmatic intervention, but it would be an interesting way to organize these in a condensed format, making adding new variations much easier. There is no way I can get even a fraction of the variability from 200k prompts. I may try firing up a local LLM and have it analyze the structure of the prompts to determine some syntactical categories that the tagger used and see if it can sort the tags by what it finds. The problem is that tag count-wise, you will lose a ton of unique tags, as I don't think any LLM I can run will have a high enough context size to capture tags that only occur a handful of times. If anyone reading this happens to know a model I can load in less than ~20G VRAM with a large enough context size to analyze the prompts, let me know.
Hello,
Thx for your pack. It's very good job. The prompts are perfects!
Yes! Thank you! Generating totally new images!