Training does not require a dataset, only words.
Trained in 2400 steps (about 7 hrs on 2 a6000s) using only text with:
https://github.com/ntc-ai/conceptmod
Training prompt for the first 2000 steps:
"=trending on artstation, 8k, ultra hd|boring=exciting|drab=captivating|1woman%generic woman wearing generic clothes:0.003|1woman%woman wearing many unique tasteful accessories and a cute outfit:-0.003|@monochrome--|@black and white--|@text--|@written words--"
Pulls the unconditional towards "trending on artstation, 8k, ultra hd"
ReplacesPulls "boring" to "exciting" and "drab" to "captivating".
Update: this doesn't replace. It pulls towards. Replace needs an additional % term, see the readme.mdReduces the occurrence of "monochrome", "black and white", "text", and "written words".
Slightly blends 1woman away from "generic woman wearing generic clothes"
Slightly blends 1woman towards "woman wearing many unique tasteful accessories and a cute outfit"
Update: I think these terms might be backwards. LOL
Training prompt for the next 400 steps:
"@#|1woman=ugly manly sad wretched:-0.1|1woman=woman wearing many unique tasteful accessories and a cute outfit|nipples--:0.1|cleavage--:0.01"
Freezes the unconditional concept using the "@#" operator.
Reduces the occurrence of "nipples" by 0.1 and "cleavage" by 0.01.
Technique based on https://github.com/rohitgandikota/erasing
Model based on https://civarchive.com/models/13565/criarcys-fantasy-to-experience
Description
FAQ
Comments (8)
I'm not sure I understand. I've read the description and the github already. Would you mind explaining it further like I'm 5?
Are you changing the base understanding of the model so it's more biased to certain results? What's the goal, exactly?
Alright! So, this project(SD) is like a magic coloring book that can draw pictures by itself. But sometimes, you want the coloring book to draw more of your favorite things or change the way it draws.
By telling the coloring book some special words, you can make it draw more of what you like or change how the pictures look. For example, you can ask the coloring book to draw more pictures with bright colors or more pictures with your favorite animals.
The goal is to make the coloring book listen to you, so it draws the pictures you want to see. This way, you have more fun and enjoy the pictures the coloring book creates for you!
-- gpt 4
@ntc Lol, I certainly asked for it. Thank you.
Would I get anything different from simply using the words I like, adding some weights? Is the intention that it naturally skews one way so I don't have to keep using certain words to get results?
@Polygon It depends on what you are doing with the words. If you are doing 'concept++', the concept gets exaggerated across all images I think. Adding 'concept' to the prompt after training on 'concept++' exaggerates even further.
Concepts don't exist in isolation though, close by concepts may be modified too.
I'm still exploring it, gpt-4 wrote the library (with guidance).
@ntc LOL I was going to do the same, copy paste the github description in gpt and have it explain to a 5 years old :D
@2BlackChicken I tried for 10 minutes it but GPT did better immediately. o_O
Do you think this training process would work with only 24GB of Vram? Obviously would take longer.
Hey vram is not the main bottleneck. I'm training with 4 terms and it's using 7GB I think.
You may need to change your devices to 0,0 instead of 0,1. (see train_sequential.sh)
I have only tried on this hardware so feel free to share any errors or experiences.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.