We have updated our quality standards found under our uploading guidelines.
This was done mostly because modern AI generations are of much better quality now than they have been in the past, and we would like for this archive to focus on displaying the best work, not just any and all work. Please take a moment to review the guidelines before uploading going forward now. Older submissions will be exempt for this at this point in time, but keep in mind in the far future we might clean those up as quality continues to improve.
YaLonip
Memberwilling to spill the secret as to how you get this style?
Hamgas
MemberSure ! Here's the starting metada for this one. I think I only changed the prompt when I inpainted the background characters.
Updated
YaLonip
MemberThank you very much lol. i have a question. i cant seem to find any info of what a vpred model is or how it's different form a normal model. do you know? and also what does the 'break' do in the prompt?
Updated
Hamgas
MemberIf i'm not mistaken, Vpred is a method of training that correct a flaw in the base training of SD 1.5. Base SD1.5 doesn't have good a dynamic range of contrast (getting nights that really look dark) without using tricks or Loras to bypass the issues. Vpred models can do it by themselves.
To use a vpred model with Automatic1111, you need two things :
- A yaml file (a setting file) that can usually be downloaded alongside the vpred model. It's placed in the model folder with the models checkpoint.
- The cfg rescale extension with cfg rescale usually set to 0.7 (you can play with it, here i used 0.5)
SD1.5 convert the words in the prompt in tokens. It can only deals with the tokens in chunks of 75 and usually pay more attention to what's at the begining of a token chunk.
The BREAK keyword in automatic1111 allows you to fill the remaining tokens of a chunk with empty ones. Like that you can control where new chunks of 75 tokens start in your prompt (right after a BREAK).
I (and others) found it usually best to separate concepts in different BREAK. I usually do something like this:
Though, I'm not sure if it's still used or if i'm just outdated. I know it doesn't work with all programs used to generate images with SD.
TerraRaptor
MemberI also use BREAK. I do it like:
species and gender + pose/sex position + the main style BREAK
clothing details/expression/color details/location + reinforcement style
I try to keep it at just 2 chunks. Start stretching the AI thin beyond that (depending on the model, some can handle it tho).
Also a secondary bonus to using BREAK is that multi-token concepts don't coincidentally fall right on the chunk borders and get split up.
Login to respond »