31st January 2023 Update:
djzAirlockV15 released!19th January 2023 Update:
djzAirlockV21-512-inpainting released!
you will need the 2-1-512-inpainting.yaml config, so grab that too!
2.1-768 model: Example merging Rick-Roll, JunglePunk & Airlock:
credit: RickInversion
These "strong style" Models are intended to be merged with each other and any model for Stable Diffusion 2.1
I recommend merging with 0.5 (50/50 blend) then using prompt weighting to control the Aesthetic gradient.
example merged model prompt with automatic1111:
(Airlock:1.2) (yourmodeltoken:0.8)
if you drop the "djz" and the "V21" what remains is the token you need to call up the concept in the model. All examples shown were the Raw Token, no other words. Tokens are case sensitive and in almost all models it will match the filename.
It is possible to merge these models with each other using a different value. It is possible to pair models and then merge those resulting models. In this way we can blend abstract concepts together and then weight the tokens to achieve the result we may wish to create. Of course to eliminate all those tokens, you can simply train a new custom model from the outputs, which means you are back to a single token.
A video explanation will follow, but for now the above explanation should do. We are focused on getting as many style/aesthetic models into artist hands to enhance the creativity already at their finger tips.
Art Freedom for all!!
[all original artwork used for training with full permission from Drift Johnson]
Description
SD 1.5 version
FAQ
Comments (10)
Ty for the 1.5 versions of your work! Just curious; any idea why the test images on this specific one came out soft? Regardless, thanks for providing multiple format options!
we were testing a soft focus technique by oversizing the dataset in training, if i recall correctly. All our models were designed for 2.1-768 at 1024x1024 for extra sharpness. I believe that we were training this model at 768x768. Its possible this might get a revisit if the 512x512 dataset sharpens it up, but there is no way to know for sure as SD1.5 produced less sharp results. We wanted to make an Airlock model that could be used to be inpainted over the top with other sharper models and give a nice effect for those backgrounds. Each model had a hidden "test feature".
Please, would you tell me how you create a 2.1 inpainting model? I've tried the analogue of 1.5 merge trick (using 2.0 base and 2.0 inpainting models), but it doesn't work (only produces white clouds). I've been searching through the internet to find some info, but couldn't find anything...
Also, you mentioned "the 2-1-512-inpainting.yaml config". Where can I get it? I literally googled "2-1-512-inpainting.yaml" and this page is the only result. Did you mean the "v2-inpainting-inference.yaml"?
if you visit the CivitAI discord I wrote and explainer in the resources section with a diagram showing how to make inpainting models.
You use "add difference" merge mode.
A = Target model (1.5/2.1-inpainting model)
B = Your Model you want to make an inpainting model
C = The base model you trained your model on (must match the version of A)
Use 1.0 for the Alpha value.
additionally, there is no Inpainting for 2.1-768 only for 2.1-512
"SD V2.0-512-inpainting.yaml"
https://civitai.com/questions/116/sd-2021-model-wont-load-config-yaml-files
https://huggingface.co/webui/stable-diffusion-2-inpainting/tree/main
"512-inpainting-ema.yaml"
When you are on the inpainting models and then look on the right side and open the "FILES" tab then you will see 2 files. one is the safetensor, the other one is the yaml config file.
"if you drop the "djz" and the "V21" what remains is the token you need to call up the concept in the model. All examples shown were the Raw Token, no other words. Tokens are case sensitive and in almost all models it will match the filename."
I noticed some of your prompts have different capitalization from the file name--for example, you used spirecity instead of SpireCity. Was that a different version?
Also, what checkpoint should I have loaded/selected in Automatic1111? One of the checkpoints, or does it not matter?
ah, this is because i trained them all with the capitalisation and later it was proven that the tokeniser did not care about capitals, i wanted people to get the exact same trigger word. The advice i gave was accurate, but as it turned out the capitalisation is not strict as i thought it was initially (this text is written in january and the same is on most pages). In the case of spirecity, you do not need the capitals. i never edited the descriptions because it still works and there were literally hundreds of those to edit.
if you are using djzAirlockV21, you need to use that model.
However check out my account as i retrained all these into a single model Johnson Diffusion Zenkai, which is also available as a Lora and a Lycoris on V2.1-768 (or lora only for 1.5)
@driftjohnson thanks! I read the style triggers txt and couldn't find CyberCity Aerial, Dieselpunk City, etc--are these still included just under different trigger words?
@Servo_Scribe cybercity is just cybercity alone, there were 3 variants originally, but i think they are just one model in JDZ. dieselpunkcity is all one word and should be in there, but there are some styles (note: 230 is not all of them) that were removed by design.