STOP: This is a lora/embed or even checkpoint train on OUR art, fear not we own the dataset.
We've done this now 8-10 times. We want to see what creative ways either you can use our data in a new content OR re-train and better tag the lora in full force.
Basically the base model can be anything you please for training. I usually use NAI/Any Lora to complete the full style, but sometimes i've used SD Base.
However: This is up to you how you do it, just the only stipulations are you don't steal our art for commercial gain (the dataset) and you don't make illegal content with it. (This is all laid out in the lisc document attached).
If you ARE going to use something besides SD 1.5 or SDXL please make sure ONE of your loras or models is SD based so it's useable.
YOU CAN train the following:
Embed
Lora
Full Model
Realistically YES this is a LORA creation, but go nuts, go wild.
Suggested trainings: Full style for at least ONE of the items, and an item or two that INCLUDE a large portion of the same dataset but merged with other data to make a different style.
Please make sure you read the full text file, if you're too lazy basically that reads;
1. Don't do anything against Civit ToS, or anything against local or international laws and regulations even after the bounty with the dataset
2. I don't own your outputs OR your trained file.
3. However credit would be amazing on your final product when you upload it.
4. Don't STEAL the dataset as your own for commercial gain.
5. DO use the final product/trained item for commercial outputs, but do not sell the model itself. (AKa: Commercial image generation is totally fine)
6.Ignore the mac OS thumbnail folder, make your own NEW folder with the images and tag it your best without the gunk the zip file does.
7. MICROSOFT MIGHT CLAIM ITS A VIRUS BECAUSE OF THE THUMBNAIL FILES - trust me, MACS ARE NOT A VIRUS, it's not a virus I swear it's just images and the ".whatever" type files it spews out.