This model is able to generate images in a, what I call, dark flat art style. First I tried creating a LoRA out of the 122 images I used, but that didn't work out so I trained a checkpoint which is working the way I want it to.
Triggerword:
(darkflatartai)
I hope you have fun with it and create awesome images in this dark flat art style! Let me know what you think of it in the comments below.
Regards,
KoningWouter
Description
Trained for 5000 epochs on 122 images of dark flat art.
FAQ
Comments (17)
Very cool style. What kind of things did you use as your training material?
@justmaier 122 images in this exact style. I spent hours and hours just editing and cropping the images xD But I'm very pleased with the results.
Can you teach me how you trained this particular style? I always have problems when training special styles
@Aye9494
The way I do it is:
1) Get as much images of your particular style as you can
2) Get rid of watermarks and borders, crop it as good as you can (try to fit important pieces of the artpiece in your 512x512 square)
3) Capture using CLIP Interrogator and edit captions to match the image
4) Try creating a LoRA out of it
5) Test your LoRA
if it doesn't work out
6) Train a checkpoint. I always set it to do 5000 steps with saving every 500 steps
7) Test your checkpoint go from high steps to low until you don't have artifacts anymore.
But note this is my way of doing it, I'm new to this AI think just a couple of weeks in.
I hope this can help you out a bit.
Regards,
KoningWouter
@KoningWouter Thank you for your reply!
My English is not good ToT, I hope you can understand what I mean.
I would also like to ask a question.
For example, this safetensors format file you trained on, which model did you train on?
Because the images generated by safetensors are always distorted every time I train them.
I look forward to your reply! Thanks!
@Aye9494 You're English is fine! ;)
I trained on the base StableDiffusion 1.5 model.
If you finetune an existing model you have the problem that you can overtrain the model. That gives artifacts and distortions.
I don't think I had good results finetuning an existing model. But that can be me being new to this scene.
A thing I can think of what you can do is first train on de SD1.5 base model. Then merge your model with another model you want to finetune. Maybe that gives better results.
Good luck!
@KoningWouter Thank you for your reply and have a nice life!XD
Very cool style. Looking forward to digging into this. GJ.
Did u try extracting a Lora from this Checkpoint? Should work. Checkpoints way to huge just to get an art-style
@epinikion Yes I tried making a LoRA first but it didn't give me te results I wanted. I don't know how I can extract a LoRA from this checkpoint file. Could you give me a hint?
@epinikion I managed to extract the LoRA but it had the same problems as my earlier trained LoRA. It quickly neglects the style and puts colors in it. This was the reason I went for a checkpoint.
@KoningWouter i'll give it a try then
@KoningWouter it works on some prompts with high weights. Did u train on the 1.5 original model? It seems not, could also explain why your Lora training doesn't work. Or its way undertrained.
@epinikion yes I trained on sd1.5
@KoningWouter then its undertrained pretty much, don't know much about checkpoint training tho
@epinikion it's trained for 5000 epochs. Could it be still be undertrained?
@KoningWouter don't know how checkpoint training works in detail, but as i see the results from Lora and from the sample prompts, all weight adjusted