Since all these SD3 shenanigans don't seem to find a satisfying end, and people asked for some Pony Loras. Let's give it a try!
As always, keep in mind: I'll try to use the absolute most basic prompt in my samples, so you can see the baseline that's possible (and I absolutely suck at prompting).
Tell me what you think and have fun!
Description
FAQ
Comments (4)
Great lora, do you have a certain method training your lora's? I'm trying to up the quality of my own loras.
It's pretty much straight forward OneTrainer with masked out faces. It all comes down to the quality of your source images.
As you can see in this model with the buckles, I should've spent more time preparing the dataset x(
@CrunchyBanana would it be possible to share your onetrainer settings? Like a json export of some sort
@wxcvbnw I don't know how useful this would be, since the settings change a bit depending on the training and input.
Pick whatever scheduler you like. Go with learning rate 1 for the adaptive schedulers. I still like AdamW8Bit with a LR of 0.002 - 0.0001. All common known schedulers yield good results, there's no magic here.
Ema: Yes
Rank/Alpha: 16/8 (the alpha is just another way to dampen your LR. So adjust accordingly)
Masked Training: Yes
Train both TE with 10-50% of your Unet LR (I'm still experimenting. Since you can deep fry the model pretty quickly with the TEs)
Batch: 1 to as much as your VRAM allows. But it depends again on your dataset. Can't say I found a clear direction here. Some datasets train better with BS1 some are better with BS4. I don't know why, I just try it out and accept the result :D
The rest is pretty much on default.
Make a backup every 100steps and compare them. Some trainings are good in 300 steps, some need 2000. There's no fixed value here.
There's only one rule that applies to all trainings: Have high quality source images. Dont expect a good model with shitty source images.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.














