Synne - Fictional Norwegian Russ
The newest GPT image model is the first model I have found that can almost recreate a typical Norwegian russedress properly, so I decided to make a few character LoRAs around that idea for fun.
The LoRA ended up not been as good at making a russedress as i would have hoped for. But im very happy with the character and the level of detail i managed to get.
As usual, I got a bit perfectionist with it. This LoRA was trained for 10,000 steps with GA set to 5, so it has been pushed much harder than most other characters-loras. (20 hours on a RTX 5090). And therefor probably a bit repetetive on simple prompts.
The result is a very robust character LoRA that holds up surprisingly well. It can also be combined with other LoRAs to some extent, where many other character LoRAs usually start falling apart. And ofcourse it can be combined with a unlimited amount of my slider LoRAs.
Much better at retaining the original image compared to most other character loras:
This was mostly made as an experiment since its my first propper character LoRA, but it ended up working better than expected.
Trained on only SFW images.
Recommended range 0.9 to 1.1
(But i have pushed it as high as 2.0 in combination with other LoRAs)
Description
FAQ
Comments (5)
Hello! I hope you're doing well. I just wanted to say that your work is excellent—as always! If it's not too much trouble, would you mind sharing how you achieved this? Perhaps there is a manual or some instructions you could point me to? I would be truly grateful for any guidance. Thank you so much!
Not anything special in this one. As i said 10,000 steps at Gradient Accumulation at 5 witch makes each of the training step compare against 5 ref images instead of 1. So 10,000 steps is in a sense 50,000 steps.
As you can see on my dataset, all my images are 3072 res. And i train at all the resolutions that are available in ai toolkit. fra 256 all the way up to 1536.
I make a dataset for each resolution, and use higher weight on the higher resolutions. i belive i used 1.3 on 1024 and 1.2 on those above.
If you have good settings on your training, the noise graph should be gradually move lower and lower. If its not stable and flatlines for a 1000 steps, you have a bad setup. Mine gradualy went lower the whole 10,000 step run.
@Loraholic Thank you very much