My first Lora created on RTX 5090.
Wan2.1 Milf Luna, Trained 130 different images, 6500 steps, on 576 x 1024 and 1024 x 576 resolutions.
This is the first time creating a lora, so go easy on me yeh.
Please show me some of your creation(s) to improve current or next future lora character.
Description
FAQ
Comments (5)
looks like its a bit overtrained. You dont need 6500 steps for character lora training. Only 2500 should do. About 25-30 pictures is sufficient.
thnx bud, i'll remember that :)
there is absolutly no rules about steps, because there is other params. you can do 10000 steps at a low learning rate if you want. Maybe with his settings and his optimiser 5000 should be enough, or if he has the time for 6000 steps, i just can try to lower his actual LR by 10-20% to see if the result looks better at 6000.
Do not agree with random advices about something unknown.
@argotic True, i did train 130 different images at low learning rate, i think it was 0.00003, so 6500 steps was a good choise. I could increase the learning rate to 0.00008 and decrease the steps by half (3200 steps) but the results would be different
@OmegaWPN ok good to know. I am currently struggling with default AI Toolkit settings (3000 steps at 0,0001 with adamw8bits) and it overtrains quickly, destroying the render etc.
I think 6000 of 50% LR speed is better than 3000 of 100%, supposed to be the same, but it's like cutting a square to obtain a circle, more precise you are, more the time you take, but the result is more a "circle". (in my opinion)
I personnaly have a preference for prodigy, because it can do in 3000 steps the result of 6000 (It’s a figure of speech) by starting big and ending with small adjusting its lr dynamically (cosine helps).