Description
FAQ
Comments (10)
This is a technical question, but what changes have you made between v1.0 and v2.0?
v2.0 is very good, but I think it is wonderfully complete at v1.0.
I have started making Lora myself and would like to refer to it if you don't mind.
The v2.0 has:
-Better contrast
-More accurate details
-Greater consistency of the outfit
-Greater compatibility with other pony based checkpoints
So it's more of a QoL change than a complete rework.
As for the changes I made, I cut down the number of images I used from 100 to 50, while using a stricter personal criteria for image selection.
@symplexity Thanks for your reply!
The making is deep...
Nijisanji Lora, I'm so glad!
@blackbird0006 I'm glad to be of help
@symplexity I'm surprised that it's not just a matter of having a lot of images.
@blackbird0006 Depends on the optimizer you use. I use Prodigy, so 40-60 is ideal. But I have made LoRAs with 500+ images using AdamW8bit in the past.
@symplexity I didn't know the details of the difference between the optimizers, so I kind of used them.
I see, I'll have to look into this.
Thanks a lot!
for the Inui lora, i think the hardest part is to maintain the small inu hair ornament's face. The small face still breaks sometimes in v2 (squeeze long, have 3~4eyes on face, or not even face...etc), but its much better than v1.
@AmadeusAI Exactly. These minor details are often hidden so only half of the dataset may contain them at a time.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.