Details
This version was trained with around 1500 images. I will soon make another version with around 2000 images and a different training method outlined here:
All training settings are visible within the file metadata.
I gave up on trying to make a lora that does not affect style. I just cannot figure it out.
Usage
Works best with models that contain NAI as this was trained on NAI.
If you really want more 2D style rather than 3D, add the following to the negative prompt:
(3D, realistic, render)
Description
Details
The provided file is epoch 30/44. There could be a better version but based on my limited testing, epoch 30/33/35 are the most consistent.
All training settings are visible within the file metadata.
This version shouldn't affect the style too much, but it will modify composition a lot.
Usage
Works best with models that contain NAI as this was trained on NAI.
Please use alabasterBlend_v10.safetensors [43381b29f9] or kotosmix_v10.safetensors [0f789fc3c4] or abyssfuta.safetensors [62b68f1181] or AbyssFluffyKemono.safetensors [1f5b74ae12] with DPM++ 2M Karras @ 20 steps minimum + 512 x 768.
If you really want more 2D style rather than 3D, add the following to the negative prompt:
(3D, realistic, render)
Changes
DAdaptAdam
Cosine with repeats (3)
Lower repeats, higher epochs
Increased training data from 253 images to 687 images
Added 2D images to the training data to prevent the style from changing too much.
Double the amount of 2D images compared to 3D images
FAQ
Comments (4)
Love to see an update! Anything we should know about the new version?
I have updated the description with the changes for this v2
Could you tell me about the "abyssfuta" model? I couldnt find this anywhere else.
Sorry for the delay. Here it is: https://civitai.com/models/155279/abyssfuta