I'm trying to use regularisation images lately, but I'm not sure if it makes things better or worse.
Description
PICS: 37
BLIP: 20 - 100
REG PICS: 1258 uncropped pics of males from other models I made.
steps: 10% 127/1270 [02:51<25:41, 1.35s/it, loss=0.136]
steps: 20% 254/1270 [05:51<23:24, 1.38s/it, loss=0.133]
steps: 30% 381/1270 [08:50<20:37, 1.39s/it, loss=0.136]
steps: 40% 508/1270 [11:49<17:44, 1.40s/it, loss=0.149]
steps: 50% 635/1270 [14:47<14:47, 1.40s/it, loss=0.128]
steps: 60% 762/1270 [17:44<11:49, 1.40s/it, loss=0.129]
steps: 70% 889/1270 [20:41<08:52, 1.40s/it, loss=0.138]
steps: 80% 1016/1270 [23:39<05:54, 1.40s/it, loss=0.127]
steps: 90% 1143/1270 [26:36<02:57, 1.40s/it, loss=0.125]
steps: 100% 1270/1270 [29:44<00:00, 1.41s/it, loss=0.129]
FAQ
Comments (2)
I think it looks a little bit like that actor from the Terminator 2 film, but is missing some details.
What do you mean by "regularisation images"?
It´s the first time I hear this term.
...
I often just use HQ-HD images and in various sizes, from 512x512, 768x768, 1024x1024, 1536x1536, and 2048x2048 max-size. I will go no larger, as this is great enough for a lot of fine details.
(I get so many notifications I don't see when people comment.)
You can supply a separate folder of "class" images that are of similar subjects to what you are training, so the AI knows what you are attempting to pull from the "training" images and doesn't veer off.
I don't fully understand it, and they don't seem necessary. But others have said it can dramatically improve their results.
I read of them here: https://old.reddit.com/r/StableDiffusion/comments/11qkbfy/kohyass_lora_finally_improved_the_final_output/
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.