v0.2
This one knows about much more things. I also included some general nsfw images, because why not.
Samples were generated with DynamicThresholding (so real, not distilled, cfg) and lying sigmas from Detail-Daemon - its not a necessity, but I'm getting better results with them.
Will upload more samples soon.
v0.1
test run, it more or less learned concepts from the dataset
all samples were generated with higher than standard guidance (5.0 instead of 3.5), 3.5 more often gives washed out colors and broken anatomy
Description
FAQ
Comments (20)
how many concept were you trying to teach it?
Apart from the ones on photos I've uploaded (and thats like a dozen of concepts or so) in the dataset there was a set with general mix of pictures. This set was mostly for it to learn general look, but maybe it also learned some concepts from it. I'm still testing as there was several hundreds of pictures in the dataset
67000epochs?
steps :P
So, how many epochs, how many images, how many repeat?
the rank is much to high or it is 32bit saved ?
100 MB should be enough ;)
its rank 32 in 32bit
@shndibfd847e691 so who can work with lora 32bit? ^^
but for first try ;)
@sevenof9247 to be honest I just left in on default in onetrainer without thinking much about it
@shndibfd847e691 i see ... is onetrainer more stable than kohya ?
@sevenof9247 haven't tried anything else and didn't have any noticeable problems. What stability problems did you observe?
@shndibfd847e691 all fine... so i never tryed o-train ...
btw ... which weight for your lora ?
@sevenof9247 I'm using 1.0. Haven't checked the current version, but like 10k steps before it was already giving distorted outputs at 1.2
thanks!
works really good with character lora
hope you do more nsfw poses
Have you saved the versions with fewer steps?
"all samples were generated with higher than standard guidance (5.0 instead of 3.5), 3.5 more often gives washed out colors and broken anatomy"
Lowered the LoRas strength to 0.6, and it worked without changing the guidance.
When trying to install, InvokeAI returns an error "Unable to determine model type". Does anyone else have the same problem?
I have similar problem.
EDIT: Not solved in InvokeAI. This LoRa only works with ComfyUI.
I tried using here with the default workflow provided but ComfyUI just ignores the lora and doesn't use it. Any ideas why?
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.