An attempt at creating a "turbo" version; it works with 8 or more steps.
The int8 version was taken as the basis, there will be no bf16 version!
Description
This version is less realistic, but cleaner and sharper.
FAQ
Comments (7)
SAM models are awesome, but did you consider publishing them without the DMD2 Lora built in so we can set our own strength for the Lora? I know it's for convenience but if you burn it in for us we lose a lot of possibilities to experiment.
You can simply use the separate DMD lora with negative weight.
@mphobbit Hmm okay but first we'd need to know what weight the author used to burn it to get the original model. Plus I have automated workflows that iterate many models, so with the dozen or so regular ones I already have the DMD2 Lora at the strength I like. Switching it around would break the automation that I have running for hours. More broadly, it would be better if these things were standardized instead of having all kinds of things out there and having to keep track of what each model specifically requires. I've tested like 200 Krea models, Moss doesn't have that many but soon I'm sure it will.
Hmm okay but first we'd need to know what weight the author used to burn it to get the original model.<
Lora's documentation recommends a weight of 0.7 (latest variant) or 1 (earlier) . So something around -0.7 - -1 .
Okay Wow! This is super fun! Great job! :-D
enjoy :)
The best! Thank you!