Re-arranged UI to make more sense from trainer pov.
Added bunch of features from my older fork that were tested thoroughly by few trainers on it, and provided benefit, in one way or another.
Those features are:
min snr (idk why it disappeared in rework, @saunderez added it while back)
debiased estimation loss
multires noise
ip noise gamma
loss curve offset (my own thingie)
Scheduled noise offset (Shown to be especially beneficial in longer trainings)
Randomized noise offset
Issue ticket number and link (if applicable)
Not applicable
Checklist before requesting a review
[ ] This is based on the /dev branch (Or a fork of it)
No, because main is most up to date. Im probably too stupid to make it fork dev at this point, idk how to use github, lmao.
[x] This was created or at least validated using a proper IDE
Visual Studio?
[x] I have tested this code and validated any modified functions
I did test this code, and training is successful.
I cannot say that XL training is working, because i don't have 24 gig gpu to test, but it seemed pretty identical, so i'll assume it works.
[ ] I have added the appropriate documentation and hint strings if adding or changing a user-facing feature
Ah, eto... Bleh <(e^e)>
Here is example, just so you can be sure i actually did train xD
All are underbaked, because i used twice less steps than needed. Loras don't work in post-XL rework after all, gotta do dreambooth...
First is all features on, second is pre-noise-stuff added and fixed min snr, third is zero min snr
P.S.S. Yes, i have no idea how to use version control and updating whole files.
Here is some screenshots of how it looks right now:
First, order of tabs(I touched only Parameters UI):
Let's look at accordions now.
Intervals were not changed.
Lora was minified a bit with row.
Batch size and image size were merged in single tab, alongside with their utilities.
Learning rate tab was only moved.
Performance was strongly reworked, and extended with new features.
Prior loss was minified too.
Saving, Image generation and Extras were not touched.
Now, whole basic UI is really small, but i probably f'ed up here and made it hide some relevant stuff? Idk, i don't recall how it looked before that tbh.
Describe your changes
Issue ticket number and link (if applicable)
Not applicable
Checklist before requesting a review
Here is example, just so you can be sure i actually did train xD All are underbaked, because i used twice less steps than needed. Loras don't work in post-XL rework after all, gotta do dreambooth...
First is all features on, second is pre-noise-stuff added and fixed min snr, third is zero min snr
P.S.S. Yes, i have no idea how to use version control and updating whole files. Here is some screenshots of how it looks right now: First, order of tabs(I touched only Parameters UI):
Let's look at accordions now.
Intervals were not changed.
Lora was minified a bit with row.
Batch size and image size were merged in single tab, alongside with their utilities.
Learning rate tab was only moved.
Performance was strongly reworked, and extended with new features.
Prior loss was minified too.
Saving, Image generation and Extras were not touched.
Now, whole basic UI is really small, but i probably f'ed up here and made it hide some relevant stuff? Idk, i don't recall how it looked before that tbh.
