40 epoch is amazing
40 epoch is amazing
MCan you detail how many repeats, epochs, total steps and batch size please?
F256 images, batch size 7, epoch 40 = 256 * 40 / 7 steps = 1462
Ffine tuned on RTX A6000
F
MThis is Dreambooth finetuning on Kohya, right?
Fdreambooth in kohya for flux
F256 images
F256 images is a total overkill can be reduced a lot
MYou didn't mention repeats, so your folder was 1_ohwx man, right?
Falways 1 repeat
Ffor flux
MWell ,maybe that's where I went wrong, I did use repeats
F
Fwhy?
AHas anyone had issues with the vae when trying to finetune in kohya? It looks like it's loading the full flux dev checkpoint as the vae rather than the vae I'm pointing it to in the config...and so it fails to generate latents. I think I can work around this by just not caching to disk and by turning off "use latent files".
Just wondering if anyone has seen this. kohya_ss is all up to date.
Just wondering if anyone has seen this. kohya_ss is all up to date.
MI thought I was using your settings, but maybe I increased the repeats in addition to the epochs to get more steps? I can't recall 100% but I guess it was a mistake on my part
Fwhat do you train?
Fflux_?
Frepeat always 1
AYes, Flux using one of your presets on 24gb,
Fthen you need to give vae path
Ffrom your pc
AYep, I am, and I can see that it's in the .toml file under ae=.
A
A
Fyour vae is inaccurate
F
Fuse my downloader downloaded vae
AI have a few in that directory, I also have ae.safetensors which is straight from huggingface
Ftry mine
ARunning the downloader now just to be sure though!
Fif something not working use mine
AThis is the issue I'm seeing...I'm pointing it to your vae now, but when kohya runs you can see in this output that it keeps trying to use the flux checkpoint as the vae. For some reason it's not use the vae in the gui or in the toml
Fshow me gui screenshot please
A
Falso did you install kohya via stability matrix?
Fthis is loading flux dev model as vae
F
F
Fget fresh config
Ffrom zip file
Fput itno c
AKohya was installed via your scripts, my flux models just live under stability matrix. I just updated it to point to the flux model from your downloader but same thing, it's still setting the vae path to the full model. I'll try a git hard reset just to make sure kohya isn't stuck on an old commit
Fwell thjis message
Floading inaccurate
F
AI know, that's the strange part. I mean it's definitely the issue, but the toml file shows the ae being set to the one I'm specifying, not that path. So it's confusing why it's suddenly switching it at runtime
AIt's similar to this issue: https://github.com/kohya-ss/sd-scripts/issues/1535
You can see in his screenshot that the VAE is listed as the full checkpoint
You can see in his screenshot that the VAE is listed as the full checkpoint
GitHub
Getting this error when I try to start a finetune training Happens when I try with the Full Flux Dev as well, what am I doing wrong? Am I using the wrong base model somehow? 2024-08-30 12:28:38 INF...
Fif you upgrade gold membership i can connect and check out if you need
