do we have a queue in kohya ss?
do we have a queue in kohya ss?
LWhen I install your caption models, where do they store? I tested some models and I don't like them so I want to get rid of Gigs and Gigs of models downloaded to my disk.
AAlright, I think I have my first epoch, or I will later today.
Can you show me how to resume training?
I want to stop training so that I can make some sprite-style images, and then resume training later?
Can you show me how to resume training?
I want to stop training so that I can make some sprite-style images, and then resume training later?
Fwhat do you mean by that?
Fhugging face cache unless downloaded in local folder
Fcheck this article
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözükara on Patreon
Fi chnaged my cache to this folder
Fthe resume means
Fyou use your checkpoint.safetensors file
Fas a base model
AAnd then I just train it with the same exact training data ?
Fyes if you want to have effect of continue
AThank you, Doctor!
Xlike we can stack up several jobs and it does those one by one?
not sure I explained it well
not sure I explained it well
Fkohya dont have
XI see
TReddit
Explore this post and more from the StableDiffusion community
Fyes i am following
Fcurrently very experimental and hard to use
Fwe need kohya or onetrainer
HThey just improved to 24gb training and Mac support
HGitHub
Flux It's here! Runs on 24G cards using Quanto's 8bit quantisation, or 25.7G on a Macbook system (slowly)! If you're after accuracy, a 40G card will do Just Fine, with 80G cards being s...
Fyep as i guessed
Ffull training will be doable
Fsoon
ELooks like flux training is up and running and the Lora comes out with amazing quality.
Though it is very VRAM hungry min 24GB VRAM. I setup a session on mass compute last night in less than 20 minutes.
Only available on Simple tuner at this stage but works a treat.
Comfy UI included flux Lora support yesterday.
Though it is very VRAM hungry min 24GB VRAM. I setup a session on mass compute last night in less than 20 minutes.
Only available on Simple tuner at this stage but works a treat.
Comfy UI included flux Lora support yesterday.
HNice, what are you training? Style or person?
EI was gonna do person first since it is easier to compare results vs source images
EI'll try style after that works 
Fyes person training is best to see if works
Dfucking hell, things move fast
SI tried training a family member. I can actually tell it captured immense details but doesn’t want to output them in normal image generation settings. It stays blurry. So I crank the CFG way up I can see that it’s exactly right down to each individual tooth or freckle, but high CFG becomes like you spilled paint cans all over it. I can’t find a balance.
SWhen I trained it I used images of differing dimensions but did not enable buckets because I was trying to use the settings on patreon exactly as shown. Should I try again with buckets enabled?
SRtx 4090
Fhello on which model you have trained?
Fif you had different dimensions you have to enable bucketing
For best is make all 1024x1024 with manually cropping
SOk I will enable bucketing and redo 
and I got real vis 4 for the training
and I got real vis 4 for the trainingSYou have an additional sdxl VAE chosen in the dreambooth training settings as well?
Fyes but many models already have it embedded
Xis lokr training slower than lora training?
in lora for 1step it got like 2-2.3sec. in lokr it gets 4-5sec for a 1step
in lora for 1step it got like 2-2.3sec. in lokr it gets 4-5sec for a 1step
Xand not sure what are these
Xtraining a style for animagine
Fi always prefer fine tuning or dreambooth
Fbetter than lora 
HFirst Flux Lora Test running 
Fgood luck
Flooks like using dynamic LR
HIt worked!