I don't think I agree that LR 8e-6 is better than 1e-5 on OT, on my test dataset. some more skin det
I don't think I agree that LR 8e-6 is better than 1e-5 on OT, on my test dataset. some more skin details on 8e-6, but likeness is reduced
Flower LR means more training necessary but not 1 for all 
Fdepends on dataset
Fi try to balance
Dthe best one I've found is a cosine wave between 0.5e-5 and 1.5e-5. it's got the best of both, skin details and same likeness as 1e-5. that's not so easy to configure though
Dand I haven't tried any other values. it was just an idea but it's surprisingly great
Fsurely configs can be specialized
Fi am trying to make config that wont very undertrain or overtrain on every dataset :d
DI don't know yet, but a wave around a midpoint might even be more general for every dataset than a fixed value. I don't like the cosine annealing with restarts and such, they waste to much steps at the bottom. but this one seems to work because it stays mostly around the LR you want
Fthe config gives very good likeliness and generalization between 100 epoch and 300 epoch that means flexiable
RHello I have a question what is the difference between dreambooth and finetuning in koya. If possible I want to know the details.
Fdreambooth is a technique where you apply regularazation images
Fwith prior loss rate
Fotherwise it is also fine tuning
Fwhen you dont use reg images it becomes fine tuning
RAh and also I'm new to this regularization images I believe that is it supposed to represent a class and what I'm curious a out it is that do the training resolution need to match the resolution of that regularization images? Caz I often see some people use the images generated from the Training model as regularization images. Is that a bad practice or regularization images doesn't matter in quality. Well the question in general is what does it actually do.
Fi use ground truth
Fcollected from unsplash
FMedium
I have been keep testing different scenarios with OneTrainer for Fine-Tuning SDXL on my relatively bad dataset. My training dataset is…
Fbut flux reg images doesnt help
Fyet
RAlright thx
MHey Doc, question about training if you please… In the OneTrainer mega-tutorial (awesome, BTW) you repeatedly mention that, in each epoch, regularization images should be as many as the training images. This makes a lot of sense. However, in the Kohya tutorial (also great) we end up setting the same amount of repeats (1, if I'm not mistaken) for reg and training; still, if I'm not mistaken, you don't say in the Kohya tutorial that we should only use as many reg images as there are training images. The net result is that there's an imbalance of reg vs. training images.
Is it me who didn't understand the Kohya tutorial, or is it just an oversight?
Is it me who didn't understand the Kohya tutorial, or is it just an oversight?
Fin kohya we simply cant use as many as
Fbecause kohya doesnt have 1 to 1 system
Fwe have to use repeat
Fbut still we aim to use
F1 reg image for every step
Fso lets say 200 epochs 15 images requires 3000 steps
Fso we need 3000 reg images
MSo you just select the first 3000 from your (again, super-awesome) 5200-pack?
MBTW, an article you pinned on qwen-flux-wan-such insists that we should always use repeat 10. Do you agree with that? (I know you say rep 1 in the videos…)
Fit uses first 3000
Flogic is
F200 repeat 15 train images
Fit fetches first 3000
Fnope no such thing
Fi dont know where they come such ideas
MIt's actually the second pinned article in the channel, but I do see your point
https://blib.la/blog/mastering-lora-style-training-a-comprehensive-guide
https://blib.la/blog/mastering-lora-style-training-a-comprehensive-guide
Discover the secrets of LoRA style training for AI art. Learn dataset selection, optimization, and detailed captioning for stunning AI-generated art.
Fi explained in latest video why kohya implemented repeating
Fit was for balancing inbalanced multi concept datasets
Fbut he made it poorly that we cant set 1 to 1 reg images like we can in onetrainer
Fbut one trainer doesnt have dreambooth with prior loss rate so not exactly same
HI followed your tutorial on dreambooth for runpod using automatic installation files on patreon. But now I cannot find extensions such as civitai browser+ on extensions download tab
Fcan you link me which tutorial
HYouTubeSECourses
With following this video you can install and use famous Automatic1111 SD Web UI DreamBooth extension on RunPod for full DreamBooth training. Installer scripts are here : https://www.patreon.com/posts/runpod-auto-of-84716845
Click to download scripts
https://www.patreon.com/posts/runpod-auto-of-84716845
SECourses Discord To Get Full Support...
Click to download scripts
https://www.patreon.com/posts/runpod-auto-of-84716845
SECourses Discord To Get Full Support...
HI also want all the controlnets and latest updates on a1111 but maybe it's not possible if I want dreambooth together with it on the same pod
Fdreambooth is broken