How To Properly Setup OneTrainer Concepts And Some Other Options - Quick Tutorial Before Big One

[@SECourses]

OneTrainer Stable Diffusion XL (SDXL) Fine Tuning Best Presets : www.patreon.com/posts/96028218 The Very Best OneTrainer Workflow & Config For SD 1.5 Based Models DreamBooth / Full Fine Tuning : www.patreon.com/posts/97381002 Full Workflow For Newbie Stable Diffusion Trainers For SD 1.5 Models & SDXL Models Training With DreamBooth & LoRA : www.patreon.com/posts/full-workflow-sd-98620163 5200 Ground Truth Regularization Images for Man & Woman (each one has 5200 images) : www.patreon.com/posts/massive-4k-woman-87700469 OneTrainer repo : github.com/Nerogar/OneTrainer

[@1Know1tHurts]

Waiting for a public tutorial. Thanks for your videos, man.

[@SECourses]

I am waiting massedcompute to give me VM we agreed to make it

[@SECourses]

It is published : kzbin.info/www/bejne/ZqWYnWl5hZ6YeMkfeature=shared

[@reapicus557]

@@SECourses You are a gentleman and a scholar! I am excited to dig into this video. :D

[@SECourses]

@@reapicus557 thank you so much

[@hakantarget5328]

Good solutions, great work Keep on watching u

[@SECourses]

thank you

[@zappazack]

Thx a lot for this brief intro waiting for the full tutorial!

[@SECourses]

thank you so much. it is in production. i am waiting massed compute to put more A6000 gpus into the pool

[@Royvcent]

At 2:49 you say that enabling 'Latent Caching' will increase the speed and reduce VRAM. The way I learnt it is that it does increase the speed but as a tradeoff it also increases VRAM, not reduce. So this option won't be good for everyone. Edit: seems I was wrong on this one! It does save VRAM space.

[@SECourses]

i asked this to the developer. actually onetrainer does latent caching in any case as far as i understood. so either you pre cache into disk with this way, or it will recache at every epoch which will make it very slow. also caching is used to reduce vram usage as far as i know and speed up. i will ask your comment to him on discord now

[@Royvcent]

@@SECourses reading upon it more I believe you are right. Sorry for the false statement and thank you!

[@SECourses]

@@Royvcent he also replied me from discord that my comment is accurate

[@Dany-w3g]

Thanks, subscribed to patreon and will support you for a long time, you have great content.

[@SECourses]

awesome thank you so much

[@orboan]

Training images can be mixed of different sizes? Or is it mandatory all have to have same size, like 768x768?

[@SECourses]

I strongly recommend 768x768. once you have a base model with such training, you can try bucketed training, different resolutions and compare results of both

[@KnutNukem]

Amazing, thank you

[@SECourses]

you are welcome thanks for comment

[@marcusmeins1839]

can i use one trainer with a gtx 1050 TI? i tried using one trainer but now it is too slow .

[@SECourses]

it will be slow with 1050 TI. how much VRAM you have? you probably should rent a cloud service for training. massed compute is great

[@p_p]

OneTrainer is downloading an incredible amount of model.safetensors of big sizes, wtf is going on? i can't even find were this files are going in my hard drive

[@SECourses]

the files get downloaded into HF cache you can learn more here : kzbin.info/www/bejne/qJu7pH1oadCEh8k

[@vityapelevin]

when I set your settings on float32, my model does not work in sd, writes "NansException: A tensor with all NaNs was produced in Unet. Use --disable-nan-check commandline argument to disable this check." I put a check mark in the settings for "Upcast cross attention layer to float32", but it doesn't work either, it gives out a black square. nvidia 4060ti 16gb, I trained on the basic 1.5 model

[@SECourses]

you mean you tried SD 1.5 base? can you try hyper realism v3?

[@vityapelevin]

@@SECourses I train on v1-5-pruned-emaonly.safetensors so that I can then mix (Checkpoint Merger) my model with some other popular model. it works fine only when I set the default settings from the list.

[@SECourses]

don't use it is very bad one.

[@ollyevans636]

what platform should i use to run my code on mac? when i click 'install.sh' on my mac, i don't know which software i should use in order to run onetrainer? thanks!

[@SECourses]

I suggest massed compute with our coupon - 31 cents per hour : m.kzbin.info/www/bejne/ZqWYnWl5hZ6YeMk

[@Sevan-k4y]

Thank you very much, good job

[@SECourses]

thank you so much for comment

[@zappazack]

1 question about the reg pictures. In the video u said 768 res for SD1.5 training but should it not 512res? What is about the 512x768 res images how they are used for training?

[@SECourses]

it depends on used base model. my used base model was able to handle 768 very well. 512x768 should work good. is your training images also 512x768?

[@zappazack]

Only 1 set is 512x768 the other is 512x512. I am currently doing a research about training with higher res images and 9:16 ratio​@@SECourses

[@SECourses]

@@zappazack i see. i prefer single resolution but sure give it a try

[@zappazack]

Is there any reason why u are reluctant to use the bucketing training?@@SECourses

[@zappazack]

Why u are so reluctant of bucketing training I am asking coz at ur huggingface u have a lot of various regularisation images@@SECourses

[@streamtabulous]

how do you change training steps when not using second traing set, plus doesn't have repeat like yours, has balancing?

[@SECourses]

Balance is new feature. So instead of % you type any number. Training steps same with or without. So let's say you have 16 training images make and using reg images make balancing 16 and it will use 16 reg images every epoch

[@MoonDoes]

Is there any benefit to having "TensorFlow" installed? When starting OneTrainer, it mentions that it will be limited due to not finding a TensorFlow installation. I installed OneTrainer through StabilityMatrix, could some packages have been omitted during the installation?

[@SECourses]

yes don't install via StabilityMatrix. i don't know if missing TensorFlow would make difference. but as i said don't install via StabilityMatrix. OneTrainer already has one click install. just make sure that your default python is 3.10.11

[@MoonDoes]

@@SECourses thank you

[@SECourses]

@@MoonDoes you are welcome

[@Grifter]

Does the amount of training images and regularization images effect training time with fine-tuning?

[@SECourses]

total time of course but not step speed. you know like x/y second / it

[@vityapelevin]

It turned out to create a photorealistic model of different eyes, good results in txt2img, I needed it to restore photos via img2img, and in img2img it works poorly, produces almost the same image in quality as it needed to be restored. What settings should I set for a good img2img during training?

[@SECourses]

we dont use img2img during training. improve your dataset quality if your results are not good enough. also this is the big tutorial for onetrainer : kzbin.info/www/bejne/ZqWYnWl5hZ6YeMk

[@brandonwells2401]

3 seconds in when starting the training --------> error:check the console for more information

[@SECourses]

please message me from discord and let me know the console message : discord.com/servers/software-engineering-courses-secourses-772774097734074388

[@vityapelevin]

The realistic vision model says it was created on 1.5, how do I create the same cool model to generate different photorealistic people, detailed faces, landscapes and objects? no matter how hard I try, nothing works out, it's just ugly, although I tried very hard on the dataset.

[@SECourses]

i think you need a private lecture for this. or you need to watch my all videos historically. this is a big task

[@zappazack]

Just to clarify all training images have to be resized right?!

[@SECourses]

if you do that yes better

[@nowshinnur]

will there be a kaggle tutorial?

[@SECourses]

for kaggle not possible you need desktop. i am working on massed compute tutorial. a6000 gpu is 31 cent per hour

[@Vigilence]

Can you cover art styles as well?

[@SECourses]

yes i plan to cover. i did few good style trainings.

[@MS-gn4gl]

Having seen your comparisons on the patreon of the outputs vs kohya - I've got to say there's a level of "what you look like now" that's much more apparent in the onetrainer output. Kohya at times makes your face just slightly too narrow compared to the face you present on youtube.

[@SECourses]

you should compare it with the training images. currently i gained some more weight :D

[@streamtabulous]

can you make a stand alone of the wd14 you have in your koyha ss, as someone with dyslexic and reading writing issues that tool has been fantastic. using koyha just for wd14 to use in onetrainer as my rtx3060 would not do sdxl in koyha, and stand alone wd14 tool would be very helpful. or recommend tool if there already is one.

[@SECourses]

Yes I can make. Actually we might already have I need to check. Can you message me from discord so I can let you know once I have time

[@HanSolocambo]

3:10 Given that the name we give to a concept "training" or "regularization" could be any name, how can OneTrainer know, when we activate 2 concepts, which one are the training images and which one are the regularization images ? I guess we still have to respect kohya's training folders (1_trigger girl / 1_girl) and those folders names are the only way for OneTraininer to know what's what ?

[@SECourses]

it doesnt. it doesnt have dreambooth so it treat every image same.

[@HanSolocambo]

@@SECourses Wow. That's uttterly terrible. Holly fuck... 2 full days trying to train something because this tutorial showed such a promising app.... but not so much as we CAN setup regularization images in OneTrainer BUT those images won't be used... oh wait, worse: they WILL be used but as training images basically. Fucking up the whole likeness training. Thanks for your answer, but that's the worst news of the day: This tutorial made me believe OneTrainer would replace Kohya UI, but OneTrainer = straight to the trash now. What's the use of those concepts we define in Onetrainer then ? :/ Well... I'm gonna have two days of electric power to re-train everything with Kohya... Really pissed off against OneTrainer now... Cheers.

[@SECourses]

they are making dreambooth effect . they are helping i compared

[@cesarurrutia1016]

fantastic video, thanks for sharing, I learn a lot with your videos, do these settings work for clothing, I am dedicated to ecom photography and I need the clothing to be 100% 100% the same as the original, is this possible with this type of training, or is there any software that can do this? Thank you.

[@SECourses]

yes it would work but the training images would change. perhaps you may needed captioning as well. did you try IDM VTON it is also very powerful : kzbin.info/www/bejne/o2XTlHybdruEepI

[@Illusiongraphy]

lol what happened at the end of tjhe video rushing to end it without completing the tutorial? nature calling? xD

[@SECourses]

@@Illusiongraphy this was a short one :) you can see 133 minutes full tutorial here kzbin.info/www/bejne/ZqWYnWl5hZ6YeMksi=f69O_0aSiZStiTvG

[@sneedtube]

Is this dreambooth or lora?

[@stevebruno7572]

Lookd like dreambooth bc you can see finetuning in upper right corner

[@SECourses]

this is fine tuning. dreambooth also fine tuning but a slightly specialized way. with adding second concept, we make the effect of dreambooth almost

[@SECourses]

this is fine tuning. dreambooth also fine tuning but a slightly specialized way. with adding second concept, we make the effect of dreambooth almost@@stevebruno7572

[@AIPixelFusion]

How does the quality compare to dreambooth?

[@SECourses]

@@AIPixelFusion almost same. i did a comparison here : medium.com/@furkangozukara/experimenting-with-onetrainer-onetrainer-vs-kohya-realism-vs-stylization-reg-images-vs-0438950e9515

[@suveniro4ka]

Чепуха, файлы то где?

[@SECourses]

on patreon or you can watch this public tutorial (config is shown): kzbin.info/www/bejne/ZqWYnWl5hZ6YeMk OneTrainer Stable Diffusion XL (SDXL) Fine Tuning Best Presets : www.patreon.com/posts/96028218 The Very Best OneTrainer Workflow & Config For SD 1.5 Based Models DreamBooth / Full Fine Tuning : www.patreon.com/posts/97381002

[@moc_santos]

when is not an indian, it is a russian making tutorials

[@SECourses]

I am Turkish wrong guess this is my profile : www.linkedin.com/in/furkangozukara/

[@moc_santos]

@@SECourses Please improve your english!

[@SECourses]

@@moc_santos I am trying. Where are you from?

[@HelenSchwieger]

You've never heard a real Russian accent, but you've already made a silly attempt at guessing and got it wrong. Ironic.