Software Engineering Courses (SECourses)•8mo ago

make sure to save once every n steps

FFurkan Gözükara SECourses make sure to save once every n steps

Escanor•5/8/25, 10:51 PM

Ok, and is the tier1 48gb config optimal for A100 graphic card with 80gb or I should change something to make it better and use more available memory? Batch size around 4-7? Do you recommend onetrainer finetune over kohya_ss dreambooth?

Thank you for the quick answer

EEscanor Ok, and is the tier1 48gb config optimal for A100 graphic card with 80gb or I sh...

Furkan Gözükara SECoursesOP•5/8/25, 10:53 PM

i dont recommend change anything it is best quality

Escanor•5/8/25, 11:01 PM

I'll try to give results here if I find the best sweet spot for my number of training images!

I do think it could be useful to have somer kind of pinned message with recommended setups of epoch #, repeat # and such specific to the # of images, or at least a ballpark close to it

Like 8 images: 400 repeats, 1 epoch

EEscanor Ok 🙂 I'll try to give results here if I find the best sweet spot for my number...

Furkan Gözükara SECoursesOP•5/8/25, 11:13 PM

sadly there isnt much but i usually recommend 200 epoch

Furkan Gözükara SECoursesOP•5/8/25, 11:13 PM

however your train image count too low

Furkan Gözükara SECoursesOP•5/8/25, 11:14 PM

more is better

Escanor•5/8/25, 11:39 PM

Yeah I only have 8 images, I'm hoping I can finetune a model and generate decent enough images to be able to train with more images... Not sure how that will go lol

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•5/9/25, 9:08 AM

hi, for quick trainings I often use civitai. I remember you talked about rank of lora in kohya, here it is both rank and alpha, and not sure about these. The default seems to be the same for both flux and sdxl trainings: Default is also not adafactor, but the Adam...any recommendations here? realistic person training

EEscanor Yeah I only have 8 images, I'm hoping I can finetune a model and generate decent...

Furkan Gözükara SECoursesOP•5/9/25, 1:52 PM

you can also do this

Furkan Gözükara SECoursesOP•5/9/25, 1:52 PM

1st fine tune generate pick good ones

Furkan Gözükara SECoursesOP•5/9/25, 1:52 PM

and fine tune again

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭hi, for quick trainings I often use civitai. I remember you talked about rank of...

Furkan Gözükara SECoursesOP•5/9/25, 1:52 PM

well the combination of all parameters make a good workflow

Furkan Gözükara SECoursesOP•5/9/25, 1:53 PM

so you should research good params there with testing all combinations

FFurkan Gözükara SECourses and fine tune again

Escanor•5/9/25, 1:55 PM

Yeah, I'll first try to finetune with only 8 images, will try to find the best # of steps sweet spot until not too much overfit and then generate new images and finetune again

TrumpYouAll•5/10/25, 4:11 PM

@Furkan Gözükara SECourses I am not 'urging' u to make a video now but I would like to know if Kohya can be used to train HiDream model (with my personal pics) like Flux ? Can I do it in Massed Compute ?

Escanor•5/10/25, 5:50 PM

If I train model A for 3200 steps and model B for 5000 steps, both having the same base model, would the result of generating an image be the same for both if for model B I use a checkpoint that's near 3200 steps like model A (save every x epoch/steps)?

And if I train a model for 3200 steps and then retrain it for another 1800 steps for a total of 5000 steps, will that model be as efficient as if I trained a model to 5000 steps directly?

Thank you

EEscanor Yeah, I'll first try to finetune with only 8 images, will try to find the best #...

Furkan Gözükara SECoursesOP•5/11/25, 11:30 PM

great

TTrumpYouAll <@205854764540362752> I am not 'urging' u to make a video now but I would like t...

Furkan Gözükara SECoursesOP•5/11/25, 11:30 PM

that is also on my list but so many stuff

EEscanor If I train model A for 3200 steps and model B for 5000 steps, both having the sa...

Furkan Gözükara SECoursesOP•5/11/25, 11:31 PM

yes it should be same 3200 + 1800 = 5000 directly

Furkan Gözükara SECoursesOP•5/11/25, 11:31 PM

since our learning rates are all constant

Escanor•5/13/25, 3:38 PM

@Furkan Gözükara SECourses Is the kohya dreambooth sdxl 48gb config usable with the kohya sdxl lora extract configuration that you have in one of your patreon posts (not sure if I can post image here)?

I am asking because if I leave the default settings of Clamp Quantile at 0.99 and Minimum difference to 0.01, I think it says the text encoder does not need to be extracted or something similar and when I try to use the lora, it simply does not show anything of my trained model compared to if I use the settings in your image of 0.1 for Clamp Quantile and 0.00001 for Minimum difference where now it works fine.

It just made me wonder if that screenshot of the kohya sdxl lora extract settings was up to date with the latest 48gb config of kohya dreambooth settings!

I did see a recent post for lora extract, but it was for flux so I wasn't sure

Thank you for the great support!

EEscanor <@205854764540362752> Is the kohya dreambooth sdxl 48gb config usable with the k...

Furkan Gözükara SECoursesOP•5/13/25, 4:37 PM

make it 0.000001 to extract

Furkan Gözükara SECoursesOP•5/13/25, 4:37 PM

also flux post is updated too but flux extract not very good

JAV_Pomoe•5/14/25, 1:54 AM

Doctor, could you please tell me the best settings for LoRA training for SDXL? I saw your previous post about fine tuning... Have you already posted about LoRA training?

JJAV_Pomoe Doctor, could you please tell me the best settings for LoRA training for SDXL? I...

Furkan Gözükara SECoursesOP•5/14/25, 2:12 AM

we have dreambooth

Furkan Gözükara SECoursesOP•5/14/25, 2:12 AM

i recommend after that extract lora

FFurkan Gözükara SECourses i recommend after that extract lora

JAV_Pomoe•5/14/25, 2:32 AM

I see. I'll give it a try. Thank you as always.

Escanor•5/14/25, 11:15 AM

@JAV_Pomoe Personally, on my end and from what I see in all the guides, I get better results by training either with kohya_ss dreambooth or onetrainer and then extracting the lora from the finetuned model with koyha_ss. The resulting extracted lora seems better to me than one training directly trained as such. Plus, you also have the finetuned model you can keep too

It's crucial to use the good settings for extracting the lora and finetuning the model though. For extracting the LORA, the settings can be found in an image at the very bottom of this patreon post

However, do note that in the image, the Clamp Quantile and Minimum difference don't have red rectangels around them, but they do have different values than the default and you have to use those precisely.

Good luck!

Patreon

Full Workflow For Newbie Stable Diffusion Trainers For SD 1.5 Model...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

uisato•5/14/25, 4:25 PM

Kinestasis Stop Motion / Hyperlapse - [new WAN 2.1 LORAs]

https://www.youtube.com/watch?v=3rYPtYxq_3c&ab_channel=uisato

YouTubeuisato

Kinestasis Stop Motion / Hyperlapse - [WAN 2.1 LORAs Project Files]

Technique consisting in a fine-tuned AI model [WAN 2.1 / txt2vid]. Hundreds of hours of training and testing. Still far from good, but hopefully getting somewhere. I'm a huge fan of this technique.

Music by @klsr-av

You can access these new WAN 2.1 LORAs, through my Patreon profile.

#stopmotion #machinelearning #design #aiart

petecemetery•5/14/25, 10:42 PM

looking for some advice on this regularization workflow for training this flux concept. my concept im training is faces pressed against glass. So i have 75 images of faces like the above. Then I have captioned each one without any mention of pressing against glass, and have generated regularization images with those captions so that each training image has a corresponding regularization image to help isolate the concept. With this workflow, do i want to keep repeats on both training and reg images at 1?

petecemetery•5/14/25, 10:45 PM

also as long as my txt files are in the same folder as my training images , they should be used to train right?

petecemetery•5/14/25, 10:50 PM

im using the lora config for this 1voteup

Ppetecemetery looking for some advice on this regularization workflow for training this flux c...

Furkan Gözükara SECoursesOP•5/14/25, 10:54 PM

you dont need reg images

Furkan Gözükara SECoursesOP•5/14/25, 10:54 PM

make sure every person in dataset is different

Furkan Gözükara SECoursesOP•5/14/25, 10:54 PM

it should work perfect

petecemetery•5/14/25, 10:54 PM

ok great, thanks

Ppetecemetery ok great, thanks

Furkan Gözükara SECoursesOP•5/14/25, 11:04 PM

you are welcome

EEscanor <@1320367778317865031> Personally, on my end and from what I see in all the guid...

JAV_Pomoe•5/15/25, 7:32 AM

Thank you so much for the great information!

Stav•5/15/25, 11:13 PM

Hi, how many epochs are recommended for a 15 pictures training with dreamboth?

SStav Hi, how many epochs are recommended for a 15 pictures training with dreamboth?

Furkan Gözükara SECoursesOP•5/15/25, 11:33 PM

at least 200

Furkan Gözükara SECoursesOP•5/15/25, 11:33 PM

but you can go up to 300 and compare

FFurkan Gözükara SECourses at least 200

Stav•5/16/25, 11:42 AM

Thank you doctor

Escanor•5/16/25, 1:48 PM

Honestly I'm always wondering how to know the number of epochs and such. Should we just aim for a certain # of total steps and play with the # of epochs and repeats?

For example, with 15 images 200-300 epochs but how many images repeats? Or is it better 1 epoch and tons of repeats?