and fine tune again
and fine tune again
Fwell the combination of all parameters make a good workflow
Fso you should research good params there with testing all combinations
EYeah, 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
T@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 ?
EIf 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
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
Fthat is also on my list but so many stuff 
Fyes it should be same 3200 + 1800 = 5000 directly
Fsince our learning rates are all constant
E@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!
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!
Fmake it 0.000001 to extract
Falso flux post is updated too but flux extract not very good
JDoctor, 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?
Fwe have dreambooth
Fi recommend after that extract lora
JI see. I'll give it a try. Thank you as always.
E@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!
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
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
UKinestasis Stop Motion / Hyperlapse - [new WAN 2.1 LORAs]
https://www.youtube.com/watch?v=3rYPtYxq_3c&ab_channel=uisato
https://www.youtube.com/watch?v=3rYPtYxq_3c&ab_channel=uisato
YouTubeuisato
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
Music by @klsr-av
You can access these new WAN 2.1 LORAs, through my Patreon profile.
#stopmotion #machinelearning #design #aiart
Plooking 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?
Palso as long as my txt files are in the same folder as my training images , they should be used to train right?
Pim using the lora config for this 
Fyou dont need reg images
Fmake sure every person in dataset is different
Fit should work perfect
Pok great, thanks
Fyou are welcome
JThank you so much for the great information!
SHi, how many epochs are recommended for a 15 pictures training with dreamboth?
Fat least 200
Fbut you can go up to 300 and compare
SThank you doctor
EHonestly 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?
For example, with 15 images 200-300 epochs but how many images repeats? Or is it better 1 epoch and tons of repeats?
EI also wonder if with like 15 portrait images we should be using captions or not 
EAlso wondering because the sdxl kohya 48gb config uses 1 epoch
SHey, for flux full finetuning is there a solution faster than L40S for the 48gb config file ? Thanks
Frepeat is always 1 for flux
Fwe had to use repeat for sdxl to use more reg images due to how kohya bad coded reg images logic
Fbecause with sdxl we want to use more reg images
Fand only way is more repeat with kohya
Fwhat do you mean?_
EAhh I see. Then for sdxl is it better to use onetrainer instead in your opinion? Considering I have only 8 portrait images and the 5000 reg images of your patreon. Also unsure if I should use captions or not
Thank you
Thank you
SI mean would an H100 run the training significantly faster than the L40S ? I'm currently at 4.8sec/iteration
Fyes onetrainer also has EMA
Fwhich works good for SDXL
Fyes it is faster
Fare you batch size 7?
F4.8 sec is low for H100
EOk for onetrainer, will try.
Is FLUX really that much better than SDXL though even for just 8 training images? Wondering because it's so slow to run on my m2 macbook pro..
Is FLUX really that much better than SDXL though even for just 8 training images? Wondering because it's so slow to run on my m2 macbook pro..
Fvery very better
Fso much difference
