i recommend after that extract lora
i 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
EUrgh ok. Because on mac I need to use a weird flux model or else tons of errors. Model is flux1-dev-Q8_0.gguf 
Fi would use mac if only i was developing an app for iphone :d
EMeh, for developing from backend to frontend it's great, ai stuff it's shit lol
ECompatibility is atrocious
EEspecially with arm/sillicon chips
Fyep that is why
EWhat is the best workflow for preparing images?
Let's say I have portrait images that are mediocre quality and not 1024x1024. Which of the following is the best workflow:
Workflow A: joycaption,supir 5x upscale with caption,outpaint image sides,downscale to 1024x1024
Workflow B: joycaption,supir 5x upscale with caption,downscale to 1024x1024 with transparent sides
Workflow C: joycaption,supir 5x upscale with caption,downscale to 1024x1024 with black or white sides
Let's say I have portrait images that are mediocre quality and not 1024x1024. Which of the following is the best workflow:
Workflow A: joycaption,supir 5x upscale with caption,outpaint image sides,downscale to 1024x1024
Workflow B: joycaption,supir 5x upscale with caption,downscale to 1024x1024 with transparent sides
Workflow C: joycaption,supir 5x upscale with caption,downscale to 1024x1024 with black or white sides
EAlso, I'm unsure which caption type + length I should use with joycaption
EAnd for having square images, for my current dataset, I had used Fooocus to outpaint the sides since the original images were portrait, but would it be acceptable or good to just resize to 1024x1024 and have the sides be black,white or transparent..? Or will that confuse the training?
ESorry for many questions, thank you very much for the help!
SNo I mean I'm currently using the 48go config with an L40S and getting around4.8 sec. I have a deadline and would need the training to be twice faster, so I wondered if that was achievable with an H100 SXM. Sorry if I explained myself wrong
Ffor flux i dont use captions i dont recommend 
Fah yes go with h100 it will be faster directly
EWell it wasn't for flux specifically, it was to have captions from image to upscale them with SUPIR!
What I'm unsure of is the best workflow to prepare a dataset of mediocre quality to increase its quality...
And also if I can just resize to 1024x1024 a portrait images and just leave the sides transparent
What I'm unsure of is the best workflow to prepare a dataset of mediocre quality to increase its quality...
And also if I can just resize to 1024x1024 a portrait images and just leave the sides transparent
Ffor supir caption matters