Is it possible to extract a lora from finetuned model with Onetrainer?
Is it possible to extract a lora from finetuned model with Onetrainer?
J@Dr. Furkan Gözükara Thanks to your great guides I now have a really great dreambooth model that reproduces the face correctly in most situations 
. However I still have big problems with hands 
. I need to generate a lot of images of the person and I dont have time to manually fix certain images. That's why I have to sort out about 80% of ALL images because the hands are misformed. Adetailer does not do a good job at fixing hands too. Do you have any recommendation for a workflow that either avoids hands in the image or at least fixes them sometimes? I cannot use "hands in pocket" because the images are NSFW mostly... Thanks for your ideas. You do an amazing job, really!
. However I still have big problems with hands . I need to generate a lot of images of the person and I dont have time to manually fix certain images. That's why I have to sort out about 80% of ALL images because the hands are misformed. Adetailer does not do a good job at fixing hands too. Do you have any recommendation for a workflow that either avoids hands in the image or at least fixes them sometimes? I cannot use "hands in pocket" because the images are NSFW mostly... Thanks for your ideas. You do an amazing job, really!
Atry to use embeddings: they don't do miracles, but sometime prevent big mistakes (that adetailer can fix more easily); try to search specific embeddings on civitai and play around...
Fit would take few minutes
Funless it uses shared vram
Fi can test it
Ftell me your final resolution that you see
Fbut i would compare
Fboth original and upscaled
Fyes but i prefer kohya better
Fhands issue coming from the base model
Fand thanks
R@Dr. Furkan Gözükara in kohya, can "Epoch" be set to "0" or empty if I set Max Steps to 10000? I am having issue when i set sample every 100 steps but leave Sample every "0" epochs. Should I leave sample every x epochs? I'm struggling finding what is the accepted parameter
Rstandard_1.json4.4KB
V@Dr. Furkan Gözükara can webui still be used?
KCan you explain what this means? do you mean do an original supir and then an upscale supir?
Aone trainer won't train what am I missing
Dtried a fresh install?
Fnever tried
Fyes
Fi mean do 1 training with original
F1 with upscaled
Fand compare
Fi dont know you can ask developer on their discord
Fhe is super active
Fpurchased Poco X6 pro and transferring all installed apps taking huge time
Ppoco x6 is good? i have a poco gt x3
Pi think they only change screen type in x6 haha
Awheres the tier2 for onetrainer @Dr. Furkan Gözükara ?
Afor 24gb 3090 cards
VYou also like performance Doc...i have redmi k50 with dimensity 8100...works good..
Vdo you also root your devices doc?
Vmine is rooted
Fi have poco x3 pro
Fupgrading to poco x6 pro - going to give mine to my wife
Fit is super powerful phone
Fno more tier 2 after recent updates
Fall are tier 1
Fno i dont root
Fin past i did in few phones but not needing it
Pi want do this too git mine to my wife haha
Pthey always have the leftovers haha
F
Fwell my wife prefers getting clothes so she gets them
Vwomen usually love two things - clothes and jewellery
MHi all,
has anyone yet tried to be more accurate while training specific body parts, e.g. abs and pecs of a trained man, etc.
Or legs, waist and bottom? Obviously you cannot always have the face on the pictures if you dont always wanna take full body shots, especially for lower bodyparts.
I'm wondering what is the possible way to use closeup pictures to be more detailled.
Right now I'm not labelling / captioning my pictures.
But I dont know how the trainer is supposed to know that this bottom belongs to "ohwx man" or similarly.
Has anyone experience if this specific task can be improved by only labelling those specific images exactly with using the instance prompt? "bottom of ohwx man from behind". Or stuff like this?
has anyone yet tried to be more accurate while training specific body parts, e.g. abs and pecs of a trained man, etc.
Or legs, waist and bottom? Obviously you cannot always have the face on the pictures if you dont always wanna take full body shots, especially for lower bodyparts.
I'm wondering what is the possible way to use closeup pictures to be more detailled.
Right now I'm not labelling / captioning my pictures.
But I dont know how the trainer is supposed to know that this bottom belongs to "ohwx man" or similarly.
Has anyone experience if this specific task can be improved by only labelling those specific images exactly with using the instance prompt? "bottom of ohwx man from behind". Or stuff like this?
A
S@Dr. Furkan Gözükara I did not notice you were training up to 150 epoch for just 15 images. That makes me wonder what would be the ideal epoch to train up to for 4000 images?
SAre your sample images on early epochs looking fried?
Y@Dr. Furkan Gözükara , is there any LoRA model we can play with that shows the full implementation in pytorch ? I would like to see the layers , connections and everything to change the architecture and see the impact for instance