Furkan Gözükara patreon

Experimenting With OneTrainer - OneTrainer vs Kohya - Realism vs Stylization - Reg images vs No-Reg images

🕑 Added 2024-01-08 14:03:48 +0000 UTC

Comments

Furkan Gözükara

Sadly not because it requires a Desktop GUI. I plan to research for such template on RunPod.

Faraad Haashim

Big improvements to realism. Is there a runpod installer script for OneTrainer?

Furkan Gözükara

yes they recently merged into code and I plan to do hopefully soon

Hassan Alhassan

you can train on Pixart using onetrain, i just started the training not sure about the results

Furkan Gözükara

not yet sadly.

mike oxmaul

Did you get a chance to test with the face masking feature?

Furkan Gözükara

i am starting machines now to start testing

Idraen

no problem man, thanks for the hard work!

Furkan Gözükara

still working on that. so sorry that i still couldn't release

Idraen

Any chance to get a SD 1.5 config?

Furkan Gözükara

it is true for sd 1.5. for SDXL i dont know maybe due to incorrect implementation i got horrific results nothing good with EMA. did a lot of testing

John Dopamine

What about training EMA? I recall you praising Kohya's 1.5 version w EMA training for being one that gave the best results. The configs here (which have worked very well for me) have the EMA training. I think that is default. Is that function disabled in these configs due to not (yet) being tested, or have you determined the best option is to leave it off (at least how it works at the moment in Open trainer)?

Furkan Gözükara

i am yet to test masking

Umut T.

I would like to see comparison between masked vs usual training as well !!

Furkan Gözükara

ye kohya is great too

Umut T.

I love my Kohya, no ty. If I wanna generalization, I just train on sdxl dpo... I believe style loras are better without generalization.

John Dopamine

Thanks for the instructions on reg images!! I need to do comparisons w/ masking on vs. off. I'm not sure if it just masks the person and doesn't look at the rest at all or if it just weights things differently. If it only looks inside the mask then yea, maybe not as useful as I was thinking. I got good enough results w/ it on though so I'll try it on w/ reg images sometime.

Furkan Gözükara

1 : added our reg images as new concept but made the repeat 0.003. so 0.003 * 5200 = 15 images used each epoch 2 : i only have a txt file for training dataset that has ohwx man and for reg images just man. so a single txt file that contains the rare token and class token 3 : i didnt try masking yet. i don't think it will improve quality. do you have good results with it? i believe better dataset better than masking

John Dopamine

Do you explain anywhere how you incorporated reg images? Did you just make a concept for "Regularisation images", put 3k images in there w/o captions (or maybe a single caption txt file w/ the class name? ie "Person"?) Also have you used captions or a single caption txt for your trials so far? I got really nice results w/ your configs already, but I only put a simple txt file w/ the subject (trigger) name in it - I never captioned w/ kohya for single likenesses. Lastly, per the developer in that thread you created on their github I tried mask traininng last night and had good results. I used the "remgb" option and it quickly made masks for just the person - pretty cool. Do you know in that case if it just looks at the person, or does it look at both but can differentiate better? I don't believe Kohya has a masking feature so I wanted to try it. I'm so glad you picked this to look into. I see others asking for different things, but for me this is the best topic you've taken on after Kohya's training. Plus the OT developer is so active that I'm sure he'll be very quick to add support for future models (OT supports PicsArt training, Wurscher training, etc - dev is all in w/ the different models).