Home
Artists
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register
Partychan Telegram ThePornDude
Home Artists Posts Import Register

OneTrainer Stable Diffusion XL (SDXL) Fine Tuning Best Presets (Patreon)

Published:
2024-01-07 11:23:13
Edited:
2024-03-30 15:26:10
Imported:

Downloads

Content

Patreon exclusive posts index

Join discord and tell me your discord username to get a special rank : SECourses Discord

30 March 2024 Update:

All configs updated and weight decay parameter set in Adafactor optimizer since it improves quality

26 March 2024 Update:

22 March 2024 Update:

  • All configs are updated and Stochastic Rounding disabled

  • Stochastic Rounding makes the effect of like FP32 - float training

  • However, float like training causes overtraining with our currently found best hyper parameters

  • Therefore, for now, they are disabled until a better hyperparameters are researched and found

  • How to use these configs quick tutorial : https://www.youtube.com/watch?v=yPOadldf6bI

9 February 2024 Update:

  • Settings are updated for the latest OneTrainer update

  • You don't need to put optimizer prefs anymore

  • You can open configs json file and look inside to understand logic of how it works

  • You need to change workspace_dir, cache_dir, output_model_destination, edit your training concepts

  • I did set default model as the SDXL 1.0 base model hosted on Hugging Face

  • You can change model to any model you want from your computer or Hugging Face repo

  • Hopefully I will make a full tutorial that includes how to train on Windows and RunPod

I have been experimenting with Nerogar OneTrainer for days right now.

OneTrainer repo : https://github.com/Nerogar/OneTrainer

We have 3 configs.

The configs will save checkpoints every 25 epochs during training and save a final checkpoint. Don't forget to change that behaviour or fix your final checkpoint path.

Clone the repo into any folder : https://github.com/Nerogar/OneTrainer

Double click install. bat

Then double click start-ui .bat to start it after putting presets into accurate folders

Kohya config : https://www.patreon.com/posts/very-best-for-of-89213064

Load our preset according to your GPU VRAM

In the general tab set your workspace directory and cache directory

In the model tab set your SDXL model path and output destination

In concepts tab add your training images and set their training captions. I trained with ohwx man

I don't see sampling is useful. It uses poorly performing sampler by default

In backup you can set your parameters as you wish. Like save after every 25 epoch. Saved checkpoints will be inside workspace/save folder

As I said I will make hopefully a big tutorial for OneTrainer

What Each Config Means

tier1_10.4GB_slow.json : This file uses Fused Back Pass and disables Autocast Cache for minimum VRAM : 10.4 GB which is amazing. Same as best Tier 1 we had previously.

Tier 1 slow is 3.72 second / it on RTX 3060 - 12 GB GPU

Tier 1 slow is 1.58 second / it on RTX 3090 TI - 24 GB GPU

tier1_15.4GB_fast.json : This file don't use Fused Back Pass and enable Autocast Cache

Tier 1 fast is 1.45 second / it on RTX 3090 TI 24 GB GPU

Gradient Checkpointing do not affect quality but only make training slower yet uses lesser VRAM

I also added a full resolution comparison.jpg between default SDXL preset of OneTrainer vs my 14_5GB preset vs Kohya best config ( https://www.patreon.com/posts/89213064 ) we already use

Used training images dataset - 15 images

This dataset is at best medium quality. You should try to add different backgrounds, different clothings, different angles, different expressions and different distances.

For regularization images, I have used 5200 images dataset that we have. Set repeating 0.003 so 15 reg images used in each epoch. : https://www.patreon.com/posts/massive-4k-woman-87700469


Files

Comments

Hassan Alhassan

my model is getting over trained quickly, i used the 18gb and 14.5 gb both are getting over trained

chriseppe

the results seem promising. Would love to see if it looks even more realistic without a surreal background. Did you simply generate the images without using Adetailer on top?

Anonymous

Regardless of which Tier presets I use I get the CUDA out of memory error. I have a 16GB card and have ran other trainings with different settings just fine before. What setting in the presets may be doing this? I'm mainly trying the 13GB one.

Furkan Gözükara
>>129967459

Can you verify your VRAM usage before trying preset? Also are you using latest version of the OneTrainer? If you message me from discord i can connect via anydesk and check it out

Anonymous

How do I verify my VRAM usage? As far as I know it's the latest but I'll check. I'm on the discord as OhTheHueManatee.

Javi dltr

how do we use captions here for reg and for data?

Javi dltr

in sdxl config 1 ema is off but decay is 0.999?

John Dopamine

There's a new branch (just about merged) w a new "fused base" feature that allows for a significant decrease in the amount of VRam needed to train. If you haven't seen this yet, keep an eye out. With 24gb you can now train w float32, or large batch sizes like 14. It also should allow people who have cards under 16gb train now also. Mentioning this after seeing you updated after reviewing stochastic rounding. This discovery kinda came from that (which came due to Cascade not training at all). Perfect timing w SD3 right around the corner (please let them not backtrack).

Karl B

I have a question, on the Onetrainer wiki it says if "latent caching" is on as it is in your file, then you need to use more than 1 on the "image variants" setting under concepts. I have about 50 photos ready to train, Onetrainer doesn't really explain what do set that number to. Any ideas?

George Gostyshev

That's really great work done! Appreciated) OneTrainer big tutorials could be nice. Also, can I ask\suggest to sometimes add not so big videos? Like videos about some small but useful technics. Anyway thanks again)

Marian Ban

Hi, what are the correct settings for network rank and dimension? These are not included in the json files. I'm asking about the Kohya settings. Thanks

Furkan Gözükara
>>133606003

hi that is lora. i dont suggest lora . these are fine tuning : https://www.linkedin.com/posts/furkangozukara_why-i-dont-research-lora-training-because-activity-7164700874097856512--Ecj/?utm_source=share&utm_medium=member_desktop

mike oxmaul

Why not use EMA? Previous post said using EMA made this better than Kohya.

Tom Bloomingdale

Hello - this is working in that it is training and finishing. The results are fairly consistent and look like the same person (kind of) but nothing like the me (the subject). I have tried a few different models as a base, and gone through the settings as you describe in your video/writings several times. Any idea why I am not getting results as expected? This is SDXL, the tier1_10.4GB_slow settings, 200 epochs.

Furkan Gözükara
>>134187827

are you using adetailer to fix the face after generation? that is super important. please let me know after trying