Now You Can Full Fine Tune / DreamBooth Stable Diffusion XL (SDXL) with only 10.3 GB VRAM via OneTrainer
Added 2024-03-26 04:44:32 +0000 UTCNow You Can Full Fine Tune / DreamBooth Stable Diffusion XL (SDXL) with only 10.3 GB VRAM via OneTrainer — Both U-NET and Text Encoder 1 is trained — Compared 14 GB config vs slower 10.3 GB Config
Full config and instructions are shared here : https://www.patreon.com/posts/96028218
Used SG161222/RealVisXL_V4.0 as a base model and OneTrainer to train on Windows 10 : https://github.com/Nerogar/OneTrainer
The posted example x/y/z checkpoint comparison images are not cherry picked. So I can get perfect images with multiple tries.
Trained 150 epochs, 15 images and used my ground truth 5200 regularization images : https://www.patreon.com/posts/massive-4k-woman-87700469
In each epoch only 15 of regularization images used to make DreamBooth training affect
As a caption only “ohwx man” is used, for regularization images just “man”
You can download configs and full instructions here : https://www.patreon.com/posts/96028218
Hopefully full public tutorial coming within 2 weeks. I will show all configuration as well
The tutorial will be on our channel : https://www.youtube.com/SECourses
Training speeds are as below thus durations:
RTX 3060 — slow preset : 3.72 second / it thus 15 train images 150 epoch 2 (reg images concept) : 4500 steps = 4500 3.72 / 3600 = 4.6 hours
RTX 3090 TI — slow preset : 1.58 second / it thus : 4500 * 1.58 / 3600 = 2 hours
RTX 3090 TI — fast preset : 1.45 second / it thus : 4500 * 1.45 / 3600 = 1.8 hours
A quick tutorial for how to use concepts in OneTrainer : https://youtu.be/yPOadldf6bI
Comments
for SD 1.5 we have different config here : https://www.patreon.com/posts/very-best-config-97381002 please also watch this tutorial : https://youtu.be/0t5l6CP9eBg
Furkan Gözükara
2024-05-02 01:44:29 +0000 UTCdoes this work with 1,5?
Javi dltr
2024-05-02 00:46:58 +0000 UTCyou are welcome
Furkan Gözükara
2024-04-01 17:47:55 +0000 UTCAwesome thank you. moving much faster now, looks like about three and a half hours. Thanks for the help
Tom Bloomingdale
2024-04-01 17:44:32 +0000 UTCthat means using shared vram. use lower speed. something else could be using your vram already
Furkan Gözükara
2024-04-01 14:39:06 +0000 UTCHello! I have this set up for SDXL using the "fast" option for fine tuning. I have a 4060ti with 16gb vram. With 150 epoch, Im looking at like 35 hours training time. Am I doing something wrong?
Tom Bloomingdale
2024-04-01 13:06:29 +0000 UTCmore images requires same vram it wont make difference. but more images would require lesser number of total epochs
Furkan Gözükara
2024-04-01 03:52:56 +0000 UTCHi doc! Does more images require more vram? I don't get the consistency that I have with sd1.5 using the same photos, especially in the shape and physiognomy of the face.
Leonardo Chocron
2024-03-31 18:40:12 +0000 UTChttps://github.com/Nerogar/OneTrainer/commit/07f8198e8762eed2b1f5755bc0445f2ae1cd988b
Doc Snyder
2024-03-28 12:12:37 +0000 UTCok i asked the developer and he said it is already fixed
Furkan Gözükara
2024-03-28 12:09:59 +0000 UTCi get same error. the guy broken previous configs i have to remake
Furkan Gözükara
2024-03-28 12:07:53 +0000 UTCWith the latest version of onetrainer I get "AttributeError: 'str' object has no attribute 'is_wuerstchen'" while changing to the added preset and it doesn't load.
Doc Snyder
2024-03-28 04:50:10 +0000 UTCSDXL is better with 12 GB
Furkan Gözükara
2024-03-28 02:36:48 +0000 UTCno on computer but i am gonna introduce something better hopefully
Furkan Gözükara
2024-03-28 02:36:37 +0000 UTCHi Furkan, would you rather do a full training on SD1.5 with your preset for 12GB cards or a full training on SDXL with your new config? Which training ist better for realistic faces whilst still enabling flexibility in the generated pictures?
C. Jonas
2024-03-27 09:48:37 +0000 UTCare you tried on runpod?
Joel Maynard
2024-03-27 04:37:42 +0000 UTCye kaggle is picky. i prefer runpod over paid google colab. cheaper and better. also there is a new cloud computing i am going to introduce. 31 cents per hour for A6000 gpu - 48 gb
Furkan Gözükara
2024-03-27 02:17:45 +0000 UTCI'm not impressed with Kaggle Notebook because I got blocked. And they said I violated community rules. Even though I use the script to test.
Gen Zero
2024-03-27 02:11:00 +0000 UTCah i see what you mean. it can read safetensors or ckpt as well
Furkan Gözükara
2024-03-26 23:58:10 +0000 UTCI understand. Does onetrainer only take diffusers for this low vram method?
Hey Ooo
2024-03-26 23:54:17 +0000 UTCThis is standalone app OneTrainer.
Furkan Gözükara
2024-03-26 23:53:04 +0000 UTCDoes this only work with diffusers
Hey Ooo
2024-03-26 19:22:40 +0000 UTCi think both are really good. i would compare both to see which one is better for you
Furkan Gözükara
2024-03-26 13:38:09 +0000 UTCSince I don't have premium colab I can't test. but our kaggle notebook should work fairly well on pro colab have you tested?
Furkan Gözükara
2024-03-26 13:37:49 +0000 UTCyou are welcome
Furkan Gözükara
2024-03-26 13:37:12 +0000 UTCthank you so much
Furkan Gözükara
2024-03-26 13:37:07 +0000 UTCThanks, I'll give it a try.
Alex
2024-03-26 10:07:11 +0000 UTCwonderful work, Furkan!
Art
2024-03-26 09:51:05 +0000 UTCI want Stable Diffusion - Automatic11119 The latest one is 2024 at Run on Colab. Please send me some tutorials. Thank you very much.
Gen Zero
2024-03-26 08:04:41 +0000 UTCThx for this. How does it compare to latest kohya db TE training? I get excellent realism on kohya. From your earlier post these are very good but maybe facial detail not as good w/onetrainer. Thoughts?
Ec Jep
2024-03-26 04:59:04 +0000 UTC
