@Dr. Furkan Gözükara have you tested results of data photos having a beard vs no beard?
@Dr. Furkan Gözükara have you tested results of data photos having a beard vs no beard?
Swondering if i need to shave lol
RHey @Dr. Furkan Gözükara got amazing results for 12 image character LoRA using AdamW8bit 8000 steps--testing different batch sizes quality vs speed of training. Did you test LoRA Adafactor settings or only Full Finetune? I tried LoRA Adafactor with your settings with LR 0.0004 and it learned very little, color discoloration etc.. Can you point in right direction for LoRA--Adafactor?
RI see you are checking Adafactor SDXL TE problem on Issue #855. Does this mean Adafactor is broken at the moment, in light of TE? I thought it was beneficial to exclude SDXL Text Encoder from learning? Is Dadaptation a better option?
Rcan someone explain why google colab gives me this error:
ImportError Traceback (most recent call last)
<ipython-input-7-fdf29cbd2fa1> in <cell line: 22>()
20 import io
21 import math
---> 22 import timm
23 from IPython import display
24 import lpips
11 frames
/usr/local/lib/python3.10/dist-packages/PIL/ImageFont.py in <module>
32
33 from . import Image
---> 34 from ._util import isDirectory, isPath
35
36 LAYOUT_BASIC = 0
ImportError: cannot import name 'isDirectory' from 'PIL._util' (/usr/local/lib/python3.10/dist-packages/PIL/_util.py)
---------------------------------------------------------------------------
NOTE: If your import is failing due to a missing package, you can
manually install dependencies using either !pip or !apt.
To view examples of installing some common dependencies, click the
"Open Examples" button below.
ImportError Traceback (most recent call last)
<ipython-input-7-fdf29cbd2fa1> in <cell line: 22>()
20 import io
21 import math
---> 22 import timm
23 from IPython import display
24 import lpips
11 frames
/usr/local/lib/python3.10/dist-packages/PIL/ImageFont.py in <module>
32
33 from . import Image
---> 34 from ._util import isDirectory, isPath
35
36 LAYOUT_BASIC = 0
ImportError: cannot import name 'isDirectory' from 'PIL._util' (/usr/local/lib/python3.10/dist-packages/PIL/_util.py)
---------------------------------------------------------------------------
NOTE: If your import is failing due to a missing package, you can
manually install dependencies using either !pip or !apt.
To view examples of installing some common dependencies, click the
"Open Examples" button below.
JHello guy's, How can I run 2 instances of SD ( a1111 ) at the same time ?
EHey @mikemenders , try RV5 (not 5.1) with the 8400 VAE. I personally had better luck. And yes, SD1.5 vs. RV models are worlds apart.
ERegardless though your pics look great!
Sso the --share command thing is broken i believe
SCould not create share link. Please check your internet connection or our status page: https://status.gradio.app
Startup time: 13.1s (import torch: 1.2s, import gradio: 0.5s, import ldm: 0.3s, other imports: 0.8s, load scripts: 4.6s, create ui: 0.3s, gradio launch: 5.4s).
Startup time: 13.1s (import torch: 1.2s, import gradio: 0.5s, import ldm: 0.3s, other imports: 0.8s, load scripts: 4.6s, create ui: 0.3s, gradio launch: 5.4s).
EHey @mikemenders , this is interesting. Do you know the difference between this vs. adetailer, or have you tested them against each other?
TI also want to know this. Do I need different SD folders?
TIt will only work if SD is on local. You can't use that command on RunPod or Kaggle
SI only use local
SDo I need an extension or something as well?
PQuestion for any of you: I have trained LORAs for my wife, myself, and my two kids. Whats the best way to compose a family portrait or group photo? Any good tutorials out there? I haven't found much.
Si thought about doing something lie that but my approach was going to train multiple concepts within the same model with dreambooth
E@xPIXELPAPIx you'll need to use the Loras in a single prompt. So for example it'll be something like "husband:lora" sitting next to "wife:lora" with "son:lora" chasing "daughter:lora" in a living room. Something along those lines. I'm not too experienced with complicated Loras myself but it should be something along those lines and there are a few youtube videos out there of people using multiple loras.
Fhi
Fi dont think you need
Fi can get shaved photo :d
Fi will explore lora adafactor later. currently finding best hyper params for dreambooth.
Fi got super interesting finding with TE still researching. but we really need different LR for TE : https://twitter.com/GozukaraFurkan/status/1710416135747748150
Furkan Gözükara (@GozukaraFurkan)
Twitter
•
SDXL text encoder turns out to be very strong and becomes so quickly overtrained
10/6/23, 10:05 PM
Fwhich colab?
Ryea, i saw the thread. Sounds like very difficult to test for TE? Have you got some idea about it? I thought there being two TE's made it impossible practically
Fbut it really doesnt worth unless you have 2 gpu
Fi dont know still searching
Fye it wasnt working yesterday
Fand today for a while
RI am testing AdamW8bit quality with batch sizes 1, 2, 4, 6. I am seeing what is the best sweetspot for quality vs speed of training
Fyou can try latent couple extension
Fthis is necessary for companies
Fi will also research this for a company
Fit adds beard hard 
i tested not easy to get without a beard
i tested not easy to get without a beardRi got pretty impressive results with batch 2, which drove SDXL adamW8bit training by by 30% and bs 4 by 60%. For me it's somewhere in there. Need to check flexibility, but i am sure there are other experimental variables to test. Here's dataset of 12 images, vs xy plot with batch 2 up to 9 epochs (5000 steps), 5 hours vs 9 hours batch size 1
Rinteresting the level of variance in the dataset gets picked up in dim 128/alpha64. I guess the ultimate test will come when i run this LoRA on a Art Style model
R@Dr. Furkan Gözükara Hello, I'm trying to train LoRa on objects but it's not working. What's the reason? All the sources I saw were for faces
Fwell it works but object training is a different methodology
Fi can give private consultation
RHow?
KI am looking for any dev task.
Cold you help me?
Cold you help me?
Fsorry i dont have any tasks
P@Dr. Furkan Gözükara If I am training a model of my son, should I use reference images of boys his age instead of the ones I would use to train my own LORA?
S
Slooking better now.
Fif you can find yes better class dataset
FBS 2 brings 82% speed increase on A100
S
