@Dr. Furkan Gözükara kohya dont train v1.5 models ? i tried realistic_vision_5.1 and got errors in
@Dr. Furkan Gözükara kohya dont train v1.5 models ? i tried realistic_vision_5.1 and got errors in load unet model.
FTensorRT wont work on AMD. it is NVIDIA feature
Fit supports
Fyou need to uncheck sdxl option
Fin first page
Pi uncheck this and checked v_prediction then booth unchecked?
PI'll try, another thing I noticed is that in sdxl with 200 repeat and 1 epoch I got 1280 steps. Without sdxl I got 6500 steps. My dataset is 16.
Fit makes 3200
Fwithout reg images
PEven if I provide regularization images?
Fif you provide
Fit becomes 6400
Pyes that's what happened, sdxl then generates fewer steps. So it must be right, I'll try to run it with SDXL unchecked.
Pi can specify vae-ft-mse-840000 for training ?
Pthere must be a bug or unsupported parameter.
CDoes Kaggle allow Automatic or Comfy? I tried Furkans notebook for Fooocus which was horrible UI to use 
୧Yes it allows automatic
୧There's a notebook for it too
CAlright I'll search for that notebook thx then I know I won't be wasting my time 
Zanyone heard of a program called unstable diffusion which is uncensored?
Pi think this should be the right name for stable diffusion
Phaha
Zlol
P
JAhh okay. Yeah the 6GB was less than expected. I restarted and changed the batch sizes to 1, and then it said I had 10GB. I’ll fiddle around a little more. Thanks so much for making the video. It’s been really helpful.
NHello Furkan,
I did all the installation again from scratch however, when I double click “webui-user.bat” in order to launch SD I get the enclosed pip error.
What did I do wrong and how do I rectify it?
Thank you for your help.
I did all the installation again from scratch however, when I double click “webui-user.bat” in order to launch SD I get the enclosed pip error.
What did I do wrong and how do I rectify it?
Thank you for your help.
![Digital [Starburst]](https://cdn.discordapp.com/avatars/220961909447458817/8ecbbb3132fc6e44312f3d6d2ae3df69.webp?size=40)
DThere's a live summit right now about Google Cloud and how to use it with something called Vertex. https://cloudonair.withgoogle.com/events/summit-applied-ml-summit-23/watch?talk=kn_demo
Discover how you and your fellow developers, data scientists, and ML engineers can build a new generation of applications and boost productivity with the latest generative AI research and technology from Google Cloud.
Bdoes fine-tuning on a pruned fp16 model work?
KFu**ing hate this:
KOnly some models suddenly
Fthis happens when model cooked as well
THello everyone.
NDear Furkan, I solved the problem. TY
T@Dr. Furkan Gözükara are you available for paid consultation?
TPlease check your DMs.
Fyou can
Fit allows automatic1111
Fwe have notebook for that
FFi would reinstall python
Fif it cant find pipo
Fyes
Fhello
Fyes i am checking
RHi I'm trying to train an SD1.5 model with my face on Kohya SS Dreambooth (checkpoint training, not lora) and wanted to see what settings you would recommend. Appreciate it! Here is my current setup:
CPU: AMD 5900x (12core/24processors)
GPU: RTX 3090 (24gb vram)
RAM: 16GB
Training images: 16 images / 40 repeat
Regularization images: 5200 images from your patreon post / 1 repeat
Image size: 768 x 1024
Kohya setting:
accelerate launch
--num_cpu_threads_per_process=24 "./train_db.py"
--pretrained_model_name_or_path="C:/Users/foodmass/Stable-diffusion/t3_Ver09.safetensors"
--train_data_dir="C:/Users/foodmass/Documents/asdf-t3-v2\img"
--reg_data_dir="C:/Users/foodmass/Documents/asdf-t3-v2\reg"
--resolution="768,1024"
--output_dir="C:/Users/foodmass/Documents/asdf-t3-v2\model"
--logging_dir="C:/Users/foodmass/Documents/asdf-t3-v2\log"
--save_model_as=safetensors
--full_bf16
--output_name="asdf"
--lr_scheduler_num_cycles="8"
--max_data_loader_n_workers="1"
--learning_rate_te="1e-06"
--learning_rate="0.0001"
--lr_scheduler="constant"
--train_batch_size="4"
--max_train_steps="2560"
--save_every_n_epochs="1"
--mixed_precision="bf16"
--save_precision="bf16"
--cache_latents
--optimizer_type="Adafactor"
--max_data_loader_n_workers="1"
--clip_skip=2
--bucket_reso_steps=64
--gradient_checkpointing
--xformers
--bucket_no_upscale
--noise_offset=0.0
CPU: AMD 5900x (12core/24processors)
GPU: RTX 3090 (24gb vram)
RAM: 16GB
Training images: 16 images / 40 repeat
Regularization images: 5200 images from your patreon post / 1 repeat
Image size: 768 x 1024
Kohya setting:
accelerate launch
--num_cpu_threads_per_process=24 "./train_db.py"
--pretrained_model_name_or_path="C:/Users/foodmass/Stable-diffusion/t3_Ver09.safetensors"
--train_data_dir="C:/Users/foodmass/Documents/asdf-t3-v2\img"
--reg_data_dir="C:/Users/foodmass/Documents/asdf-t3-v2\reg"
--resolution="768,1024"
--output_dir="C:/Users/foodmass/Documents/asdf-t3-v2\model"
--logging_dir="C:/Users/foodmass/Documents/asdf-t3-v2\log"
--save_model_as=safetensors
--full_bf16
--output_name="asdf"
--lr_scheduler_num_cycles="8"
--max_data_loader_n_workers="1"
--learning_rate_te="1e-06"
--learning_rate="0.0001"
--lr_scheduler="constant"
--train_batch_size="4"
--max_train_steps="2560"
--save_every_n_epochs="1"
--mixed_precision="bf16"
--save_precision="bf16"
--cache_latents
--optimizer_type="Adafactor"
--max_data_loader_n_workers="1"
--clip_skip=2
--bucket_reso_steps=64
--gradient_checkpointing
--xformers
--bucket_no_upscale
--noise_offset=0.0
KIrritating when it suddenly starts doing this ona model it has not done it with previously
NThank you Furkan, I did exactly that and the problem is solved.
Fsadly i dont have best parameters for sd 1.5
Fi only have for sdxl
Bi have the problem where i cant generate images, is my problem not having the correct checkpoint?
