then i would be better off training SDXL at lower resolution?
then i would be better off training SDXL at lower resolution?
Fi think that wouldnt be good
Fi would compare supir low text cfg upscaled
Fvs raw training
Fworking
F
F
J@Dr. Furkan Gözükara You said some things will be fixxed in new supir one click installer for runpod right? Do you know when you will release it or do you edit the file in the patreon?
Frunpod is working nice
Fif you mean runpod proxy i dont know if fixed
Fuse gradio live share
JNo sorry, had some errors last time dont know if you remember
Jbut they seem to be fixed somehow
FKohya is a little bit faster than OneTrainer when training Stable Diffusion XL (SDXL) with no VRAM optimizations and with half precision BF16
F
R@Dr. Furkan Gözükara do you know kohya update « Fix issue with vae path validation » and « Add new setup menu option to install Triton 2.1.0 for Windows » does for training optimisation?
Fpath is working if you mean vae he said fixed
Ftriton i didnt test
Dcan anyone reccomend ways to get sam detectors in comfy to recognize non human objects?
Dthere is a paper about a zero shot sam detection but cant find anything ported to comfy , being able to load an image to guide the sam detection for masking would be awesome
RIl just not sure how vae is used in training. Never heard anyone talk about it in training
RIs it beneficial for training? With which vae model
Fno you can ignore thinking it
Fit will use embedded vae
RRight
A
AWhat did I do wrong
Fthis is a bad dataset
Fyour config is also inaccurate not the one shared on patreon
KCan models be trained without text-file with OneTrainer?
Afixed this part
Fi wonder if conflicting
Fyou are providing multiple values
Fuse interface
AWhat multiple values?
Fthose values you provide additional arguments
CThank you so much! But I ma getting the following error ## fine_tune.py: error: unrecognized arguments: --no_half_vae ## the following is the full training commad ====== :: ======= accelerate launch --num_cpu_threads_per_process=4 "/kaggle/working/kohya_ss/sd-scripts/fine_tune.py" --max_grad_norm=0.0 --no_half_vae --ddp_timeout=10000000 --ddp_gradient_as_bucket_view --bucket_no_upscale --bucket_reso_steps=64 --cache_latents --cache_latents_to_disk --dataset_repeats="40" --enable_bucket --min_bucket_reso=256 --max_bucket_reso=2048 --in_json="/meta_lat.json" --learning_rate="7e-07" --learning_rate_te="7e-07" --logging_dir="G:\automatic1111\stable-diffusion-webui\models\Stable-diffusion\sd15train\logs" --lr_scheduler="constant" --max_data_loader_n_workers="0" --resolution="768,768" --max_train_steps="26240" --mixed_precision="no" --optimizer_args scale_parameter=False relative_step=False warmup_init=False weight_decay=0.01 --optimizer_type="Adafactor" --output_dir="G:\automatic1111\stable-diffusion-webui\models\Stable-diffusion\sd15train" --output_name="tier_2_quality_speed" --pretrained_model_name_or_path="MonsterMMORPG/sd15_best_realism" --save_every_n_epochs="1" --save_model_as=safetensors --save_precision="float" --train_batch_size="1" --train_data_dir="/kaggle/input/test-all" --train_text_encoder --xformers
Fremove --no_half_vae
Fyou loaded dreambooth tab right?
CI was trying to fintune so loaded
CI am using using image captions to fintune , can you please point out if that is okay
CThank you again!
CDatasets: 800 images resolution 350by 450, topic: Various paintings,
AOoh onetrainer seems much better than kohya
AI did an export, how can I load those configs next time? I dont see import
CCan we use onetrainer on Kaggle or unpod of colab?
Fthat is why probably
Fdo dreambooth
Fdont use reg images
Fit will be same
CThanks!