Fye you need to understand how training happens. long thing to explain. you need to watch all tutorials 
FarXiv.org
Large text-to-image models achieved a remarkable leap in the evolution of AI, enabling high-quality and diverse synthesis of images from a given text prompt. However, these models lack the ability to mimic the appearance of subjects in a given reference set and synthesize novel renditions of them in different contexts. In this work, we present a...
Lwhen completing training with dreambooth, why do we merge the trained model with the original model used?
Fwe dont merge
Lany reason why some people merge and some do not?
Fthey could be merging to get other model style
Fbut no need to merge with base model you trained on
Ecould you tell us which models did you use to train your lest set for the costumer you said about ealier? when i tried a few models from there it wasnt very accurate. Also, as discussed here before, why use captions when you use regularlazation images?
D@Dr. Furkan Gözükara Sir are we talking about this 
Fi use SDXL 1.0 base
Fcaption didnt affect much
Fi think most important part is training images quality and selection
Eoh i meant what captioning model sorry
Fit is a long thing to explain here
Fi used blip2
Fyou know i shared on patreon
Eactually the set you helped me in private (eliminated a few images) - turned out my best model ever (out of around 10 made)
Dok ok , i will just go through them.. i thought you are speaking about the captioning.. NVM , will look into your tutorials ..
Ei understand its blip2 but when opening the captioning webui there's a really huge list of models thats the one im talking about
Fas you watch more you will understand more haha i can also make private session and explain you if you are interested in professionally
i can also make private session and explain you if you are interested in professionallyEbut if you say captioning didnt matter much it doesnt matter
Fblip2 is stand alone
DI am now seeing your controlnet tutorial sir. Going ot to take the nexct step into controlnet , and then dreambooth as you suggested
F
DYouTubeSECourses
I show how to install Automatic1111 Web UI & ControlNet extension installation from scratch in this video. Moreover I show how to make amazing QR codes and inpainting and out painting of ControlNet which are very similar to Photoshop generative fill and Midjourney zoom out. Furthermore, I explain and show what are Canny, Depth, Normal, OpenPose...
EI've actually tested so many models but forgot to document (or rather was in a hurry from excitment) - is there a way to take a checkpoint and somehow check it to see if i used captions to train it?
L@Dr. Furkan Gözükara training with dreambooth with lion, can I use it without using classification images?
Lfor this setting i set it to 0 or leave it to 5?
Fdont use it
Fuse my class images
For generate from txt2img tab
Fmanually
Lfor 30 images how many images in my classification dataset and how much "class images per instance images" should i set it to?
M1/1 - for one classification image one instance image. No more need. Use Furkan's reg images.
MAfter the recent Prodigy testing, I have now styled with AdamW. Using the same image set. So far the Prodigy has won.
MHere is my config for style. I used 10 images with 10 repeats, captions and 10 epoch
Fwith dreambooth extension
Fif class images are good quality
Fi suggest 100 per instance
Fso you would need 3000
Li will use your class images
Fok make it 100
Lok so i set it to 100 and use 3000 images
thank you
thank you
L@Dr. Furkan Gözükara my dataset is 512px x 768px and I have trained max resolution 768px. Any particupar reason why the generated images have more ressemblance with my dataset when generate at 768px x 768px compared to those generated 512px x 768px?
GHi all, I'm a brand new Patreon supporter. I'm following the tutorial to make a LoRA training of myself in Kaggle. But I get the following error:
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 14.76 GiB total capacity; 13.17 GiB already allocated; 7.75 MiB free; 13.46 GiB reserved in total by PyTorch Here is my training command: accelerate launch --num_cpu_threads_per_process=2 "./sdxl_train_network.py" --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" --train_data_dir="/kaggle/working/results/img" --reg_data_dir="/kaggle/working/results/reg" --resolution="1024,1024" --output_dir="/kaggle/working/results/model" --logging_dir="/kaggle/working/results/log" --network_alpha="1" --save_model_as=safetensors --network_module=networks.lora --text_encoder_lr=0.0004 --unet_lr=0.0004 --network_dim=32 --output_name="kaggle_glpr123" --lr_scheduler_num_cycles="8" --no_half_vae --learning_rate="0.0004" --lr_scheduler="constant" --train_batch_size="1" --max_train_steps="6400" --save_every_n_epochs="1" --mixed_precision="fp16" --save_precision="fp16" --cache_latents --optimizer_type="Adafactor" --optimizer_args scale_parameter=False relative_step=False warmup_init=False --max_data_loader_n_workers="0" --bucket_reso_steps=64 --full_fp16 --xformers --bucket_no_upscale --noise_offset=0.0 --lowram --max_grad_norm=0.0MIf I'm tagging with WD14 and it's a drawing style, but I don't want it to be an asian/manga type model, which one should I train on?
MMeinaMix maybe?
Fwhich model? also did you enable bucketing?
