วีดีโอ

How can we log more items like time_step etc?

Thanks a lot

cant even open the file browser to pick a model "TCL was probably not installed correctly" holy shit trying to do anything with this makes me want to commit suicide

horrible tutorial

hello does kohya_ss upload my training images to a server or something? i have alot output traffic if i try to train with it

Can you give us New Tutorial they is change everything now :(

2:45 So g-r-e-a-t, thanks. I can finally check my tensorboards even after a Kohya/Onetrainer training is finished. Didn't how to access those logs folders in tensorboard anymore. Cheers for that. Very helpful video (b^-^)b

Juste efficace, merci 😉

help please it says File "D:\AI\KohyaTraining\kohya_ss\sd-scripts\train_db.py", line 118, in train text_encoder, vae, unet, load_stable_diffusion_format = train_util.load_target_model(args, weight_dtype, accelerator) File "D:\AI\KohyaTraining\kohya_ss\sd-scripts\library\train_util.py", line 4113, in load_target_model text_encoder, vae, unet, load_stable_diffusion_format = _load_target_model( File "D:\AI\KohyaTraining\kohya_ss\sd-scripts\library\train_util.py", line 4067, in _load_target_model text_encoder, vae, unet = model_util.load_models_from_stable_diffusion_checkpoint( File "D:\AI\KohyaTraining\kohya_ss\sd-scripts\library\model_util.py", line 1008, in load_models_from_stable_diffusion_checkpoint info = unet.load_state_dict(converted_unet_checkpoint) File "D:\AI\KohyaTraining\kohya_ss\venv\lib\site-packages\torch n\modules\module.py", line 2152, in load_state_dict raise RuntimeError('Error(s) in loading state_dict for {}: \t{}'.format( RuntimeError: Error(s) in loading state_dict for UNet2DConditionModel: size mismatch for down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for down_blocks.0.attentions.0.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for down_blocks.0.attentions.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for down_blocks.1.attentions.0.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for down_blocks.1.attentions.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for down_blocks.2.attentions.0.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for down_blocks.2.attentions.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for up_blocks.1.attentions.0.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for up_blocks.1.attentions.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for up_blocks.1.attentions.2.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for up_blocks.2.attentions.0.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for up_blocks.2.attentions.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for up_blocks.2.attentions.2.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([640, 768]) from checkpoint, the shape in current model is torch.Size([640, 1024]). size mismatch for up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for up_blocks.3.attentions.0.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for up_blocks.3.attentions.1.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for up_blocks.3.attentions.2.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([320, 768]) from checkpoint, the shape in current model is torch.Size([320, 1024]). size mismatch for mid_block.attentions.0.transformer_blocks.0.attn2.to_k.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). size mismatch for mid_block.attentions.0.transformer_blocks.0.attn2.to_v.weight: copying a param with shape torch.Size([1280, 768]) from checkpoint, the shape in current model is torch.Size([1280, 1024]). Traceback (most recent call last): File "C:\Users\isodu\AppData\Local\Programs\Python\Python310\lib unpy.py", line 196, in _run_module_as_main return _run_code(code, main_globals, None, File "C:\Users\isodu\AppData\Local\Programs\Python\Python310\lib unpy.py", line 86, in _run_code exec(code, run_globals) File "D:\AI\KohyaTraining\kohya_ss\venv\Scripts\accelerate.exe\__main__.py", line 7, in <module> File "D:\AI\KohyaTraining\kohya_ss\venv\lib\site-packages\accelerate\commands\accelerate_cli.py", line 47, in main args.func(args) File "D:\AI\KohyaTraining\kohya_ss\venv\lib\site-packages\accelerate\commands\launch.py", line 1017, in launch_command simple_launcher(args) File "D:\AI\KohyaTraining\kohya_ss\venv\lib\site-packages\accelerate\commands\launch.py", line 637, in simple_launcher raise subprocess.CalledProcessError(returncode=process.returncode, cmd=cmd) subprocess.CalledProcessError: Command '['D:\\AI\\KohyaTraining\\kohya_ss\\venv\\Scripts\\python.exe', 'D:\\AI\\KohyaTraining\\kohya_ss/sd-scripts/train_db.py', '--bucket_no_upscale', '--bucket_reso_steps=64', '--cache_latents', '--enable_bucket', '--min_bucket_reso=256', '--max_bucket_reso=2048', '--learning_rate=1e-05', '--learning_rate_te=1e-05', '--logging_dir=D:/AI/kohya destination\\log', '--lr_scheduler=cosine', '--lr_scheduler_num_cycles=1', '--lr_warmup_steps=200', '--max_data_loader_n_workers=0', '--resolution=512,512', '--max_train_steps=2000', '--mixed_precision=fp16', '--optimizer_type=AdamW8bit', '--output_dir=D:/AI/kohya destination\\model', '--output_name=last', '--pretrained_model_name_or_path=D:/AI/sd.webui/webui/models/Stable-diffusion/dreamshaper_8.safetensors', '--save_every_n_epochs=1', '--save_model_as=safetensors', '--save_precision=fp16', '--train_batch_size=1', '--train_data_dir=D:/AI/kohya destination\\img', '--v2', '--xformers']' returned non-zero exit status 1.

oldie but goldie

I have a folder with 200+ images in it, and Kohya_SS says that there are not images in that folder.

Hello Bernard, I followed your excellent presentation step by step, but I cannot obtain models under the "safetensors" extension! In the "Models" folder there is only one .json file... Do you have a solution to offer me? Thanking you.

awesome, a tutorial that actualy tells you what you need to know. Well done.

Thank you for the tutorial Bernard. I was wondering if it was possible to generate a smaller file (I am thinking LoRA size with dreambooth results). Do you know if that can be accomplished?

I just had the exact same phenomenon over my house! Hundreds of confused geese. It was weird! So I googled "confused geese" and saw your video. Gettysburg, PA. 12 Jan 2024. 1615

Hola gracias por el tutorial. Hay alguna alternativa gratuita a Gigapixel que pueda hacer lo mismo???

This videos obsolete now, all can be done on tensor art.

hey! I recently started following your channel and watching your videos about Kohya, LoRa and other stuff... but I'm having difficulty with some things... I would like to know if you can help me create a LoRa and in the meantime clarify some doubts that I'm having... I'm willing to pay for your time. btw congratulations on your videos... they are very enlightening and dynamic haha

Hi Bernard, Thanks a million for your efforts and your tutorial, I watched so many tricks thanks to your video! I was wondering I could I create a specific part of a body I would like to focus on man's chest (muscles and hair) and I was wondering if I need to takejust chest training images or the full body pictures of a man. And what about regularization pics? Should I take just chest or face or the full body? I'm a bit confused. Thanks a million!

Merci Berny

Thank you so much for all your work! Learned a lot from this video. Subscribed!

I have a 4090, if i don't care about speed, should i do float precision for better accuracy? whats a good basic config for highest accuracy?

First, thank you for your Guide. But as other mentioned, it has been trained on a known person what makes it a bit hard to see if the values are correctly. Could you make a video with an unknown person?

Merci pour le tutorial et l'outil!!

MFW when you check OP's textual embeddings 😏 allfoursnaked corneosidedoggy sexyback bimbostyle busty_15

Amazing stuff! Thank you!

Dont work for me i dont know what the problem is in the preview i see a good face but in the last step i turns back to the bad face and the output is just a even worse oversharpend face

this works!! thank you! i was struggling so hard lol

How to train SDXL model with large number of images? like example 30k images

I know this might be outdated not but for some reason When I train using BF16 I can only use a batch size of 1 or I get out of memory. even at 512,512. for a test. How are you able to do 8 I have a 3090 by the way. followed your settings im using BF16 and 23 images.

I have trained the lora, and I have used it for stable diffusion, but I keep on getting black images, doesnt mater what i type in

Hi! @Bernard Maltais, do you have a relation whit HandyPrint?

grateful for video but damn that audio makes me wanna die

kohya_ss 和 webui 可以共用虚拟环境吗

Quebec Power

One problem that I'm running into is that it takes over three hours to train using 7500 steps, but looking at my CPU and GPU performance shows that I'm only using around 10% of my CPU and usually only around 3% or lower on my GPU.

Excellent presentation

🎯 Key Takeaways for quick navigation: 00:02 🌐 El video explica cómo recopilar imágenes de internet y prepararlas para crear un modelo LoRA utilizando Kohya SS GUI. 01:13 🖼️ Es importante seleccionar imágenes de alta resolución para el modelo LoRA, no importa la proporción. 03:19 📂 Se organiza una carpeta llamada "Jennifer Aniston" para recopilar las imágenes y otra subcarpeta llamada "image" para almacenarlas. 09:27 📸 Utiliza Gigapixel para mejorar la resolución de las imágenes y luego las guarda en formato JPEG. 24:10 🖼️ Emplea el Auto1-1-1 Web UI para mejorar la calidad de algunas imágenes, ajustando parámetros como el denoising strength. 28:03 📈 Refina las imágenes con Auto1-1-1 Web UI, manteniendo la apariencia original de los rostros para no comprometer el modelo final. 29:42 🖼️ La preparación de un conjunto de datos para crear un modelo LoRA comienza con la limpieza y mejora de la resolución de las imágenes. 32:18 📷 Después de la preparación de las imágenes, es importante agregar descripciones a cada una de ellas para enseñar al modelo sobre el tema relevante y no importantes detalles. 34:33 📝 Las descripciones se pueden validar abriendo cada archivo de texto junto con la imagen correspondiente para asegurarse de que sean precisas. 44:10 🚀 Una vez validadas las descripciones, se está listo para configurar el modelo LoRA y proceder con la fase dos del entrenamiento. Made with HARPA AI

Having issues with the Group images script, It only does 10 images at the time and only once. Windows 11 machine. Anyone else having that issue.

Your audio is fucking awful with an echo in it. We can't understand you. I just had to mute it and use CC instead.

These settings dont work. I keep getting CUDA out of memory issues even when using a 24 GB card. LoRa's train no problem. There's something off in these settings

45 minutes??

thank you, really clear and concise tutorial

Does anyone know why I continue to get - CalledProcessError: Command '['C:\\Users\\Name\\Desktop\\Vlad Diffusion\\kohya_ss\\venv\\Scripts\\python.exe', returned non-zero exit status 2. I can't find anything that fixes this issue and after watching so many videos on Kohya it feels like I'm the only one with this issue.

10:00 -- Question: Have you used the Internet since 1991 or just recently, like two years ago? Your image is smaller because, for some strange reason and logic, you think that dragging an image from Google Search into a folder will restore it to its original size which is completely wrong! You have taken the image from the temporary cache of your Internet browser! To get the right size of the image that you want to save, you need to access that image FROM the website, directly. But because you don't want to visit the whole site, you only need to use the direct link to the image. To do that: On the Google Search page, click on the image to open and see the Google preview of that image FROM the website, then right-click and select "Open image in new tab". And NOW and only NOW you have the right size image and can drag it into your folder (or right-click and save) ... not the smaller version of the image from the temporary cache!

🎯 Key Takeaways for quick navigation: The video is about creating a LoRA weight using kohya_ss GUI for model training. The process involves collecting images from the internet and enhancing them for model use. The speaker demonstrates searching for high-resolution images of Jennifer Aniston from the 1990s. The images are collected in a specific folder structure for processing. The speaker uses Gigapixel AI to upscale and clarify the images. Some images are processed using the Koya SS GUI tool to improve their quality further. The speaker discusses considerations for aspect ratios and cropping images. The images are being prepared for use in a model training process. Some images are discussed as potentially usable while others may require more improvement. The video highlights the workflow of collecting, enhancing, and preparing images for training. 32:18 The next stage involves adding text captions to the images to train the LoRA model. 32:49 Captions are added by pasting the image directory and using utility tools for text captioning. 33:20 The process involves configuring parameters like batch size, beams, and minimum caption length. 34:18 Captions start appearing next to each image file as the process proceeds. 35:01 Descriptions are reviewed and edited for accuracy and relevance. 35:26 Descriptions focus on clothing and essential details while avoiding unnecessary specifics. 40:10 Providing clothing details helps the model learn without enforcing specific attributes. 44:31 The process of captioning and validating images is complete, preparing for the next phase of training the LoRA model. Made with HARPA AI

Appreciate your tutorials. From all of your use and testing of SD, would you arrive at your best preference being Lora training and not Dreambooth ?

learned from the video, but tried 2 on the same settings and it resulted in a melted faces like the one you are showing in this video @ 80%. Is hires fix a requirement to make it work?

Thanks, now I've made 3 successful LoHA files and made a tutorial guide in spanish, if you want to look and made some corrections I would be so thankful, again, thank you very much

32:28 she looks more like a trans woman now :D