Furkan Gözükara patreon

Best Settings For The END of Photography - Use AI to Make Your Own Studio Photos, FREE Via DreamBooth Training

🕑 Added 2023-07-04 17:41:18 +0000 UTC

Comments

Furkan Gözükara

not yet but i will hopefully make research for this. stay tuned. sorry for delay

Khoa Vo

Do you have a settings file that will work for Kohya for SD 1.5 similar to the one you provide for SDXL dreambooth training?

Furkan Gözükara

Yes you can delete all

Luis Arellano

Thank you for that! any paths were I can delete the checkpoints after the model is finished? I assume I don't need to keep all of them

Furkan Gözükara

if you delete with jupyter lab they will be sent to trash bin execute this : rm -rf /workspace/.Trash-0

Luis Arellano

I've been deleting files but when I train the models, these end up taking so much space. is that normal? how much space do you usually need to train these models? i'm using 12 photos and the settings you described. Any other folders that I should delete?

Furkan Gözükara

you can delete folders and objects with rm -r file_path you can also turn off machine and expand your storage

Luis Arellano

It's weird, I did selected a different (i assume checkpoint) from the trainings and that worked. Now however, i get a different error. No space left on the device. I'm using podrun. 40 GB Disk / 75 GB Pod Volume. I though that would be enough but seems that training these models takes up so much space.. any good advice for how to clear space? should I delete some of the checkpoints from my training?

Furkan Gözükara

that is very weird. where you are trying to generate on automatic1111? you can use your trained model on the latest version of automatic1111. so have 2 installations. 1 latest 1 dreambooth extension

Luis Arellano

I just did a trainning, but when trying to generate an image I get this error: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! any ideas?

Furkan Gözükara

it increases the emphasis of the prompt explained here : https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis

Manpreet Singh

What's the role of the "1.1" in the inference prompt? Eg: (ohwx man:1.1)

Furkan Gözükara

that repo is not maintained anymore much. sadly i don't know

Manpreet Singh

Thanks!

Manpreet Singh

Do you have any recommendations on how to take these WebUI settings and use them directly on shivamshrirao's dreambooth repo (I think the DB extension is based on that) and run training and inference via a python script?

Furkan Gözükara

you are welcome. realistic vision 5.1 supposed to be better version of 2.0. sdxl is far superior than sd 1.5 if you do dreambooth training.

Manpreet Singh

Is there a significant difference between 1) using Realistic Vision 2.0 and 5.1 and also 2) using SDXL instead of 1.5? Trying to see what to include in my personal "v1". Thanks for your amazing tutorials and detailed instructions btw!

Furkan Gözükara

all same settings. i haven't found a better way yet. but i suggest you to do SDXL DreamBooth now instead of SD 1.5. hopefully a tutorial coming but config files shared here already : https://www.patreon.com/posts/very-best-for-of-89213064

Manpreet Singh

Ok, thanks for sharing. Do all other settings remain exactly the same or are there some differences?

Furkan Gözükara

i suggest to use latest realistic vision which is 5.1 : https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE

Manpreet Singh

The first setting screenshot SS_1.png shows Realistic Vision 3.0 with VAE as the model to use. Whereas the YouTube video talks about Realistic Vision 2.0. Which one would you recommend to use? And what are the differences between the two?

Furkan Gözükara

yep leave it empty as in pictures

Tiago4D

Hi! what about "advanced" options on "settings"? I can leave it empty?

Furkan Gözükara

you are welcome. thanks a lot for the info

Daniel Limia Aspas

So for everyone reading this, as Furkan pointed, reg (class) images need to be exactly the same resolution as the training images. I was using the 1024x1024 dataset instead of the 768x1024. Using the right one worked. On another note: Automatic1111 doesn't allow training Dreambooth on an inpainting model, and gives the error above, which is very verbose but it means basically that. If inpainting is required, then one can merge the resulting Dreambooth final model with the SD-1.5 inpainting model post-training. Thanks again Furkan

Daniel Limia Aspas

Interesting. Using a non inpainting version casues issues with the trained model when using it to inpaint the face of the subject. But maybe I can merge the sd-1.5-inpainting model afterwards

Furkan Gözükara

inpainting may not work. never tried it. i use regular version

Daniel Limia Aspas

Thank you, I will apply your feedback. By the way, when loading Realistic Vision 4 (inpainting) as a source checkpoint and creating a new Dreambooth model, it fails to create the model. Does it happen to you too? Error: 2023-08-01T00:06:05.203719455Z size mismatch for conv_in.weight: copying a param with shape torch.Size([320, 9, 3, 3]) from checkpoint, the shape in current model is torch.Size([320, 4, 3, 3]).

Furkan Gözükara

Your issue is class images and train images not exactly same resolution

Furkan Gözükara

If fp32 available prefer it. Check hugging face of that model. I would suggest you to test realistic vision 5.1. Just released recently

Furkan Gözükara

Yes that is why. Resolutions have to be exactly same. Use 768 x 1024 I have them here https://www.patreon.com/posts/4k-2700-real-84053021

Furkan Gözükara

Yep did a fresh install with the scripts I shared here : https://www.patreon.com/posts/runpod-auto-84716845

Daniel Limia Aspas

Also, another question if you don't mind answering: is using fp32 base models necessary? or using fp16 is sufficient? Asking because currently one of the best realism models (epiCRealism) does not have a fp32 version

Daniel Limia Aspas

Can it be because i am training 768x1024 images? Maybe i should use the 768x1024 reg images dataset?

Daniel Limia Aspas

Disabling class matching seems to work? (not sure): https://github.com/d8ahazard/sd_dreambooth_extension/issues/1180

Daniel Limia Aspas

Thank you! I am using the same dataset. Are you using the latest Stable Diffusion runpod template? (v10)

Furkan Gözükara

I just verified and it didn't generate any images. used 1024x1024 training images. used the images downloaded from here as class images : https://www.patreon.com/posts/4k-2700-real-84053021

Furkan Gözükara

with 8 GB vram unfortunately you can only do LoRA training. i have a LoRA tutorial here : https://youtu.be/TpuDOsuKIBo

Furkan Gözükara

hello. are your training images are also exactly 1024x1024? can you verify that? so sorry for late reply i was on vacation . i will test this myself now

Alex May

It turns out that if I have an 8 GB video card, then I will not be able to use this training? Or is it possible, but with different settings?

Daniel Limia Aspas

Hi! when using your classification images dataset posted, Dreambooth does not seem to pick them up well and it always starts generating new ones. I am using the 1024x1024 dataset, max resolution set to 1024px. I downloaded them to runpod using wget and they're in the right folder.

Furkan Gözükara

awesome ty so much. sorry for late reply i am on vacation. hopefully even better videos coming soon

Furkan Gözükara

awesome. yes cropper will just crop and leave max size. then you need to resize.

Crypto Melon

Thanks, I think wget actually did solve it :) I had already used A1111 to resize as shown in the video. Love your content by the way. As someone learning this for fun, it's always very clear and easy to follow - thanks for all that you do!

Furkan Gözükara

how exactly did you upload them to the runpod? drag and drop may not work very well. runpodctl works great. moreover you should use wget to download into runpod. cropper.py will not resize them into the correct resolution. it will only crop to the target aspect ratio as much as possible with highest resolution. so please use automatic1111 to resize them into the exact resolution as i have shown in the video

Crypto Melon

Hi, I have two problems (using runpod) 1. Dreambooth isn't finding my class images and is trying to generate new ones. My max res is "1024", class images are your "768x1024", and I used your cropper.py + SD training to make my training images "768x1024" also. I thought I got around this by turning on "Disable Class Matching" in the testing tab, but t hat created problem #2: 2. log error: 2023-07-23T02:58:08.048664108Z ValueError: 'Training Images/man_training/00812-0-eric-ardito-GdWwogJ4TOk-unsplash.png' is not in list. *I checked this file and it's able to be opened, everything seems fine with it. I've googled my way through other problems along the way, but I can't get past this. Any help would be massively appreciated! *edit: I just noticed that I didn't use your instructions to use the class images in runpod. I just drag and dropped them into a runpod folder that I created, after downloading and extracting on my pc. They do seem fine to me with this method, however.

Daniel Chen

Yes, you're right. It's the problem of the resolution of the training images. But, I was processed these images by cropper.py, I thought it was correct but which is not. After I prepared my training images with birme, everything was OK.

Furkan Gözükara

yes Kohya can do both but implementation differs. DreamBooth extension fo Automatic1111 web UI can also train both DreamBooth and LoRA. so Kohya and that extension can be compared

Furkan Gözükara

hello i am so sorry for late reply. the patreon didn't show me your reply. do you still need help? please private message me

Furkan Gözükara

hello. sorry for late reply. this usually happens when the resolutions of images are mis matching. can you verify that resolution of class images, training images are all same? also you have set the max resolution of training?

Daniel Chen

I have generated about 2700 images by using t2i, and after I click the training button, cmd checked those images, and said we need about 2400 classification images and start generating … why? I think we are fine with the numbers

Максим Сутковенко

RAW photo, a portrait photo of (ohwx man:1.3), traveler clothes, standing in the forest, natural skin, 8k uhd, high quality, film grain, Fujifilm XT3 Negative prompt: (deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime), text, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, BadDream, UnrealisticDream Steps: 25, Size: 640x960, Seed: 441072722, Sampler: DPM++ SDE Karras, CFG scale: 7

Furkan Gözükara

Can you write here prompts? I can test on my training and give you feedback

Максим Сутковенко

But I trie to use realism prompts...

Furkan Gözükara

Hello. I haven't tested thoroughly. But with realism not every prompt works. I am working on a styling workflow as well. Styling is different than realism.

Максим Сутковенко

Hey, did you try another interesting prompts, which generate you? Because I can not any good results with other prompts, except for yours (provided in video.)

Furkan Gözükara

ok probably that is why. if your train images are 512x512, make max resolution 512 and use 512x512 class images dataset

Samuel

Sure. Max res 1024. Image res train 512x512. 14 images training

Furkan Gözükara

your setup looks correct. may i learn your set max resolution? your training images resolution? also how many images you have? so sorry for late reply

Samuel

Why after putting : https://gyazo.com/07bea5d8a2b4831a354a3828ac18a31e And having: https://gyazo.com/c4871748b12e61de7797bf6352a56956 It's telling me I need to create 1400 images class and start generating??? I have the class images... I dont understand.

Constantine Tenma

Since Kohya can train both Dreambooth and LoRA, as well as anything else, really, what exactly are you going to compare?

Furkan Gözükara

Actually I find Kohya inferior than DreamBooth extension. I will hopefully make a new video for LoRA better settings as well and I plan to compare Kohya and also DreamBooth extension

Broadband

Wanted to thank you so much for this tutorial. I've been trying to train myself for ages, but as a long hair clean shaven guy the output would always end up too masculine using "man" class or feminine using "person". Using male reg images and the man classifier It looked great at 4200 steps. I know I didn't have has much luck with your LORA tutorial last time, but will try again. In that you're using 8bitadam instead of Lion like here. Do you have an updated configuration of training lora with Lion or will you update in the next few months? Also do you find Koyha is always best for LORA training?