Software Engineering Courses (SECourses)•3y ago

@Furkan Gözükara SECourses what is the best approach for Dreambooth training in terms of time and r

@Furkan Gözükara SECourses what is the best approach for Dreambooth training in terms of time and resource (RAM/VRAM) consumption for a single GPU (3060 12GB). Are the Kohya_ss and A1111 the most efficient for model training or does these libraries ColossalAI (https://github.com/hpcaitech/ColossalAI#AIGC) and VoltaML (https://github.com/VoltaML/voltaML-fast-stable-diffusion) offer better training speed and efficient memory utilization? I don't mind the OS, be it windows or Linux.

saudhach•5/25/23, 8:14 AM

Hii all of you

Ssaudhach Hii all of you

ashley•5/25/23, 8:20 AM

Hello

ashley•5/25/23, 8:20 AM

How does one go about cloning the configuration/persistent volumes for a RunPod pod? Is that done using Cloud Sync or something else?

AvidyaCM•5/25/23, 8:21 AM

@Furkan Gözükara SECourses https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html
very interesting article from popular startup ceo making chip company. curious ur thoughts as he says NVIDIA is like apple and they will make android (working with AMD)

the singularity is nearer

the tiny corp raised $5.1M

Here we go again. I started another company. The money is in the bank.

AvidyaCM•5/25/23, 8:22 AM

code here: https://github.com/geohot/tinygrad

GitHub

GitHub - geohot/tinygrad: You like pytorch? You like micrograd? You...

You like pytorch? You like micrograd? You love tinygrad!

GitHub - geohot/tinygrad: You like pytorch? You like micrograd? You love tinygrad!

Aashley Hello

saudhach•5/25/23, 8:27 AM

hopeyan•5/25/23, 9:35 AM

@Furkan Gözükara SECourses when choose lion optimizer , what is the best setting for batch size ,gradient accumulation steps ,learning rate

majidzendaniyan•5/25/23, 9:46 AM

Is it possible to create such an image with stable diffusion? What tips are needed to make it?

ManicWulf•5/25/23, 9:54 AM

I've just seen the Video to Anime tutorial and now I wonder, would it be possible/feasible to apply this to something other than a person? Specifically I'm thinking about taking footage of me cooking (no person in frame, only hands/food/tools) and style change it to anime style.
If it's doable, would I have to do the training for each new recipe/video or could I just do the training with a few videos and then use the same training data for more?

Furkan Gözükara SECourses•5/25/23, 10:36 AM

wow i am having so many meetings with companies recently

DDawood @Furkan Gözükara SECourses I basically binge watched all of your dreambooth trai...

Furkan Gözükara SECourses•5/25/23, 10:36 AM

did you try to train with dreambooth?

Furkan Gözükara SECourses•5/25/23, 10:36 AM

i think it works better

NNull does anyone know what is Stable Diffusion by Automatic1111, why do we need it if...

Furkan Gözükara SECourses•5/25/23, 10:37 AM

we use automatic1111 for inference. kohya for training

NNull Click to see attachment

Furkan Gözükara SECourses•5/25/23, 10:37 AM

you need to add more memory optimization

Furkan Gözükara SECourses•5/25/23, 10:37 AM

there are advanced settings

Furkan Gözükara SECourses•5/25/23, 10:38 AM

gradient checkpoint memory attention xformers

Kkevh Nevermind, I'm able to upload, I just run into this error: PIL.UnidentifiedImage...

Furkan Gözükara SECourses•5/25/23, 10:38 AM

it still works. are you using dev version of the atuomatic1111? try with dev branch. i tested the video workflow on runpod few days ago for a patreon subscriber. it worked very well

DDeadMan1999 @Furkan Gözükara SECourses what is the best approach for Dreambooth training in...

Furkan Gözükara SECourses•5/25/23, 10:39 AM

I havent tested them but i believe most optimized script is the dreambooth extension of the automatic1111

AAvidyaCM @Furkan Gözükara SECourses https://geohot.github.io/blog/jekyll/update/2023/05/2...

Furkan Gözükara SECourses•5/25/23, 10:40 AM

i think they will take money and get disappeared :d

Hhopeyan @Furkan Gözükara SECourses when choose lion optimizer , what is the best settin...

Furkan Gözükara SECourses•5/25/23, 10:40 AM

best for training person is batch size 1, gradient 1, 1e-7 lr for lion

Mmajidzendaniyan Is it possible to create such an image with stable diffusion? What tips are need...

Furkan Gözükara SECourses•5/25/23, 10:41 AM

100% possible but i dont know what prompt can be used

MManicWulf I've just seen the Video to Anime tutorial and now I wonder, would it be possibl...

Furkan Gözükara SECourses•5/25/23, 10:41 AM

you can do training only 1 time, train your visible body parts that will be used in video, and follow the logic

Furkan Gözükara SECourses•5/25/23, 10:42 AM

it should work very well

Furkan Gözükara SECourses•5/25/23, 10:42 AM

i am also working on an easier way that can also be used for your purpose

FFurkan Gözükara SECourses wow i am having so many meetings with companies recently

ashley•5/25/23, 10:42 AM

Can we have a meeting with you too?

Aashley Can we have a meeting with you too? 😁

Furkan Gözükara SECourses•5/25/23, 10:43 AM

sure i am open to private consultation

FFurkan Gözükara SECourses did you try to train with dreambooth?

Dawood•5/25/23, 10:45 AM

Yes im training with dreambooth but i trained on rev animated which i guess is a merge of soo many models so that's why not good

DDawood Yes im training with dreambooth but i trained on rev animated which i guess is a...

Furkan Gözükara SECourses•5/25/23, 10:46 AM

yes some models may not work good for training

Furkan Gözükara SECourses•5/25/23, 10:46 AM

for example realistic vision works good

Dawood•5/25/23, 10:46 AM

Would you know a way to know which ones would or is it just trial and error?

Furkan Gözükara SECourses•5/25/23, 10:46 AM

trial and error

Dawood•5/25/23, 10:47 AM

Also have you ever tried stylizing your images using control net before training and then using them as training images

Dawood•5/25/23, 10:47 AM

If you want a specific stylized look

Dawood•5/25/23, 10:48 AM

Wouldn't that mean the model welcomes these training images better because they look like the same style as the model you train them on

ashley•5/25/23, 10:49 AM

I also didn't have much success with revAnimated, made the person look like a doll or something

FFurkan Gözükara SECourses best for training person is batch size 1, gradient 1, 1e-7 lr for lion

hopeyan•5/25/23, 10:49 AM

for style what is the best settings

FFurkan Gözükara SECourses best for training person is batch size 1, gradient 1, 1e-7 lr for lion

ashley•5/25/23, 10:51 AM

1e-7 for both learning rate and text encoder learning rate?

hopeyan•5/25/23, 10:55 AM

Some version of dreambooth ,when check lora,there is no text input filed for learning rate. Only Lora Unet learning rate and Tex encoder

brenkhof•5/25/23, 11:05 AM

Hey, what is the current best solution for this task? Thanks!

Bbrenkhof Hey, what is the current best solution for this task? Thanks!

felixsanz•5/25/23, 11:50 AM

inpainting

brenkhof•5/25/23, 12:11 PM

And should I use separate models? Or merge them after training?

Dawood•5/25/23, 12:50 PM

After training, i am converting it to ckpt but then I can't download from g drive idk what's happening it keeps saying error when it's almost done, so i have downloaded the weights how do i convert them to ckpt?

Bbrenkhof And should I use separate models? Or merge them after training?

Dawood•5/25/23, 1:07 PM

No don't merge you will lose both, try to train the persons face and then a lora for clothes

Dawood•5/25/23, 1:07 PM

Then try to prompt them together

Dawood•5/25/23, 1:08 PM

You can also do this with controlnet and inpainting but that's if the model and the clothes are already in a pose you need

Kallamamran•5/25/23, 1:15 PM

I think we need a tutorial for this -> https://www.youtube.com/watch?v=r-ZYOlZTLDA

YouTubeTroubleChute

🤯 2X SPEED BOOST FOR STABLE DIFFUSION | HUGE NEW NEWS |

Nvidia has announced HUGE news: 2x improvement in speed for Stable Diffusion and more with the latest driver. Using it is a little more complicated, but the speed boost is there! Exciting things coming in the future of AI. This video covers installing and using the new ONNX/Olive models and converter, as well as converting models, generating ima...

FFurkan Gözükara SECourses gradient checkpoint memory attention xformers

Null•5/25/23, 1:47 PM

these?

Null•5/25/23, 1:48 PM

still not working :\

bigbird - ビッグバード (fal.ai)•5/25/23, 2:07 PM

hi everyone! wanted to share a demo of running A1111 with a fal-serverless backend. The idea here is that you just run the UI locally but the GPU parts happen in the cloud. So instead of having a machine on all the time, you would only have a machine running when you are generating images which should save some $. Here is a video of it in action: https://www.youtube.com/watch?v=EIxF7TKk6wg

I am looking for feedback. Would you want to use this?

YouTubeFeatures and Labels

Serverless AUTOMATIC1111

In this video we demonstrate a version of AUTOMATIC1111 StableDiffusion UI that runs in a serverless manner using fal-serverless.

Bbigbird - ビッグバード (fal.ai)hi everyone! wanted to share a demo of running A1111 with a fal-serverless backe...

ashley•5/25/23, 2:13 PM

This is really cool, will be nice if you can implement all the other tabs and especially add support for Dreambooth

@Furkan Gözükara SECourses what is the best approach for Dreambooth training in terms of time and r

Similar Threads