First if all u rocks. Thank u!! I did textual with 4 photos of myself 6 tokens 1500 steps... Unbelievably acuurate almost every time... I think that something git wring with your owl...
@@zonas7915 i dont think its the right way... I believe U should use 3-5 pics. And save a pt file every 500 steps and check the quality of the picture. For me i did 16000 steps, 1500 steps is the best. More than 1500 its starting to degrade... 6 tokens for me, and i get good results and i can manipulate the picture really easyly...
Hmm…apparently a personal taste. I think the dramatic voices and uncanny valley avatar draw my attention away from the content and make it hard to follow. But, from the comments, looks like I’m maybe in the minority.
Appreciate the content! The tech is moving so fast it really helps to have comparisons like this to keep up with all the new techniques available. Thanks again!
Think that's a fair assessment. There are some very clear differences between them. That said I think Automatic1111's Textual Inversion has potential, and I like that it has a bit of a modular aspect to it where you just "plug in" the embedding you want to use. The disadvantage there, I guess, is that each embedding pushes up the number of prompt words.
Not only that but you can actually run it on lower end hardware, if you want to do it locally it's the only option at the moment for a lot of users. It takes it's time though...
On the flipside with the 1111 Embeds you're not having to have a bunch of seperate 4+GB Model files for every inversion filling up your SSD/HDD. Plus when 1.5 publicly releases the embeds should presumably still work. Hopefully the code will improve over time.
@@NerdyRodent As a matter of interest have you tested whether there was any differential betwween the full fat and low fat Models? I've shied away from the 2GB version as it talks about a quality drop? Also loving the content. :)
I'm like a Celebrity 😎. Now that I got the clout, have you done Ubuntu Dreambooth install? Is it any different from the Windows WSL install you did or is it the same?
Interested in adding things to your AI Art? Try these! Dreambooth Playlist - th-cam.com/play/PLjC8P1vEncQD-QYBsYcyYq9JUAxQSZXnW.html Textual Inversion Playlist - th-cam.com/play/PLjC8P1vEncQDSDLKuPAEguajtdVKZpr9Y.html
QQ.. what is your exact video for a complete Linux local installation of Dreambooth Diffusers(shivamshrirao)? I was trying to follow the steps from multiple videos, but with no success :( There is always something that does not compile or such. Would it be feasible to run the Win WSL video commands in an Ubuntu VM(with GPU passthrough) ? What is the exact steps you follow for your best/fastest instance installed? TIA
A very interesting video, I'm starting with text2image, but I don't know much about programming, even so I have automatic1111 ui installed locally on rtx3070 8GB, it works fine for me. I'm testing dreambooth on runpod following a tutorial, but I don't see much documentation on how to use dreambooth with google colabs and get cpkt (convert) models; videos on this would be great. On the other hand, I trained a person on dreambooth and when trying the optimised model, comparing colab with runpod version, the runpod version (I guess it is with dreambooth unfrozen, the joepenna one, , I do not know) was much better when aplying prompts to generate portraits etc...In any case, your channel is a reference for me, thanks for your work...
Isn’t the Runpod version defaulted to 2020 learning steps where the Colabs one is 800? There might be a difference in floats vs half-floats or something too, but as long as the code runs, whether it’s run on Runpod or Colabs shouldn’t make a difference. I did see a video that declared the Runpod result “better” but the result was basically a direct reproduction of a training image, which is not the goal.
I noticed today too that the readme on the diffusers version talks about using around 500 class photos, but the notebook is set to 50. Also, I got myself into Colab time-out by increasing the training step, which generating 500 class images would likely do as well. Runpod might be the way to go if you don’t have a GPU capable of running it locally.
I just learned from your video how to run the Dreambooth Diffusers on my 3080-10gb with prior preservation using WSL2. Many thanks for your awesome video! Did some successful ones and tried them out on automatic's SD. It's cool! Today I f**ed up though: I tried using better representing class images. Instead of person, I tried "photo of a 20 year old woman" in the class prompt. I planned to just generate the class images and then re-run the launch script for a proper class name, like "woman" but I forgot to change it. Now 2 hours later my training is done. Is it possible to change the class name back to "woman" in the finished model? I don't want to retrain for 2 hours if possible.
@@NerdyRodent RuntimeError: CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 10.00 GiB total capacity; 6.24 GiB already allocated; 0 bytes free; 7.58 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF Steps: 0%| Only thing that's running is the terminal window
Either use Google colab or another paid service or check out their video on how to set it up in Windows using the Linux virtualization. It was linked at the end of the video (the left card).
Aitrepeneur has a video on hypenetworks. At the end of the video he gives his opinion that dreambooth gets you the same results with much less time and effort.
Best avatar yet
Agreed!!!
Most fitting the accent, indeed 🧐
Spiffing! 😉
I whole-heartedly agree with this sentiment.
First if all u rocks. Thank u!!
I did textual with 4 photos of myself 6 tokens 1500 steps... Unbelievably acuurate almost every time...
I think that something git wring with your owl...
Poor owl 😥
I can't make textual inversion work on 1111... I Always get extremely off results, 16 images and 8 tokens, 100000 steps
@@zonas7915 ikr! It can produce strange results sometimes 🙁 Often adding or changing training images can help
@@zonas7915 i dont think its the right way...
I believe U should use 3-5 pics. And save a pt file every 500 steps and check the quality of the picture. For me i did 16000 steps, 1500 steps is the best. More than 1500 its starting to degrade...
6 tokens for me, and i get good results and i can manipulate the picture really easyly...
i love your narration in this. the fight announcer style is so funny! its genius to think of doing it in that style :) ty my ro
Glad you enjoyed the things! 😉
Hmm…apparently a personal taste. I think the dramatic voices and uncanny valley avatar draw my attention away from the content and make it hard to follow. But, from the comments, looks like I’m maybe in the minority.
Appreciate the content! The tech is moving so fast it really helps to have comparisons like this to keep up with all the new techniques available. Thanks again!
This experiment is pure gold 🥇
I appreciate this. I don’t know enough about all of this and to have a comparison make’s my understanding better. Thank you nerdy rodent 😊💚
hold on to your papers that brings a bell XD
Think that's a fair assessment. There are some very clear differences between them. That said I think Automatic1111's Textual Inversion has potential, and I like that it has a bit of a modular aspect to it where you just "plug in" the embedding you want to use. The disadvantage there, I guess, is that each embedding pushes up the number of prompt words.
Not only that but you can actually run it on lower end hardware, if you want to do it locally it's the only option at the moment for a lot of users. It takes it's time though...
@@HugoM946 "slow" just means you need to set it up before you go to bed and let it keep the house warm all night. :D
@@TOEC Marvelous technology! Great success! **cries in RTX off
@@HugoM946 as I quickly move a cardboard box in front of the clear panel on my computer to hide my 1070ti
On the flipside with the 1111 Embeds you're not having to have a bunch of seperate 4+GB Model files for every inversion filling up your SSD/HDD. Plus when 1.5 publicly releases the embeds should presumably still work. Hopefully the code will improve over time.
Yup. The 2gb files are a lot bigger than those nice, small embeddings!
@@NerdyRodent As a matter of interest have you tested whether there was any differential betwween the full fat and low fat Models? I've shied away from the 2GB version as it talks about a quality drop? Also loving the content. :)
@@KadayiPolokov Not much of a difference that I can see!
Can you recommend good voice changer? Is there any ai voice changers or tts?
Tortoise TTS is good as cloning. Not sure sure about voice changers…
I'm like a Celebrity 😎. Now that I got the clout, have you done Ubuntu Dreambooth install? Is it any different from the Windows WSL install you did or is it the same?
Pretty much exactly the same! Of course, a new version of CUDA is out now 😉
Now what would be the big editability difference between the two dreambooths? They use thew same code - so maybe their initial settings are different?
The code is a fairly similar, but not exactly the same. The diffusers libraries is one fairly significant difference…
Thanks for this awesome comparison!
Can you maybe make a video on how to create a diffusers dreambooth model locally?
Interested in adding things to your AI Art? Try these!
Dreambooth Playlist - th-cam.com/play/PLjC8P1vEncQD-QYBsYcyYq9JUAxQSZXnW.html
Textual Inversion Playlist - th-cam.com/play/PLjC8P1vEncQDSDLKuPAEguajtdVKZpr9Y.html
Você aplicou AI Voice Changer_ em sua voz?
Great video!! Quick question, how do you create these amazing avatars?
Here's a video he did on it. Very cool tech. th-cam.com/video/zZTOsm6Wm2w/w-d-xo.html
You don't believe in watching the videos till the final end cards links, do you?!
😆
@@PawFromTheBroons hahah, I do now! Thank you :)
Doing God's work here
Does dream booth only do object faces or does it also do styles of art?
I’ve personally only done objects but I hear you can do styles
YES!!!
IA trainers fight, I liked it!
Finish him!
Go Hawkeyes!? 😉
I only looked at the avatar the whole video
Probably best 😆
What is the URL for the git of Diffusers Dreambooth Shivam ?
Links in the description. See the previous videos also for more info 😀
how to do this same avatar please?
Do you not have to reload SD after you switch between models?
Nope, they fixed that ages ago!
@@NerdyRodent Thats good to know. Things are changing lightning fast.
Another great video. Nerdy did you change your.. umm.. mushst... ha.. ey.. nos... umm... hairstyle??
thanks
Glad you liked the things!
What is this camera filter that made you into a silly 19th century notable?
The power of thing plate spline 😉 Animate your avatars -
th-cam.com/video/Z7TLukqckR0/w-d-xo.html
How can you see the style boxes? I can’t see it on my webui. Guess it’s just a Linux thing?
Thank you
You're welcome!
Do you know if Diffusion DB can work on Win now ?
Shivam’s will work (partly) on Windows. Linux still has the best DB experience
How do we install all of these different training models? What's the procedure to train each, so you have any videos on them?
Yup! Links in the description and Stable Diffusion playlist 😉
QQ.. what is your exact video for a complete Linux local installation of Dreambooth Diffusers(shivamshrirao)? I was trying to follow the steps from multiple videos, but with no success :( There is always something that does not compile or such. Would it be feasible to run the Win WSL video commands in an Ubuntu VM(with GPU passthrough) ? What is the exact steps you follow for your best/fastest instance installed? TIA
It’d be exactly the same for a normal Linux install, though obviously you’d not select WSL when installing cuda 😉
@@NerdyRodent Thanks!!!
Okay, the avatars starting to go to your brain
Srsly tho gonna have to dual boot again and chuck ubuntu back on my main rig to do some 3090ti training
silly, but refreshing :) like it.
cool
Ikr! ;)
A very interesting video, I'm starting with text2image, but I don't know much about programming, even so I have automatic1111 ui installed locally on rtx3070 8GB, it works fine for me. I'm testing dreambooth on runpod following a tutorial, but I don't see much documentation on how to use dreambooth with google colabs and get cpkt (convert) models; videos on this would be great. On the other hand, I trained a person on dreambooth and when trying the optimised model, comparing colab with runpod version, the runpod version (I guess it is with dreambooth unfrozen, the joepenna one, , I do not know) was much better when aplying prompts to generate portraits etc...In any case, your channel is a reference for me, thanks for your work...
It's super easy to convert diffusers models to ckpt format :) See th-cam.com/video/_e5ymV4zY3w/w-d-xo.html
Isn’t the Runpod version defaulted to 2020 learning steps where the Colabs one is 800? There might be a difference in floats vs half-floats or something too, but as long as the code runs, whether it’s run on Runpod or Colabs shouldn’t make a difference. I did see a video that declared the Runpod result “better” but the result was basically a direct reproduction of a training image, which is not the goal.
I noticed today too that the readme on the diffusers version talks about using around 500 class photos, but the notebook is set to 50. Also, I got myself into Colab time-out by increasing the training step, which generating 500 class images would likely do as well. Runpod might be the way to go if you don’t have a GPU capable of running it locally.
Thanks everyone, I have heard that joepenna is working in a Google collab version that could work fine, but I do not know if true...
@@rincondesalva We’ve already got it running in colab free so not really a need for it, but hey!
what about a diffusers tutorial for windows (dummy edition)
You mean like this? 😆 Add yourself to any AI Artwork for FREE!
th-cam.com/video/w6PTviOCYQY/w-d-xo.html
I just learned from your video how to run the Dreambooth Diffusers on my 3080-10gb with prior preservation using WSL2. Many thanks for your awesome video!
Did some successful ones and tried them out on automatic's SD. It's cool!
Today I f**ed up though: I tried using better representing class images. Instead of person, I tried "photo of a 20 year old woman" in the class prompt. I planned to just generate the class images and then re-run the launch script for a proper class name, like "woman" but I forgot to change it. Now 2 hours later my training is done.
Is it possible to change the class name back to "woman" in the finished model? I don't want to retrain for 2 hours if possible.
Unfortunately not as far as I’m aware ☹️
@@NerdyRodent Thanks! Gonna cancel and start training again. This time with the correct class :)
How the hell are you people utilizing the 10G 3080 with this? The OS always OOMs me on Windows 11, latest drivers, latest WSL2.
@@cyberskunkworks4117 Just run the training only and nothing else that uses VRAM.
@@NerdyRodent RuntimeError: CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 10.00 GiB total capacity; 6.24 GiB already allocated; 0 bytes free; 7.58 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Steps: 0%|
Only thing that's running is the terminal window
Wheover this character is, he needs to do more videos and exist.
What are you using for your avatar?
That’d be the thin plate spline motion model for animation 😉 You’ll need to watch to the end to find the hidden links…
bum, bum, bum.
I love your videos ! I've spent my last 3 nights playing with this :O pls stop I need to sleep ;) thx for your hard work
It’s addicting! 😆
link for dreambooth diffusers? so linux only, but is their a colab? thanks!
Down in the description and there is indeed a colab!
8:34 woah! So any way for mere mortals without a linux machine to get this level?
Either use Google colab or another paid service or check out their video on how to set it up in Windows using the Linux virtualization. It was linked at the end of the video (the left card).
Shivam’s Diffusers Dreambooth will also run on WSL2. You can also use one of Google’s Linux machines via Google Colab 😀
Ahh Ok, cool, I use google collab, didn't know it was Linux friendly. Thanks!
@@audiogus2651 Yup! Colab = Linux 😀
Nuuhddy Rhoad'nt
Please share your training settings for both dreambooth models
All just the default with the phrase given 😀
can you do one for hypernetworks?
Aitrepeneur has a video on hypenetworks. At the end of the video he gives his opinion that dreambooth gets you the same results with much less time and effort.
Fantastic analysis
Awesome vid. Thanks!
Glad you enjoyed the things! 😀