Frankly I haven't seen any real difference except for the visual style of the UI itself which is all in all similar, it doesn't really provide any new tools, nor new ways. So effectively it's just another skin on already existing and working stuff :/ Aaaand this is also a comment for the great gods of youtube to give the vid some traction, you do the good thing with NNs/AI explanation

I kind of like it; it seems fast and stable. I need to play with it more, but I like the name Vlad Diffusion.

@@StyxGamingProject Same, I can’t see a lot difference.

The UI is ugly.

Not enough differences/improvements to go through the hustle of reinstalling everything, moving models and loras and installing your extensions. I just hope A1111 gets some love again at one point.

Haha, yeah, you're right. I for sure take SD for granted now and expect close to perfection or I'll throw the image out.

yeah it's insane how normal and rote text to image is now, and how fast we got totally used to having it available. the world hasn't exploded yet, so maybe we're okay. it's nuts how fast a revolution can happen, then everyone can quickly adapt and even be bored by it.

What is even the value in art anymore, all the generations no matter how breathtaking they once were, have all become just an old image.

​@@quercus3290 Hopefully not in an Idiocracy way of things 😅

@@westingtyler1 The faster the war is won, the longer the time it takes for the population to feel the ripples of the new government.

I don’t know what cfg is but I noticed the lower the better

I'm a photographer that has mostly been prompt engineering for the most photoreal results I could get. I got it to the point that I didn't think it could get better, then on a whim I switched from CFG7 to CFG3, and the results were magnitudes better. Then I switched from Euler A to DPM++ 2S a Karras and it jumped even farther. But the biggest difference, for me, in photoreal images, was going from the standard recommended 7 down to 3.

I agree, finetuning stuff like that when you're an advanced user can make incredible changes. Glad you found something that works for your workflow and your images! And thanks to the OP here providing the equation. I learned something today! 🌟

Thank you mate, really useful coming from a math background. But I wonder how your equation works since cfg is not between 0 and 1, does a 7 cfg mean you're beyond the negative prompt ?

The calculus of diffusion.

It was already destroyed as the piece of shit it is. I wish Auto would DIAFIRL and we can move on to a real developer.

​@@generalawareness101 dude wtf, you wouldn't be seeing this video without auto1111. You should chill

​@@vintagegenious For real! So crazy how entitled people get over free shit :/

SDP increases generation speed by quite a lot compared to xformers. Optimized for Torch2. Quality of life stuff like restarting is super fast when no changes are made and more settings available than in a1111. It's basically a1111 that Vlad improved on.

It's posed as a question Bro.

@@pedro3000 well, the question implies a case to be made for that, by the video creator, but that case wasn't really made to much.

@@sebastiankamph Thanks for the quick reply! Exactly the clarification I was looking for. :) Keep up the great work!

​@@sebastiankamph what is SDP? how is xformers doing better in one not the other? what is Torch2 doing for other tools do we really need it? don't be sad about a1111 not updating, it's having a half year break mandate , let not push it! thanks 😅

>_>

Not to mention last time I looked at Dreambooth it had updated its env in such a way as to completely break SD. I stopped paying attention to it after that rudeness.

too much noise for nothing

Yeah, a lot is behind the scenes, mainly running and generating faster.

Clip skip can be put to main window in a1111 too btw

@@fixelheimer3726 yes it can be put in the quick setting at the top but not in the main block of settings for txt2img and img2img without using an extension to move it.

@@sebastiankamph Interesting subject! I would love to see a video where you compare the improvement in image generation speed and how it relates to the amount of VRAM😊

Your error probably was that you didn't download and load any ControlNet models :) Happened to me. ;D

Ahh....you know what.....I'll bite the bullet, install it and let you guys know what I think instead. Will reply in a couple of hours. Cheers.

@@Dave_AIYeah please let us know. I'm very interested in rendering animations as fast as possible, thanks for testing it out! :)

@@pedrogorilla483 -- Disregard this, I made an idiotic mistake -- So I've been using it for an hour, and I'm getting some interesting results. Installation is very simple if you've already got A1111 installed - all I really had to do is paste the file paths to my models in Settings>>System paths, and it recognised all of my models, loras, textual inversions etc.... Here's where it gets interesting. Rendering out a single image @ 1024x1024 takes about 1.9 seconds (average) with vlad, and about 8.2 seconds (average) using A1111. That's a staggering difference. Rendering out a batch of 8 images (at the same size) took 10 seconds (average) with vlad, and 1 minute (average) with A1111. If anything, it seems even quicker when it comes to batch rendering. This is all on a 10GB 3080. Here's where it gets really interesting. Because I'm rendering at double the resolution that the models were trained on (1024x1024) I'm getting the usual weirdness in A1111. Double heads, double characters etc. People that have used stable diffusion for a while will know exactly what I'm talking about. Using Vlad, I haven't got a single image with a two headed character, or double subjects. Not even once. That indicates that either something else is going on under the hood that I don't know about, or I have been extremely lucky. Considering that I've already rendered out a couple of hundred pictures, it's supremely doubtful that luck has anything to do with it. Havent't had the time to test control net and some of the other features yet, but so far this has been a brilliant upgrade.

@@Dave_AI wait whaaaat?!! 🤯

​@@juanjesusligero391 I wrote a longer post detailing my findings, but the short answer is it's about 4 times faster at image generation for me on a 3080. Went from 8 seconds per image to 2 seconds. That's in text2img. Strangely, it also seems to fix the problem with double characters or double heads when you render at a higher resolution than 512x512 or 768x768. I have no idea why this is the case, but I'm not complaining! Basically, it has been nothing but positive results for me so far.

I bet you have! 😊

Torch 2 is not consistent yet. All old prompts is broken as for now.

@@PhilippSeven ahh, ok. thanks.

What big advantages do Torch 2 bring?

@@JohnVanderbeck Faster render time, you can check their documentation for further reading. i didn't familiar with this stuff either. xd

Happy to hear it! Agreed, quick updates have a tendency to break things 😅

I think it was moved into process

Yes, this is a separate installation. Can use the same models though.

Yes

yeah, same here. lshqqytiger with DirectML is the only one that worked for me with AMD. i hope in the future AMD can have a little bit more of support. it was a pain in the ass to set it up

@@Ezekiel-ny1hk The setup in my case was quite smooth, of course thanks to the instructions from Sebastian's video. The biggest disadvantage is the lack of a training function, which affects me the most now that I'm interested in embeddings. Well - I have no choice but to train at Colab...

😅

Keep an eye on the git, I belive they're actively working on a DirectML fork for AMD

Hey hey! Happy to have you aboard. Vlad is a great start because it adds some great features from the start. However a1111 will have more resources and guides for you in general. So if you're new, might want to play around with a1111 and when you know your way around, check out others, like Vlad. And make sure to check out ControlNet! That will be your best friend. I got like 10 videos on ControlNet, so make sure to check those out.

I will, thank you very much!

1. Don't know 2. Faster renders due to torch2 and sdp mainly.

@@sebastiankamph Dang, question 1 was the important one lol. When you generate an image it actually renders 3 images. A prompt image, a negative prompt image and a combined image. Somewhere along the combination process, towards the end, it begins to over sharpen and over correct to the point where it breaks/distorts the image. This is reproducible on every a111 version that I am aware of. usually 30+ will make the issue much less noticeable, but when using embeddings and other networks it will still happen. If you go to eular a, set 20 steps, use a strong negative embedding, you will see that the prompt preview image is generally better constructed than the final combined output render.

It's more than just a skin for A1111; it is just that many of the differences are either "under the hood" or look like "skin" changes. The actual importance of those differences is debated. One could make an argument that the most important difference is that VD is still seeing updates, while the last A1111 update was in March.

If you make a shortcut to the webui file you can add it after the path

@@sebastiankamph thanks. i think i understand. playing with SD has gotten me into coding just enough to know how to really break things well :)

Speed.

Thank you! 🌟😘

Speed.

You can use the same theme as in a1111, np!

@@sebastiankamph Thank you, I just installed and found the new theme support in it. Thanks for the video!

The A1111 UI is vastly inferior what're you talkin about

You can make your own colour scheme in Automatic. It’s easy

Not really, just install them in different folders if you want to test it.

Agreed!

Main part is Torch2 and sdp auto-installed. You have to manually find that for a1111. So generations are way faster out of the box. And since a1111 stopped updating, who knows what's next.

im not getting an image at all

@@kk47shooter91 I wound up abandoning it and going back to the automatic 1111.

its fixed now, i guess it was and update

Sweet, good tip!

So I assume it would technically be enough to update to 2.0 (currenlty on 1.13) and get all (or most) of the benefits? Just stumbled across 'Vlad' on the StableDiffusion subreddit, but am hesitant to "abandon ship" cause I've grown accustomed to using Automatic. :D Does updating the Torch version change anything on render results?

@@avon_c6199 I've been trying out Vlad's fork and while there are some really nice things he's doing, I've come across a lot of bugs where A1111 was pretty smooth unless I tried to do something more than my videocard can support. There is a price for bleeding edge development I suppose. But yes, if you want to update A1111 to Torch 2 it is entirely possible. I ran it just fine without the memory optimization flag which I think is a little slower and uses more memory, but it will give you deterministic results where xformers does not. I don't think it changes render results from Torch 1 to 2, but I can't honestly say that I've tested it much.

Mainly quicker with Torch2 and sdp. And easier to start up as a beginner, auto-downloads 1.5 model and extensions.

Oh no, it's much faster. You can still install torch2 in a1111, but it requires a manual installation.

Please ask somebody else about noob questions so vlad doesn't have to waste his time.

i have a low eng gpu 1050Ti and with all these fuss about vlad i thought may be i'll get some speed increase and I am literally seeing ZERO improvement in speed. BTW , you can set the directory of the models , extensions , lora , embeddings in valds to your A1111 one and don't have to re download anything

And way faster generations. And auto-installed extensions and base model.

Sweet! Then yeah, you're probably all set with sdp.

Well the speed isn't the issue, it's the compatibility. Nvidia is much much easier to get running.

I haven't made any detailed tests myself but I saw a test somewhere on Reddit where the base sdp was faster than anything with xformers. sdp reached 42 it/s where xformers were between 20-36, depending on using lowvram/medvram or not. Tested on a rtx 4090.

People in my Discord said it was meh, so I haven't even looked at it yet. What do you think about it?

@@sebastiankamph 2 million unique users in 6 days. The ChatGPT record has been broken. )

Sadly I can't test with amd myself as I don't own one. So I can only read the creator's notes on if it works with amd or not. Got no great tips I'm afraid.

@@sebastiankamph Thank you for the reply Sebastion and keep up the fantastic videos.