The Easiest UI for Stable Diffusion XL - Fooocus
Vložit
- čas přidán 13. 08. 2023
- Download the brand new Fooocus UI for AI Art: github.com/lllyasviel/Fooocus
My vid on how to install Auto1111: • How to Install and Use...
The AI film production company I'm helping build, Storybook Studios: storybookstudios.ai
Get the A.I. writing tool that I use! Sudowrite gifts you 10.000 bonus words only with this link: www.sudowrite.com/?via=albert (Affiliate Link - this supports me!)
----------------------------------------------
Did you like this vid? Like & Subscribe to this Channel!
Follow me on Twitter: / albertbozesan
![[Vowel]물고기는 물에서 살아야 해🐟🤣Fish have to live in the water #funny](http://i.ytimg.com/vi/3G9MH063Cx4/mqdefault.jpg)
This is an exciting new way to use SDXL. Thanks a bunch for sharing!
Very good experience for those just wanting to jump in without having to "learn stuff". Thanks for sharing.
"Mac/Windows(AMD GPUs)
Coming soon ..."
Can't wait!
SDXL has me blown away, Im on the verge of dropping all the SD1.5 models from my PC because none of them even come close
Love seeing KISS applied on AI work. We need this for agents, there are still too many install issues with some amazing tools like ChatDev (like a farm of developers you set on building apps).
thanks albert for sharing this .... finally i can run sdxl without any error. and i love the prebuilt styles which is very time saving and useful..
You’re welcome! Yeah, it’s a great way to use SDXL.
EXCELLENT!
Many thanks!
I hope there's a colab for this
Thanks for the video, very interesting. I don 't understand image generation and writing scripts at all . But somewhere I looked that you can take a picture insert and render pictures on this topic and already choose from them what you need. Please tell me in the Focus program is it possible to do this ?
I’m not sure I understand. Could you be a little more precise in what you’re trying to achieve? The newest version of Fooocus has image input.
thank you for sharing new ai technology
You’re very welcome!
Excellent tutorial as always.
Thank you!
hi there - very cool video. i was just wondering what style / checkpoint you used for the picture shown at 3:38 - really diggin that style. keep up the good content :) cheers Simon
Thanks! I wanted to check for you but I’m afraid I don’t have that exact image anymore…
No worries :) thanks for getting back
Hey Albert - Loved the video, i've been using focus and wanted to know if there was a way to use other models like the posing ones or Img2img on FOOOCUS.
Would be really helpful if you could show us how
I don’t think many more features will arrive in Fooocus, at the risk of it just turning into Auto1111. And for that just watch my recent video on installing and using SD :)
@@albertbozesan thank you so much fore replying.
I did do that and I think the local URl is the same, because after getting Auto1111 I I wasn't getting the link on which it was being hosted :(
Do I need to uninstall Fooocus to use auto1111
@@SuperFireburger no, that shouldn’t be a problem if only one of them is open. If you want both open at the same time for some reason, you can change which port auto opens up in. Open webui-user.bat in Notepad, and after COMMANDLINE_ARGS= add “--port 1234” with whatever number you want. Then it won’t try to share the same url.
So that Auto1111 web-ui from a previous tutorial is using this new SDXL 1.0 (super confused with all the versions available) ?
Auto1111 can also be set up to use SDXL, but it’s pretty slow. Fooocus is built for SDXL from the get-go.
Could you please explain to me what "Custom Switch" next to custom steps is supposed to do?
Generate images by that with this by default model are copyright free ? Thank you very much for your work and implication
The law is still catching up on all this AI stuff, but it does appear in most countries that if you just prompt text and do nothing else, the result is in the public domain.
It would be helpful if the command prompt could inquire whether I have an existing model before initiating the download.
If you place it in the models folder it will recognize it and not download the new ones. Make sure it’s named the same, though.
What specs did you run this on and what was the inference time? It seems very slow for me even when generating 1 image at a low res, with speed on
I’m on an RTX 2070 Super. It takes me about 35s for a quality image. Be aware that those are timelapses in the video, very heavily sped up.
@@albertbozesan it takes about 15 MINUTES per image at my RTX 3070
Sounds like something is off, my 3070 Ti pulls about 40-45 secs for one with 2 loRA's added. Does it show your loading xformers ok? - Hovers about 2 IT's@@SergheiStarus
Found any solution? Got mine but it takes more than 15 mins. Mine 3060 12 gb
is there a way to skip the checkpoints from being downloaded at start up? since i already have both xl and refiner?
Yes, just place the checkpoints in your models folder and make sure they’re named the same as what Fooocus is trying to download :)
This works well but as usual anything under 16gb of VRAM will suffer the cost of waiting time.
Hopefully in the future updates, they'll work on a solution for low spec machine owners like myself who dwell in the space lower than 8GB vram. It does work but at the cost of not being able to do anything else while its conjuring pixels out of the latent space, as well as the high waiting time that comes as a result of having a low vram machine.
You've also earned a new subscriber because of the magic words "A.I. Film Company".
I’m usually on an RTX 2070 S with 8GB vram, and can use After Effects/Blender while generating. I can’t fine-tune on this card, but most other uses are fine. What are you using?
Regarding AI Film Company - check out my last video where I talk more about it :)
I suppose 8gb is just enough, I use it on a 1650 mobile with 4gb, to achieve an usable time I use 512x512
@@sherpya SDXL doesn’t work well at 512x512 because it was trained on bigger images. If you want that low res you should stick to 1.5.
@@albertbozesan it is true, but 1024 generation takes 5 minutes and does not always work, so I think I stick with unoptimal resolution, but I would say it still works
@@sherpya it works, but I strongly recommend using 1.5 for those resolutions. The results will be better and faster than forcing SDXL into suboptimal results.
Wow, incredible software. Is there a paid version for the commercial license? Or can you even use the images commercially?
Stable Diffusion is fully open source. Unfortunately the copyright laws are far behind AI art these days - you may risk the images you create here being public domain. But many people use SD commercially!
@@albertbozesan that is awesome 👌 thanks for the info and great vid.
Nice vid, I was hoping this had both in & outpainting, maybe in the future? We'll see.
I don’t think it will, because it’s specifically mimicking the features of Midjourney and should stay simple. If you want those features, check out my video for Auto1111 :)
Thanks, gonna try it out anyway, looks fun.
Interesting. I don't need it but I might give it a go if it can use custom models.
FYI, it doesn't reduce the prompt as such, you're still prompting the same, it simply does it in the background after you pick the style and so on.
Yeah, there’s a part in the vid where I mention that :) in effect, I’d say that’s reducing the prompt 🤷♂️
@@albertbozesan True, I guess it depends how you look at it. I was looking at it as the prompt is still there, so that isn't reduced. Yet it does reduce the amount of prompt you type in, so there's that point of view. 🙂
I'm mostly just glad that I can get people into SD now without teaching them Auto1111. They don't need to know what exactly is happening behind the scenes until they're comfortable, IMO :D
@@albertbozesan Yes I agree. I'm not a YT'er and I still get a lot of questions about, so I can imagine how many you get. I usually point them to one of the websites but it's great to have something for those with the hardware but who don't want the install hassles that come with it.
is there a web version so you dont have to install? Also is there a private generation feature? what are the max dimensions? can you use commercially?
It runs locally on your computer, so it’s all private anyways. I’m not a lawyer, but barring the current copyright discussions around AI art, the results can be used how you like.
@@albertbozesan thank you. It's too bad that some of these generators force you to install rather than have some web presence
@@coldgin2023 well, I see it as a big advantage to use open source software on my own hardware for free. You can always run this in a Google Colab or use one of the many paid online services for SD. That’s not what Fooocus is for.
@@albertbozesan to have a program use up resources when you may need them for other things is a disadvantage to a lot of people. That's why an online presence is very important as it frees up your computer and opens up the generator to a lot more people.
@@coldgin2023 coding a UI and providing an online product are two very different things. Like I said, there are several ways to use SD online :) I won't name any for-profit products, but google Stable Diffusion Colab and you'll find an option to run it yourself online.
Did not realize 4gb vram was listed as the minimum, didn't expect to be able to actually run SDXL but I'll have to try it
Let me know how it goes! Hope it works well :)
I'm surprised that it is less then 16gb ram and a lot of CPU ram. In fact the low resource usage is extreme encouraging.
@@albertbozesan well for me i only got a rtx 2050 4 gb vram,and it works pretty well(takes 2 mins in average which isnt much of a issue for me).
Does anyone know why input image doesn't do anything for me, and how can i fix that? Did anyone have a similar problem?
Thank you very much for the tutorial. I tried this with my laptop . It's working perfectly. CPU : AMD Ryzen 7 4800HS , GPU : NVIDIA GTX 1660 Ti GDDR6 6GB
Glad to hear that!
It's fun, easy and pretty fast on 8gb gpu with sdxl.
This looks like a very nice tool for beginners to kick the tires on AI images without the 'hassle' of using Discord and other online / paid tools. Or the relative complexity of A1111 / ComfyUI, etc. With SDXL especially, this should allow for some nice creativity from a broader community.
Yes!! I’ve found myself using it too for quick and dirty ideas. Much faster to get going than the other options.
why is it sooo sloooow. i do not use a SSD harddrive. is that why? i do use 2080 ti RTX
@@andrer4508 what is „slow“ in your case? Anything up to 2 mins per image is probably okay.
NIce
When i go to the page it says ''this page can't be reached''
can i know how to check the seed generated in this UI? thanks
I’m not sure - and someone seems to have opened a GitHub issue about it: github.com/lllyasviel/Fooocus/issues/44
An easy seed/info ref on generated images would be nice, yet the seed should be in the image metadata so look there. I use A1111 and ComfyUI, A1111 has a nice extension for easily seeing the generation info.
Anyone tried this with an AMD GPU? Thanks for your videos Albert. We live in exciting times!
GitHub says “coming soon”, keep an eye out there :) Glad you enjoy my videos!
The cmd says this for me: The system cannot find the path specified. Can someone help me
How about pixel art? Anything to create 32x32, 64x64 ect?
I’d generate it full size with a pixel art style/prompt, then resize it down in a photo editing program with some other tricks. Trying to generate natively at low res won’t work.
style is availble in Auto111 too
Yeah, but it’s not as “hands-free” as here. Fooocus doesn’t scare beginners.
yes that is true. It is cleaner@@albertbozesan
Hello ....it's taking too much time to load one image , on 4gb vram is there any solution ?
4gb vram is unfortunately not enough.
Hey, I installed the software and opened the window in my browser, but I get the following error when I try to create an image:
safetensors_rust.SafetensorError: Error while deserializing header: MetadataIncompleteBuffer
I would love it if someone could help me.
This is a better question for the software’s GitHub page. Have you checked there if someone else has experienced it?
sdxl can run now on 4gb vram?
It’s probably slow but yeah, looks like it!
How do you get this bloody thing to train and use your own models and datasets for image generation???
I won’t claim it’s super easy. First of all, Fooocus is not the right UI for it. It’s too simple. I recommend checking out Kohya for LORA training, and be prepared to spend a lot of time with trial and error. You’ll need a good amount of VRAM in your GPU, too.
Hey, thanks for the video. I downloaded and installed focus. However, when it is opened on my chrome browser, it says "your connection to this site is not secure". Still I prompt and press the generate button but it shows "initializing" and nothing comes for ages. I have 16 gb ram and 7th gen cpu . can you help?
Ideally you can ask on Reddit for trouble like this, the community is faster than I could ever be.
having some problems trying to download it on macbook, any help?
I wouldn’t recommend using this on Mac. But if you want to, check for your specific problem on Reddit. That’s the fastest way to get support.
I like the UI but it is significantly slower than A1111 - 4090 owner. It's almost ready, but needs to be faster.
Interesting. I just installed my 4090 and will have to test it out!
it comes with controlnet?
No, this is purposely a very simple UI. For ControlNet, check out my recent video “How to install Stable Diffusion”, link in the description.
Can you make video for making 2d character portraits and unique monsters. Monster like fear and hunger 2 style. It’ll help me so much thanks
Any news on Controlnet?
I don’t think that will ever be a feature of Fooocus. It’s meant to be the simplest possible UI - they’re only working on features that Midjourney also has.
Can you explain how to uninstall it also please, take a lot of space on my pc.
Just delete the folder.
what if you don't have the Navida GPU?
The GitHub page says “coming soon”, keep an eye out there.
Thanks man!@@albertbozesan
Is there a way for me to host this as a web server so my friends can access it without a vpn?
There is a flag you can set in the config that lets you open it to your local network. Check out the Fooocus GitHub for more on that.
You can then open a port to outsiders, but that’s classic networking and beyond my core expertise 😅
it's good but there is no option for controlnet and extension
No, that’s the point. If it did it would be auto1111, which already exists :)
@@albertbozesan ya bro
😁
But we colab user cannot install a1111😑
Can we use another models??
Yes, I show that near the end of the vid.
I Have Rtx3050, 4Gb of vRam. This method is processing the prompt very slowly... I am using SDXL in confyui with no Refiner! It takes about 45 seconds to generate an image! Also SDXL is better for arts only! But for photos, human generation SD is much better and faster!
45 seconds is okay for SDXL on a card like that. I don’t agree that SD 1.5 is always better for photos - depends on what model you’re using, I suppose.
@@albertbozesan Of course, it depends on the model, and it takes only 5-7 seconds for a generation. I am using "DreamShaper" model and after transforming the same photo in img2img via "Realistic Vision", the result is fantastic!
Whats about Mac?
At the moment you need a powerful NVIDIA GPU to get any worthwhile performance. Auto1111 works on Mac, but slowly.
Why it is so incredibly slow on my RTX 3070? It takes more that eternity to take the picture? I didn't count but it is tens of minutes
It really shouldn’t be - it takes up to a minute on my RTX 2070s. Do you have any other GPU intensive software open at the same time? Have you checked the issues on the GitHub page if someone else is experiencing it?
@@albertbozesan I checked the issues but didn't find anything like that
@@SergheiStarus Did you fixed it ? because I just now I installed but it takes 3 to 5 minutes per image. I have 3060 12gb & 8gb system ram.
@@CloudVideoYT nope, just postponed it
@@SergheiStarus thanks, it you fix it in feature let me know because this UI is simple & easy to use.
At first everything is fine but when I hit generate it takes forever to create an image, the blue bar almost does not advance, I have a ryzen and a 6gb nvidia gtx 1060 video card 😭
Hm. I’m afraid that card is probably too weak for SDXL, sorry.
@@albertbozesan finally I tried other preset and it worked much better, around 1 minute or less, the only bad thing is that while im doing the process I cant use the computer for other things due to the memory consumption. But its worth it since it take very good images
i hate i cant add embeddings and hypernetworks, or vae
That’s what the other UIs are for. Adding too many features would go against the main goal of Fooocus.
@@albertbozesan but they have maps for it
My computer freezes by just opening the app, despite meeting the 8g vram and 16g ram requirements
Can you check on the GitHub page if others have had your issue?
And.....
Still using Comfui with SDXL and Face Detailer
As am I. This is just much easier way to get into it.
Whatever works best for you. ComfyUI needs some quality of life updates before I jump to it but it definitely wins in terms of speed and resource usage. Until then I'm sticking with A1111, even for SDXL.
It just simply won't launch for me. lol
Did you check your specs?
Be warned this takes absolute ages to unzip im only on 30% unzipped and its been 40 minutes, rushes to 29% then its slow going from there... will update when it's done, or if i just quit (if it goes on too long)
That definitely isn’t the case normally.
It's impossible to use! It consumes all my RAM instead of my video card! It uses my 16GB of RAM just by opening the page.
Did you check the necessary specs before installing?
@@albertbozesan Yes! My PC is much better than what is recommended there!
The problem is not with my PC! The problem is that Fooocus consumes 16g of RAM without doing anything, just opening the page. I've been using SD XL and automatic 1111 for a long time without any problems. RTX 3050 8gb - 16gb ram
@@fnsantucci ok. Then I don’t know what’s causing it. Have you checked the issues on GitHub?
@@albertbozesan Yes! I saw someone with the same problem and there was no response! I created a discussion there and so far nothing! It's all very strange because they told me that it would be lighter than the SDXL on another platform and the Automatic 1111, both of which I already use without these problems. My video card remains unused, when in fact it should pull the program from it. Thank you in any way for responding!
this video is about the old version 1
I released it soon after Fooocus came out.
Easy use is definitely NOT what i'm waiting for or is anywhere in the feature list. I think the usage is way to easy already. I need much more options in a more reasonable way to specify then Midjourney and the Discord interface.
Try ComfyUI! I don’t have a tut on that yet but it’s what I use most of the time.