Stability AI's Stable Cascade How Does It run On My Lowly 8GB 3060Ti?

In this video, we're testing the performance of Stability AI's Stable Cascade on a lowly 8GB 3060Ti. Is it worth installing? Find out in this video! I'll go over how to run Stable Cascade on your PC, install it locally and easily using the Pinokio installer.
Try Stable Cascade here huggingface.co/spaces/multimo...
Details about the model stability.ai/news/introducing...
Github details github.com/Stability-AI/Stabl...
How to install Pinokio • Pinokio The Ultimate S...
⏲Time Stamps
0:00 Stable Cascade Preview
1:08 Some info on Stable Cascade
2:53 How to install Stable Cascade with Pinokio
4:29 Settings available
5:54 How does my 3060Ti 8GB VRAM GPU perform?
*Disclaimer Affiliate Links Below*
📸 Gear I use
Sony A7C CAN amzn.to/3spWX8C
Sony 35mm 1.8 CAN amzn.to/36Apekr
OBS (FREE) obsproject.com/
Editor Davinci Resolve www.blackmagicdesign.com/ca/p...
Audio: Rode Podmic amzn.to/35sSnxv
🔦 Find us on:
Discord: / discord
Instagram: / monzonmedia
Facebook: / monzonmedia

[@MonzonMedia]

Happy to report Stable Cascade through Comfyui can run decently at 24 seconds on my 3060Ti. Basic set up but so far so good. Way better than the gradio version in this video I tried a couple weeks back...5min! hahaha! Vid to come soon!

[@user-or4ks4bs5p]

wurst is sausage, wuerstchen is a small sausage, usually soft usually brown or white, so its a mishmash of meat from different parts of an animal

[@Stick3x]

Pinokio is the bomb and so is LLM Studio.

[@NotThatOlivia]

thanks for the video!!!

[@MonzonMedia]

You’re welcome! Thanks for watching! 😊

[@jdesanti76]

Using a 16GB 4060 TI, with default settings, it takes 14.4 seconds to generate an image, 37.8s with max resolution. The model is using all the VRAM, it would explain why it is so slow on GPU with less VRAM

[@MonzonMedia]

Yeah I figured so 👍🏼 official released model should be better. But definitely need to upgrade soon.

[@anupew3276]

can confirm its memory limited. Running on 12GB 3080ti it takes 4 minutes to generate an image. It takes all the VRAM and GPU runs at around 30%

[@rionix88]

Can it offload to pc ram?

[@gynthos6368]

Wurst means literally sausage and würstchen is basically cutification of the word used for small sausages

[@pn4960]

So it’s small sausage AI ? XD

[@MonzonMedia]

Thanks for confirming! I guess I'm trying to make the correlation as to why they would name the diffusion model after a sausage 😊 I suspect because of the small file size as it's 1gb and I know sausages are a German thing so.....

[@jacksparrowramirez3312]

it means it can run on a sausage computer ?

[@bugcrazy3345]

thanks for showme how to install and down,

[@MonzonMedia]

I'm sure there is a way to optimize it more to run it a bit quicker however, for now I rather wait for the optimized commercial model. I wonder if it can be compatible with the usual platforms since it's built on a different architecture?

[@MonzonMedia]

@JayZone80 Not as of yet but saw Olivio's video on it. I could be wrong but I don't think you can use Cascade on it since it's stable diffusion architecture. Cascade is different. Although I suspect something like ComfyUI might have something in the next few days possibly.

[@nickolaygr3371]

Tell me, is it possible to use lora sdxl with playground2? it doesn't seem to work correctly. or do i need to train a new lora on playground2?

[@Mehdi0montahw]

Stable Diffusion WebUI Forge is the version that runs everything great. It has fewer requirements and has additional features. I tried it yesterday

[@MonzonMedia]

Yeah been meaning to check it out. Will be doing a video on it soon to compare it against A1111 in terms of speed.

[@Mindereak]

What do you think about Pinokio vs Stability Matrix? Thanks for sharing this new AI tech

[@MonzonMedia]

I've only tried Stability Matrix quickly and from what I saw, it works quite well, especially having the ability to download models directly from Civitai is a nice feature. I would say if anyone is just interested in stable diffusion then it's a great way to manage all the SD platforms. Whereas Pinokio it's not just a SD installer. It can install audio, video ai apps too which is great! I'll be covering those soon!

[@Paperclown]

@@MonzonMedia oh that case subscribed. Up until today I never understood why anyone would NOT use Stability Matrix as people run into so many errors trying to manage manually. However I couldn't get Cascade to work with ComfyUI the extra tools is interesting I would rather stick to Matrix personally.

[@Mindereak]

Great, can't wait to check that out then, you'll probably make me get Pinokio aswell :) @@MonzonMedia

[@sashakovtun1996]

I have outofmemory on my rtx 4070 12gb. I need to change batch size, but I don't know where I can do that.

[@baheth3elmy16]

Thanks for the demo. My GPU is 3050 8GB. So I'm not going to try it now.

[@SotosH65]

I tried it last night, with my now "oldish" system with a RTX 2070 Super and a Ryzen 7 3700x, 32GB RAM, but it was going to take a long time, probably 30 min to an hour? I don't know, I stopped it at 4% and deleted it afterwards, it does not worth all that time in my system. Maybe later, when it is optimized...

[@MonzonMedia]

I don't blame you, as I said in the video, I'm not overly impressed and why bother at this point when it's not even the official model. I mean it's nice to see it's potential but it's just a headache to run right now.

[@SotosH65]

@@MonzonMedia Absolutely! Anyway, there are plenty of apps that work fast enough for older systems, such as Fooocus, the new Forge (installed that too...), etc! Thanks for your great videos, very helpful! Greetings from Greece / Crete! 😎

[@PaulFidika]

The model weights alone are 10GBs

[@MonzonMedia]

Yeah they are big and there is like 3 of them so as I said, best to wait for the official version. 👍

[@DinoFancellu]

Takes a few seconds on a 4090. Nice quality. Will be nicer when we can choose model etc.

[@MonzonMedia]

Must be nice! 😊 Although I feel I like I can optimize it more in the backend however not going to bother until the actual model is released and support of the usual platforms.

[@ZwaetschgeRaeuber]

why do i get oom error with my rtx3060 12gb vram and you can run it fine with 8gb?

[@MonzonMedia]

What's your system RAM?

[@ZwaetschgeRaeuber]

@@MonzonMedia 512gb its a proliant dl380 g9 server

[@oysterboulevard6623]

Nvidia gave Matt Wollfe a new GPU for free, they should give you one. You do same hard work as him. Just my humble opinion 🥴

[@MonzonMedia]

I would so love that! He's a much larger channel though, maybe one day if I can ever get to 100K subs? 😁 I appreciate the kind words and support though but for now I'll have to make things work with what I have. 👍

[@TheSynaptix]

my experience is like yours, even tho I have a 4090 laptop card (like a normal 4080), A 1024 x 1024 image took around 25 seconds whereas the same prompt with sdxl took 5 sec. I'm waiting for the next version too . But the images it tproduced were very good.

[@MonzonMedia]

How I envy those of you with a 4090 😁 Thanks for the feedback. I suspect the official version to be even better.

[@BrandonMitchell84]

it runs like balls...lol i installed it and was like damn its terrible with the 3060ti , takes forever

[@MonzonMedia]

😆Right? I do feel like there is a way to optimize it manually but honestly not worth running locally right now until we get the actual official model and platforms like ConfyUI, A1111 can run it decently. The Hugging face link in the description is good enough and doesn't seem to be so busy so far.

[@dhiegoosorio8399]

I had a BAD experience with Cascade. To make a single image it took me 1 clock hour. I have a 3060Ti and 16 gb of RAM, but I get an error "pipeline_stable_cascade_prior.py:424: FutureWarning: `callback` is deprecated and will be removed in ve rsion 1.0.0. Passing `callback` as an input argument to `__call__` is deprecated, consider using `callbac k_on_step_end`" and it is SLOW. Does anyone know what is happening? Thanks

[@dhiegoosorio8399]

Please switch to PEFT backend by installing PEFT: `pip install peft`. Is Message now appear

[@MonzonMedia]

It's likely a combination of lot VRAM on your GPU and 16GB of system RAM isn't much. On average Cascade is currently using 15GB of VRAM so it will likely offset the load to system RAM. Also checkpoint itself is more than 20GB. Honestly, at this point I would just wait for the optimized model. Not worth running locally right now for people like us.

[@dhiegoosorio8399]

@@MonzonMedia thank you very much.

[@aiskills673]

is there any way to run it on google colab?

[@MonzonMedia]

Not sure my man, I don’t use collab unfortunately

[@fedorp4713]

Took just over 3 hours on a GTX 1070 8GB... yeah.

[@MonzonMedia]

😯Wow I commend you on having patience! I would've quit after 5+ minutes! hahaha! Let's hope the release an optimized version soon!

[@fedorp4713]

@@MonzonMedia I left it overnight and ordered an RTX 4060ti 16GB. I wish there were neural chips available instead of GPUs but I suspect that will change soon.

[@MonzonMedia]

Yes actually there is a company (that eludes my mind at the moment) that is working on hardware specifically for AI that can be installed on a PCIe lane, just like you would with any other accessories like a sound card, SSD using PCIe etc. I believe though we're still about 2-3 years away from that point but who knows. Very promising!

[@fedorp4713]

@@MonzonMedia I've been playing around with the RTX 4060 TI 16GB and it's very fast compared to the 1070I mean, it's nice to have an image in under 10 seconds vs hours. I like the prompt response from Stable Cascade better than most models but it's still far far away from Dalle. I'm mostly interested in simpler images and 2d vector styles. It's hard to get a good image and I've been through many models and loras. Any ideas?

[@Embassy_of_Jupiter]

* cries in 980ti *

[@MonzonMedia]

I feel for ya! :)

[@DezorianGuy]

Generating an image with Stable Cascade takes minutes with my RTX 3060 12GB, while I get a 1024x1024 image in a few seconds with Stable Diffusion. Weird....

[@MonzonMedia]

It's a new model that isn't optimized nor ready for commercial GPU's in a sense. Typically Stability AI will do this to get it out for researchers and developers to tinker with it and get some feedback before releasing the official version. From what I'm hearing currently you need 15GB of VRAM for it to run decently.

[@DezorianGuy]

@@MonzonMedia It is just that you and other creators get generation times of 5 seconds. That's why I wondered.

[@MonzonMedia]

@@DezorianGuy bruh! I didn't say 5 seconds, I said 5minutes! hahaha! Actually it's 4.3 minutes, so I feel your pain.

[@inout3394]

You say.. 5 minute to generate 😅 Now I know why when I click generate.. I have error and doing nothing 😅 A installed to my windows laptop with i5 12th Intel Integra IrisXE with 16GB ram, maybe this is the problem, because in terminal write about some error. Maybe my laptop have not minimal requirements..

[@MonzonMedia]

Unfortunately this required a dedicated GPU with at least 16GB or RAM to run this efficiently, mine barely got by. Likely the official model will be more efficient. But I still think it may not run on your system. These days 8GB of VRAM and an Nvidia card is recommended.

[@hatuey6326]

with a 3060rtx 12 go still waiting for the image lol ! so i've uninstall it

[@MonzonMedia]

Interesting, you are the second person to mention that with a 12GB 3060. I'll admit this way isn't the best. I'm sure once it's officially released, there will be better support via the usual SD platforms. There is already a node on ComfyUi for this.

[@ZeroCool22]

Mmmm, so with a 1080 TI (11Gb) not chance to run it.

[@MonzonMedia]

hhhmmm you can try but I suspect it will be slow. I'd wait for the optimized model to be released and when the usual platforms add support for it. I thought I read somewhere that there is a node for ComfyUI but haven't had the chance to check.

[@0A01amir]

8GB 3060ti is still a beast and not cheap, come on dude i just got one don't under sell it :D , optimizations will come. Wurschen was a diffuse model with the same data but never took off, it was 1gb model that did the same but with less quality, now the new Stable Cascade is using that archtect to achieve something better

[@MonzonMedia]

hahaha! I hear ya man but honestly these days even with optimized models and adjusting things on certain platforms 8GB isn't much. For image generation it's fine but once you get into things like Controlnet, inpaint, outpaint, upscale etc you really start to see the need for at least 12GB VRAM and a higher clocked GPU. But it's good that Stability is looking at "efficiency". 👍 Thanks for the explanation on Wurschen, I know it is a diffusion model, I was just trying to correlate the name to the model....is it because it's a 1GB file size? Which is small for a diffusion model, so small sausage makes sense. It's very comical hahaha!

[@0A01amir]

@@MonzonMedia I know, when i got 3060ti two months ago i was happy that i can do complex SD1.5 workflows like they were nothing and that weren't possible on my 970 but soon i hit the wall with SDXL+ipadapter+controlnet etc... :D I guess the name has something to do with how tiny the latent is and since it's made by Germans they went for that name, might be a slang or insult in Germany, i reall don't know :)))) Keep up the great work and thank you for the vid :D

[@rionix88]

5 minutes is too long. I will stay with sdxl. Fooocus can get 2 pic under 1 minute in laptop rtx 3060 6gb

[@MonzonMedia]

Context my friend, this is stable cascade not SDXL. Because the size of the model and it’s not optimized yet it needs at least 15gb of vram to run decently.

[@FelipePeletti]

My results are being TERRIBLE for logo design. Dalle smokes this one.

[@MonzonMedia]

I'm reserving my thoughts until the official release. One thing about Dalle is that they have a larger token limit which makes prompt alignment (prompt adherence) much better. Still to early to say at the moment.