Stable Diffusion 3 - How to use it today! Easy Guide for ComfyUI

How to run Stable Diffusion 3. This video shows you to use SD3 in ComfyUI. I also do a Stable Diffusion 3 comparison to Midjourney and SDXL
#### Links from the Video ####
platform.stability.ai/
github.com/ZHO-ZHO-ZHO/ComfyU...
#### Join and Support me ####
Buy me a Coffee: www.buymeacoffee.com/oliviotu...
Join my Facebook Group: / theairevolution
Joint my Discord Group: / discord
AI Newsletter: oliviotutorials.podia.com/new...
Support me on Patreon: / sarikas
00:00 Intro
00:19 SD 3 vs MJ vs SDXL
12:00 Stable Diffusion 3 in Comfyui

[@gicknardner]

* not local, not free

[@datcrazytwitch3677]

Agreed

[@quantumangel]

Sad :(

[@AscendantStoic]

Not yet sadly, but they promised it will be released soon.

[@RakibHassanAntu]

End of the stable diffusion era 😊

[@gicknardner]

Just want to add that I am referring to "use it today," nothing more.

[@sneedtube]

The API and the credits systems should have been stated at the very beginning of the video and a mention of them in the title was needed as well, even only a vague one.

[@GravityBites1]

agree, title and beginning of video (SD 3 is here... etc) totally misled me, even passed rumor of SD3 full availability on to friends after just brwosing through the video... you do great content, please don't be so clickbaity... no need for that

[@generichuman_]

Why is it always people with no content on their pages making demands... Why don't you contribute something instead just being a sponge?

[@AI_EmeraldApple]

Yeah I agree it should be stated. But to be fair a lot of other youtubers do the same to click bait and get revenue that way.

[@sneedtube]

@@AI_EmeraldApple not an excuse, still unfair

[@bigglyguy8429]

Yeah, this guy seems a bit, describes comy as a gui... it's totally silly

[@vi6ddarkking]

I honestly can't wait for the Full Open Source release. And for the Fine Tunes and Loras to start flooding in.

[@AlistairKarim]

Thanks for comparison! To state the obvious, the real game begins when we would be able to drop ipadapters, controlnets, loras on top of the model. However, prompt following is already a promising hint.

[@OlivioSarikas]

yes, totally. this will get really good when the community get's to train the models more

[@Lost_In_Entropy]

@@OlivioSarikas IF they do. Stability are putting out confusing vague references to 'self hosting' and have not explicitly said its just releasing the models as safetensors files. It keeps referring to membership, which while free would still indicate some sort of non local element.

[@OlivioSarikas]

@@Lost_In_Entropy Emad said on twitter "Stable Diffusion 3 will have an open release." - he also said that they will have a trust model for people to actually buy a license when they use it commercially. so that sounds very much like local use. otherwise why would they need a trust model?

[@Lost_In_Entropy]

@@OlivioSarikas Emad did say that but this announcement has odd language used. Memberships are free, so you can 'self host' it with a membership. OK but how do they know your a member unless its still going via an API somewhere. I am skeptical its getting a full release, as in safetensors hosted on huggingface. But hopefully I'm wrong.

[@BryceKant]

Nope. If I can't download and run it locally, with no accounts required, then not interested. Not doing this credit system game.

[@miguelgargallo]

true

[@PFnove]

this really should be in the title it's a closed beta and you pay to test it (on the cloud to avoid the model being leaked) when they finish it then you'll be able to download it for free and run it locally

[@michaelbarker6460]

Stability ai spent $99 million on compute power to get the base models for us to use. How have they made that money back. Well they haven't as they are currently on the brink of bankruptcy. Its wild that just a year or two ago we we're blown away by AI image generators and thought something like this wouldn't be possible for at least a decade or more. Now we want continual improvements and we want it all for free. Weird times.

[@CCoburn3]

And 6 fingers on the hand of the SD3 image at 1:48.

[@peacetoall1858]

I noticed that too

[@CCoburn3]

@@peacetoall1858 They STILL can't get hands right.

[@spartanx169x]

Until it can get the hands right it’s kinda pointless. It’s STILL not intelligent. And it won’t be.

[@heiferTV]

See the context?. It Must be a species of the Galactic Republic similar to humans , but with 6 fingers in each hand.XD or a mutation of human species, in the future. Hands are being, the hardest thing to fix of generative A.I. For classic art students, it's the hardest thing to learn to draw or paint properly, in the human figure.

[@PFnove]

@@CCoburn3 you can't even draw at all, meanwhile a couple billion numbers being multiplied in patterns can generate extremely high quality images in seconds it's normal for it to make mistakes since it's impressive for it to work at all and sd3 is still in closed beta so it's not representative of the final quality

[@andrewqUA]

Is this the end of the era of free Stable Diffusion?

[@OlivioSarikas]

no, it's just still in development, so they didn't release the model yet

[@AscendantStoic]

Nope, it's coming, still in training.

[@TR-707]

its over😊

[@d96002]

@@OlivioSarikas so you need to pay for a model that hasn't been completed yet, but when it will be completed it will be free ? That's weird

[@Elwaves2925]

@@d96002 You don't 'need' to, it's a choice depending on how much you want to try this right now, or wait for the full release version which will be free.

[@ImAlecPonce]

thanks :) I'm trying it out now

[@chris]

Fingers... still so bad.

[@nikolesfrances1532]

Exciting but not there yet need the open source to have fun, but it's a great party trick at the moment

[@skycladsquirrel]

Great video Olivio. The API key thing is a no go for me. I wonder what Emad thinks about this. The images I've seen don't look much better than Cascade. I'm really excited to see how it does with video though.

[@AGvfx]

It's wonderful! Will there be a local version available for training? If so, when?

[@FerikkusuOCa]

I wanna know it too

[@OlivioSarikas]

yes, in a couple of weeks they will release the model

[@Gmlt3000]

If SD is behind a paywall, it will be all over for the community (.Thumbs up for local installation, free for personal use...

[@Resqun]

yess! I was waiting for this video!

[@claudeclaude9991]

Hi Olivio, great video. Can you record a video showing how to train a Latent Diffusion Model? It would be helpful.

[@zhongxianshi3375]

Recently, there are several new models, such as Playground V2.5, SD Cascade, and PixArt Sigma, which I think provide huge improvements in terms of prompts adherence and image quality as compared to SDXL. Since SD 3 is not accessible by public any time soon, it would be great to hear your opinions on these new models (a head-to-head comparison would be awesome).

[@FightClubGarmz]

I feel that mid journey nailed the images. If SD3 is running on credits, and charges you per image, this will create frustration with bad image generations. Think about testing prompts, seeds, LORA, ip adaptors, it’s going to cost and feel annoying when it makes mistakes if you have to pay for each roll.

[@BakaTimeproject]

it is the presentation, it is true that it was said that it is going to charge for the use, but perhaps it is to manage it online, there are many pages profiting with stablediusion, it is not a bad idea to do it from the same project and promote it, at this time you can install the node that allows you to use Stablediffusion3 on your pc, but you do not occupy own resources is only an interface, I am running Stablediffusion3 in a: ryzen7 without GPU. it allows you to create 30 images, it focuses more on realism, the common hand errors remain, the new text as image generator does its job, but it is no different than adding text with an image editor.

[@HolidayAtHome]

a comparison with also SDXL base would have been nice. Now it's a bit like comparing raw ingredients (SD3) with a fully cooked dish ( finetuned SDXL ). But still interesting video !

[@OlivioSarikas]

but who using SDXL base to make image? If nobody is using it, why does it matter what it looks like?

[@FrankWildOfficial]

@@OlivioSarikasI would also like to see comparison with base SDXL, because If I compare base SDXL with finetuned SDXL, then I can roughly imagine how much better SD3 can get in the future

[@dataminearpg]

I wasn't too impressed with the output of SD3 just yet. As many already stated, the power will come with control and added community training. But I hope prompt-following will be groundbreaking with this one.

[@MrSongib]

the prompt "Style" is different for all the three model, IMO you can try test the prompt to get similar images for each model, instead of just copy paste the prompt and hopping to get same images.

[@OlivioSarikas]

Yes, but the point of this is that Stability said it needs less style words and is closer to midjourney. so the model should be able to understand what i mean without adding "raw photo, professional photo, 8k, highly detailed" and such

[@MrSongib]

@@OlivioSarikas True. I think we need prompt testing content. It's kinda interesting but can be boring as well, since no one like making a prompt. xd

[@KDawg5000]

Looks like they got text working a lot better, which is great. Can't wait till they let us download the model and run locally.

[@Make_a_Splash]

Any idea about the system requirements to use it, once is deployed open source?

[@RakibHassanAntu]

Wow!!! this is the end of stable diffusion era, thank you 😊

[@1lllllllll1]

@oliver I hope you have a chance to dive a little deeper into the tech. Rectified Flow being the innovation of SD3, according to TwoMinutePapers, deserves some of your well thought out commentary and explanation. I’m a huge fan of your efforts. And I can’t wait for the true free local install guide.

[@udippatel6147]

If this money-grabbing version is just the normal now for the company ... I'm glad I got SD set up during the v1.4 days and learned the complex capabilities for free. Thanks to Olivo and other channels.

[@OlivioSarikas]

Thank you. It's a bit strange to call this money-grabbing though. Do the people who create these models for us not deserve to be payed for their work? this is their job. this is how they pay for their childrens home, food and education

[@udippatel6147]

@@OlivioSarikas I guess they spoiled me with the free stuff, I don't think I'll ever pay for a model. My wording might be a little harsh, But even the previous ceo left because he saw that the company's vision had shifted more towards profit over the product.

[@OlivioSarikas]

@@udippatel6147 from what i understand he left becaue he wants model to be more decentralized. but that doesn't mean they are free. it only means that there is a landscape of many people developing models and using them as they wish, be that free or paid. But keep in mind that training these models costs millions of dollars and is very different from refining them with some community trained content. somebody has to pay for that and these people want to get their money back some day

[@generichuman_]

@@udippatel6147 Do you work for free? If not, are you a money grabber because you need to eat to live? lol...

[@andrewostrovsky4804]

They have removed artists from the training data, censored the model and now they charge money for its use? No thanks. They can shove it.

[@Vestu]

Gun poses are still almost impossible for AI art

[@OlivioSarikas]

not in Midjourney ;)

[@AscendantStoic]

Dall-E 3 on Bing handles it like it's nothing, very easy to get great gun poses (and it's free) .. hopefully SD3 will reach the same level.

[@97BuckeyeGuy]

@@OlivioSarikas We don't care about paid access models. Do you not understand this, yet?

[@sv122]

He says it's impossible, Olivio mentions one that can do it. What's wrong with that? Maybe get a job so you have some money to spend ;)

[@AscendantStoic]

Easily done in Dall-E 3, go try it.

[@stephanreiken9912]

They know that a third person who gets their offline model but never agreed to the terms and conditions has no obligation to follow the terms and conditions, right?

[@andrewqUA]

They probably won't give an offline model

[@OlivioSarikas]

well, that would be stealing (yes, you can steal free things too, because even they have licencing conditions)... anyways, they will release the model

[@PFnove]

@@andrewqUA they will release an offline model, there will be some rules for using it but ain't no one following them

[@namekyoyo]

Image to Image works in Stable Diffusion 3

[@mistraelify]

Until the diffusers are free for use and can be downloaded locally, i'll stick with SD 1.5 and SDXL.

[@krisrattus8707]

Love your work brother!

[@1lllllllll1]

@oliver I get soooo confused. Is SD3 simply a model checkpoint like all the other ones? It sounds like the actual underlying technology (diffusion calculations, adaptations, etc) would be what’s called “stable diffusion”. I’d love to hear commentary on this in the future, just to help clarify if we’re dealing with foundational technology updates or with incremental checkpoint upgrades.

[@OlivioSarikas]

it's officially called Stable Diffusion 3. Technically it's more like a SDXL update with new functions like very good text understanding

[@veteranxt4481]

Run on your pc but pay to run it?

[@andrewqUA]

No. you are not downloading the model. it's just a node that connects to the server. Free 3 images. Problem with hands again

[@quantumangel]

It's an api. That means you don't get to run it locally. They haven't published the model yet.

[@user-en4ts9mq4t]

​@@quantumangelwhat is api. Bro?

[@quantumangel]

@@user-en4ts9mq4t "application programming interface". It refers to the the method of communication between the client (in this case comfyui in your pc) and the server (some powerful machine with the model in a data center).

[@user-en4ts9mq4t]

@@quantumangel really thx . Bro

[@maddydon336]

Just for clarification, for sd3, you can't generate images using comfyui without buying api credits?

[@OlivioSarikas]

not at the moment, because the haven't released the model yet

[@sneedtube]

Custom SDXL, Vanilla SD3, that's not how you are supposed to compare models, this review is 🐶💩

[@jamesrcrypto]

Is paid the permanent model or an early access thing?

[@user-on6uf6om7s]

It's permanent. They will apparently release the model to be used locally but will be using the honor system to keep people from using images they create for commercial purposes if they aren't subscribers.

[@jamesrcrypto]

@@user-on6uf6om7s Thanks

[@FerikkusuOCa]

​@@user-on6uf6om7sWhat is a honor system though?

[@OlivioSarikas]

it's a early access thing. emad said the released model will come in a couple of weeks, from what i understand

[@markosz22]

Looking at the price of SD3 vs SDXL, it's probably going to be very demanding to run on local machine.

[@maxleo7556]

there will be an open version of SD3!

[@yanus_ai]

Hi there is there a way of booking a call with you for consultation?

[@DanielPartzsch]

Hopefully they will also fix hands and limbs and also it will be possible to create more interesting and authentic character photography. Not just always portraits of people starring at the camera.

[@farvardinmainyu1961]

I feel it's similar to the Stable Cascade...

[@heitorb2460]

Is the content moderation so strict because it is not on a local model? “Woman in a bikini” fails moderation rules.

[@genkidama7385]

you're comparing oranges and apples, all models are different. ridiculous.

[@GlenBland]

I don't understand why you have to pay hugging face per image to use a model that you run locally on your own machine. I really hope this does not become a normal thing. I won't be doing it.

[@valerymoyseenko]

three free pictures per api key? that's a record!

[@BakaTimeproject]

generates 30 images in normal mode, and you can always use your twin's account. Multi-pass!

[@ObscurelyTitled]

to clarify, is SD not free and local anymore ???????????????

[@BakaTimeproject]

is the presentation, and the installation is only an API, the launch is supposed to be delayed by the elections in the USA, then it will be launched, also why they emphasize security when generating image, in short it is not yet available to use it locally, it is true that it was said that it will be charged for use, but perhaps it is to manage it online, there are many pages profiting from stablediusion, it is not a bad idea to do it from the same project and promote it.

[@semisixx4967]

Model weights will be released in a couple of weeks.

[@eddiej.l.christian6754]

Charging for this version is a huge mistake on their part.

[@cristiansolano1052]

It's not them...this video generates a lot of missinformation...for the sake of views, but clicbait is a thing so...... The thing is that it's not the official SD3. that one will be released to the public for free as the previous version....

[@Mythhammer]

This wild...This is version three and they still can't get hands right. :)

[@fintech1378]

arent they launching the video diffusion?

[@ImAlecPonce]

there is a glitch haha, I made 5 16:9 images and it took just over 20 credits each and now it's not doing anything..... maybe the server is down. I don't know.

[@jamqdlaty]

I don't get something here. Why even read the prompt of the poodle dog if you're going to completely ignore if they followed it and just say everything is "very nice"?

[@godorox]

It works in a few images, now this error occurs: "Failed to fetch image: 402". Has anyone had this error? (I'm using Zho's workflow) Edit: Credit -14 =(

[@JustFor-dq5wc]

I've got mixed feelings about this. SD3 probably gonna need a lot more compute power and memory then SDXL. If it's gonna be cloud based only - ...

[@streamdungeon5166]

First thing to notice is what I call stable diffusion fingers. They just cannot figure it out completely.

[@quantumangel]

Can I haz local txt2vid plz?

[@MagnaKay]

Yeah, I thought they released the weights while I was sleeping but nah. Kinda misleading there. Fair clickbait, but a bit disappointing.

[@dascodraws6040]

This is the most asinine and scuffed way to drop a model. Now i believe the company is in trouble. Api key to run? yeaH, THANKS BUT NO THANKS

[@martin-cheers]

They promised the models in a couple of weeks from now, so I don’t know how bad it is.

[@Jeremy-sk9qh]

is it not for automatic 1111 yet?

[@HamedEmine]

It doesn't work offline, and you have to pay to use it. Unlike models who work in Automatic1111

[@semisixx4967]

Model weights should be out in a couple of weeks.

[@sneedtube]

"Hard-coded" aesthetics are a downside imho, I want a vanilla backbone to create MY OWN stuff with MY OWN rules, forced aesthetics are for casuals (that should stick with Midjourney for that purpose). And no, fiddling with the negative prompt and whatnot is NOT an acceptable trade-off if you ask me.

[@TheDocPixel]

Speaking of "hard-coded", there's rumours that censorship i.e negative embeds, are hard-coded and baked in, so that the community won't be able to fine-tune around them either with checkpoints, LoRAs or prompts. Even IMGtoIMG will be censored, since without a prompt, SD doesn't know what it's looking at or supposed to generate.

[@skycladsquirrel]

Do you have a tent in your pants! lol

[@openroomxyz]

Well if they go OpenAI way than better to start searching for alternatives, yea

[@jinouga4193]

I paid $10 for SD 3 test. Honestly it's not crazy at all. The model is better for landscapes and also for including text, but otherwise it is 0. We also always have bad hands. Dynamic poses are poorly managed... I am very disappointed personally. I would prefer that it train SD 3 for several more months than to release it in a few weeks.

[@MateoUnreal]

What happend with the rock metal intros??

[@user-ct8my8rv9c]

Stability AI is basically broke and burned all its investment cash away. I hope version 3 isn't censored, behind a paywall and API credits. But they need money. I'll believe the local release when I see it.

[@martin-cheers]

I think they said commercial use will come with a fee.

[@semisixx4967]

They say they absolutely will release SD3 for free in 2 weeks or so.

[@azmodel]

The hands... I only needed them to fix the hands...

[@sneedtube]

Also for a fair comparison you should have used the vanilla SDXL and not a custom model, because SD3 will get better trained as well in the near future you know

[@FusionDeveloper]

Too bad. I thought maybe the models were released.

[@OlivioSarikas]

in a couple of weaks, based on what emad said

[@openroomxyz]

@@OlivioSarikas thru their API or weights on hugging face ?

[@semisixx4967]

@@openroomxyz Weights

[@royandescartes]

you gotta pay for this ?

[@CecilMerrell]

This is just turning comfui into a rest api client, in no way shape or form in sd3 running locally. Also, if i wanted midjourney art, id use midjourney lol

[@omarnaser8291]

its not free

[@Nojlink]

slowly all the free stuff is coming to an end, I thought Stable Diffusion would remain loyal, but unfortunately that too is coming to an end, too bad 😒

[@miguelgargallo]

not local, no party

[@peacetoall1858]

So far it's not very impressive to me. Hopefully the final version will be way better.

[@luisbriceno5242]

I like how it looks SD3, thank you this video.

[@omarnaser8291]

your the best other videos show you credit base way

[@AI-Efast]

If you have to generate hundreds images per day, the fee is very expensive

[@Elwaves2925]

This is only for testing purposes, not the type of use you're referring to. Soon it will be free like the other Stability AI models.

[@sneedtube]

"Beautiful" is not a metric to evaluate a model result because it's subjective and not deterministic, but you heavily rely on that. Too much biased to count as an actual review.

[@TheDocPixel]

Exactly. Just because it has over-saturated colors does NOT make a picture "beautiful". Far from it IMHO. Bad composition + saturated colors + mistakes in every prompt* = TRASH. Not one keeper. SD1.5 can do better. * To be fair, the prompts were horrible: RAW and drawing in the same prompt; misspelled words, "masterpiece"-which has been proven to make horrible anime with realistic SDXL checkpoints. The list is long why this is simply shilling for StableDiffusion.

[@obscuremusictabs5927]

boo! lying by omission, boo this man!

[@screamingstarprime3503]

No thanks

[@mr.entezaee]

Sad :( not free Maybe End of the stable diffusion era

[@kawa9694]

The reason is you running sd3 online and this consume resources for stability ai ,so that why you need to pay 😢

[@froidup]

Ahora tienes que pagar. Adiós Stable Difussion

[@CoconutPete]

ugh comfyui

[@lennoyl]

To be honest, SD3 looks disappointing: texts still have errors, hands still have 6 fingers, face are still weird, etc... and I don't care it does more artistic things. AI is a only tool. the artistic touch should stay the human part (prompt, controlnet,etc)

[@user-iu4xj8mc3w]

not local - not stable diffusion

[@marcusaureliusanonymous]

Not local not free

[@DivinityIsPurity]

No A1111, no care.

[@generichuman_]

Ya, Comfyui isn't for everyone, you need to have an IQ higher than your shoe size

[@rickandmortyunofficial8986]

not free

[@Elwaves2925]

Not yet.

[@gwcstudio]

They are gonna try to paywall the genie

[@OlivioSarikas]

open source model comes in a couple of weeks though

[@openroomxyz]

Would be great

[@dsamh]

LOL credits? RIP

[@fittersitter]

Honestly, I'm pretty disappointed. I still had 20 credits, which I used for one of my standard themes (dreamy landscape wallpaper). All four images were noisy and more like a nightmare. Sure the composition is somewhat usable, but the image quality looks more like SD1.5. I won't be pursuing SD3 any further for now. It just doesn't come close to the brilliance of custom SDXL models. I don't know why I should pay money for this garbage.

[@AI_EmeraldApple]

The API version of SD3 just looks bad. The prompt adherence is nice but the overall aesthetics is lacking compared to MJ6, or even many SD1.5 models.