NEW SD3 Medium Is THE FUTURE OF AI MODELS?
Вставка
- Опубліковано 8 вер 2024
- Say goodbye to Midjourney and welcome the new AI model that's set to redefine the future of AI art generation: SD3 Medium! BUT It has issues...
In this video, I'll show you good, the bad and the ugly side of SD3 and how this release will pave the way to the future of AI models!
Have you tried SD3 Medium yet? Let me know in the comments!
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
SOCIAL MEDIA LINKS!
✨ Support my work on Patreon: / aitrepreneur
⚔️ Join the Discord server: / discord
🧠 My Second Channel THE MAKER LAIR: bit.ly/themake...
📧 Business Contact: theaitrepreneur@gmail.com
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
✨ PATREON LINK: / aitrepreneur
RUNPOD: bit.ly/runpodAi
Stable Diffusion 3 2B: huggingface.co...
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
►► My PC & Favorite Gear:
i9-12900K: amzn.to/3L03tLG
RTX 3090 Gigabyte Vision OC : amzn.to/40ANaue
SAMSUNG 980 PRO SSD 2TB PCIe NVMe: amzn.to/3oBR0WO
Kingston FURY Beast 64GB 3200MHz DDR4 : amzn.to/3osdZ6z
iCUE 4000X - White: amzn.to/40y9BAk
ASRock Z690 DDR4 : amzn.to/3Amcxph
Corsair RM850 - White : amzn.to/3NbXlm2
Corsair iCUE SP120 : amzn.to/43WR9nW
Noctua NH-D15 chromax.Black : amzn.to/3H7qQSa
EDUP PCIe WiFi 6E Card Bluetooth : amzn.to/40t5Lsk
Recording Gear:
Rode PodMic : amzn.to/43ZvYlm
Rode AI-1 USB Audio Interface : amzn.to/3N6ybFk
Rode WS2 Microphone Pop Filter : amzn.to/3oIo9Qw
Elgato Wave Mic Arm : amzn.to/3LosH7D
Stagg XLR Cable - Black - 6M : amzn.to/3L5Fuue
FetHead Microphone Preamp : amzn.to/41TWQ4o
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
Special thanks to Royal Emperor:
- TNSEE
- RG
- Gluthoric
- Peter Bernaiche
- OCGNOW.com
Thank you so much for your support on Patreon! You are truly a glory to behold! Your generosity is immense, and it means the world to me. Thank you for helping me keep the lights on and the content flowing. Thank you very much!
#SD3 #stablediffusion #imagegeneration #ai
▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬
WATCH MY MOST POPULAR VIDEOS:
RECOMMENDED WATCHING - All LLM & ChatGPT Video:
►► • CHATGPT
RECOMMENDED WATCHING - My "Tutorial" Playlist:
►► bit.ly/TuTPlay...
Disclosure: Bear in mind that some of the links in this post are affiliate links and if you go through them to make a purchase I will earn a commission. Keep in mind that I link these companies and their products because of their quality and not because of the commission I receive from your purchases. The decision is yours, and whether or not you decide to buy something is completely up to you.
HELLO HUMANS! Thank you for watching & do NOT forget to LIKE and SUBSCRIBE For More Ai Updates. Thx
👋
SD3 has enormous potential! 👍 👍 👍
I LOATHE ComfyUI. I call it RatsNestUI. But if that's the only way to use SD3 when good finetuned models come out I'll bite the bullet.
censorship is what really killed SD3.
it is SO limited that it's kinda funny
SD3 has been pretty bad in my initial testing. I'm really disappointed. I'm not sure if its just overly censored to the point where its having a lot of difficulty generating anatomy, or if there's something terribly wrong with the way were using SD3.?
The prompt comprehension is really great, though. But I was hoping it could compete with Dall-e 3/Copilot Designer, and its nowhere close.
I should mention, I just use it to play around. I dont make money from it.
The big worry is the restrictive license making it difficult for people to finetune the model. We'll see how things work out.
its the censorship, they must have removed all images of people laying down
so i guess you have not tested or dug deep enought on other models....
sd3 medium is a 2b model which is severly undertrained (stability just released it bcuz people were constantly asking for weights)
dall-e or mid or any other model is very big compared to 2b, 8b is the one real deal which we all are after or maybe people who know about these models in general
sd3 learns fast so maybe crowd funding would be a good idea we need some trustable team for it.... well in total the only thing messed about this model is anatomy and license, rest the model is clear upgrade
also a tip if you are doing those women on grass test, dont use word lying its not trained in model yet or censored so it generates blob.
@@hmmmmmm_3429 I mean, Ive been using SD 1.5 and XL for the last year and a half with good results. But SD3 has been challenging.
This is 2.X all over again. Completely crippled with overaggressive censorship. Zero understanding of anatomy. This affects not only humans, but animals and creatures as well.
On the bright side, it's great at landscapes and objects. So I'm sure some poeple out there might use it in their workflow.
For finetuning, the enteprise level requires approval from SAI. Going off what other's have been saying and SAI's responses, it doesn't look like they want anyone finetuning with NSFW. So SD3 is effectively dead.
It also affects ponys. Poor creatures.
Wasn’t able to get a good turtle, it cripples the model so bad
the community will simply not improve sd3 with that licensing
they already are...
@@Zuluknob Civitai just halted SD3 loras because of the licensing
@@aranchi20 Yeh, and a bunch of people have stabilityAI. sd3 2b was meant to be in beta not release.
@@Zuluknob theres a lot to unpack but Civitai stopped it for a lot of reason. For example you need to look at derivatives in the licensing i think it is, because if you make a model/lora , you are now responsible for what that downloading user is now doing with your model. So lets say they make an illegal looking image, well your model "let them do that", or make a lot of money using that model, etc.
It's so censored that it can't generate anime without it looking horrible.
for real ? that bad
Censored ? So it can't generate nsfw?
I am going to assume they didn't 'censor' anime, its almost certainly a lack of training data that is anime.
A pony merge or something will come out soon that fixes all your worries.
@@SorhatodTop1corruptLeaderIt not only can't, but that's exactly the main target of the censorship.
@@SorhatodTop1corruptLeader It can barely do humans period.
TLDR; SD3 dataset wasn't diverse enough and was mainly trained for aesthetics
I second this. I spent the whole day testing SD3 using type of images I use for work (slightly surreal, sci-fi, utopian fashionshoots) and it kept giving me regular dudes in rental homes standing perfectly upright. Images themselves incredibly detailed and realistic but 90% of the prompt was ignored.
the model is released in haste after community was angry for no weights from weeks and just image teases, the model is severely undertrained, so not even trained.
some people tried training it for some minutes and it started performing better, in just minutes... that is how untrained it is..... also it absorbs things fast so you guess it correctly, the dataset choosen was very very small and rigid, just aesthetics....
The model is censored into oblivion so much that it doesn't understand human anatomy at all and generates eldrich horrors instead of humans, even absolutely sfw pictures.
The license makes it impossible to create finetunings of good quality as this is a lot of work and compute that requires a lot of money, and Stability staff refuses to sell license to the biggest finetuner of SDXL and publicly insults him in their Discord when he politely asks them to sell him the license.
There won't be finetunes fixing the fundamental issues of the new model.
SD 3 is dead, and Stability AI with it.
Instead of taking copium we need to move on and rally around another project, like PixArt.
I'd say we need to wait a little. SDXL was trash when it released too. Sure, not to this extent, but this is also theoretically a much better architecture. I would still say this is closer to SDXL which was eventually wrangled into something useful, than to, say SD2.0 which seems to have been fittingly forgotten. I think there may be ways to take advantage of the absolutely massive NLP capabilities here and achieve good results with relatively little tuning.
@@df1nedsdxl had a different license though.
@@df1ned problem not with model quality. Problem with license and sharing - you can't do anything with this model. Only education
@@snatvb I fully expect that to change in the future. If it doesn't - then yeah, what you said is pretty much how its gonna go
@@df1ned People much more competent in the topic than me describe the situation like this.
When you finetune a model, you overwrite existing knowledge of the model.
If you are training the model on a topic very similar to what it already knows, this requires low amount of training epochs and thus erases low amount of model knowledge.
If you are teaching a model something from scratch, like human anatomy and poses that are pretty much completely censored out of SD3, you overwrite so much data that the model loses the vast majority of its capabilities in the rest of the concepts.
So you have to effectively replicate all the training Stability AI did without their budget, without their dataset and without their knowledge.
For SDXL it took a huge amount of effort to achieve something usable with finetuning, and it was a model with much, much better understanding of human anatomy.
censorship needs to be removed.
Let's be real, this release is way worse of SD2.0
-awful licence
-gigantic hype
-synthetic traning dataset
-months of wait for a half-baked model when they said they were taking their time to make it perfect
absolute disaster
It is censored. Terrible with human anatomy. Really disappointed
Except that SD3 is still heavily censored so who cares? Until its cracked or if, its worthless.
Yep, you just saved a bunch of people a few minutes of their lives.
censorship = downvote and onto the next video :)
Why not use stable diffusion xl models then
Are you seriously still using their older models to generate uncensored content? If so, I admire your patience. But keep in mind that CivitAI will soon be flooded with fine-tuned versions of the new checkpoint, so to answer your question, EVERYONE should care. The quality of the base model directly influences the quality of the fine-tuned models, and almost everyone relies on those fine-tuned versions.
@@anatolydyatlov963 SD3 has not been cracked at all, its till heavily censored for NSFW. So yes we are forced to use SDXL and 1.5 for that. SD3 most likely will take years to crack as StabilityAI has flat said its locked down hard and experts have said writing new models for it will be VERY difficult in comparison and could take years of training. I personally have over 40+ checkpoints I use on a daily basis my SDXL folder is around 500gb of models alone. I am far from an average AI user, working in IT and a company that wants to work in AI i spend more than half my day researching this stuff. To me on my time uncensored models is a must have. and at this point in some cases SD 1.5 STILL does better work than even the latest models in SDXL.
@@storkyfallout6516 Thats exactly what we do..
This aged badly.
No finetunes being made for SD3.
CivitAi has banned SD3 on their website.
Heavily censored training on images.
So NO, it's not the beginning of an amazing series. Your two cents can be thrown in the poubelle.
painful truth
I tested it and I prefer 1.5
Base 1.5 ? You have seriously some bad preferences man
like if you said SDXL I would have believed you, but 1.5 😂
bro what
I am guessing you mean a 1.5 merge of some kind, and not the base SD 1.5
If that's the case, wait a month and see if you like the SD3 merges coming out.
For anime style images 1.5 is unmatched.
Have you ever heard of Pony?
SD has not made any progress since 1.5, and the amazing custom models based on it vastly outperform most new ones if you know how to use them.
SD3 looks the same or even worse in many cases compared to some of the best custom 1.5 models.
I would have expected it to reach at least near DALL-E 3 levels after all this time, especially when Emad Mostaque was saying that they could make the models smaller and lighter while having the same or even better capabilities back when SD1.5 was released, but it's not even close.
for a 2b model for text, background and prompt adhésion it's not bad, especially since it's several time smaller than dall-e 3. It's just extremely bad at people, like 2.0.
My guy, its been 20 months since SD1.5 came out. I know the tech is developing fast but that is like zero patience compared to how long we have had to wait for these kinds of advancements in the past. And if you are looking at the past 20 months and saying there has not been any progress, Idk what to tell you. I guess AI has no future.
Plus, Dall-E is closed source, Stable Diffusion is Open Source. There will be merges in the near future that split the diff on what SD3 is lacking now. Dall-E you get what you got.
@@nagrom777 SD 1.5 2.0 XL are open source, model after that aren't.
@@Ethan_Fel I am pretty sure sd3 is still open source, just not a free license for commercial use.
I hadn't looked into it in much detail I am willing to admit, so I am happy enough to be proven wrong.
The main Problem with SD3 is, that due to Stabillity AI's licensing most creators won't make a community model. If SAI won't change that SD3 has no chance.
Stable diffusion 3 is great as long as you don't expect photorealistic images. Its great for pixel art, 3D and art in general. Currently working on a prompt guide.
4:12 "Really really good"???? That guy has 6 fingers! (5 fingers + 1 hidden thumb)
if you only do the basic ther is always anatomic error, the good part is the reste of the anatomy reproduction sucess
You are desingenuine ;)
No really the quality of the model really is good, base model aren't perfect and basically no one use those, they use finetune models
Let's see if the fine tuned models solve the weird poses.
@@ryzelincurve8295 I agree, other than the hands, it does indeed look good.
@@joannot6706 Fine tune models? You're clearly more experienced than I am because I don't know what that is. Okay, I look forward to more news. Fingers crossed.
@@Mumra2K Juggernaut or Pony models instead of SDXL for example. They take the base release and run it through Dreambooth or another trainer to further refine the model's knowledge (the unet for visual patterns and clip for token/text patterns.)
We got to remember that SDXL was also highly censored when it initially came out, although very impressive in terms of quality and capabilities...
Now we have Pony 🤷
Respectfully, Pony should be considered a different thing from SDXL, since their LORAs don't translate between each other. I also imagine that Pony took a lot more work on top of the original great work that SDXL put in, significant enough that I don't attribute Pony's success to SDXL.
The SD3 license prevents Pony from doing the same thing with SD3. I wouldn't get my hopes up.
@@eyriewow4297 why ?
I'll think you'll find that the censoring is an issue for you, that's why you can't make images of people lying down, even if you're not making "those" kind of images it's still affects you indirectly and very negatively. This is why we usually complain about it, it causes problems.
It's like getting a car to go to work in, but making it so it only goes to walking speed, in case you run someone over, even if you have no intention of running someone over, the limited speed makes the car useless.
I dunno how long it wouldve taken me reach that conclusion (why the woman laying in grass thing wasnt working) but you're probably right... if SAI says women are vertical, not horizontal then that's what SD3 will spit out
I'll wait for a Juggernaut(like) version of the model before really trying to do anything with SD3.
ideed, and also controlnet
i would have to wait for the next gen GPU to train that anyway. by the time training scripts mature for SD3
My question is if fine tuning SD3 will inherit the non-commercial license?
Yes. The license says that every derivative product, as they call it, must follow their license as well.
Though it should be noticed that the license doesn't actually follow the images you create with the model, only the model itself (even if the models created with these images also have the license, strangely enough)
@@gabrielbuenodossantos5203How can they know if we're using a license or not? I didn't get that right
@@FerikkusuOC cant really but if you are big enough it might be too risky
@@FerikkusuOC a) you have a website that offers this product b) you don't have a license for it. At that point somebody might become interested in what are you using to run it. Then, perhaps if the model has some fingerprint, they would be able to catch you in the act. For example, let's say that this model accidentally (or on purpose) when prompted 'qwerty' returns a specific pattern. Though I won't try to claim any particular method, I believe there are several ways to sniff the model out once there's a suspicion. Worst case - the cloud service providers like google will scan for weights.
Edit: as for the 'I have a licence' argument - they can just ask you, especially if the business you have listed on your website is not in their license database.
I'll stick to SDXL for now. I'll use SD3 for logos only.
I'd be interested in a video about comfyUI, how to use it, and extensions if those are supported.
It will ended up like SD2.0 😂
no its a overall upgrade , sd3 is cureently very very undertrained on top of being heavily censored by the "safety" team of sai, it picks concepts faster than any model so wait a month for good finetunes to appear.....
The issues are pretty huge imo.. That sucks
Kinda doubt it, seems to be a 2.0 with a confusing license
SD2.0 was changed to 2.1 fix but it was still 4 model at the end. the controlnet relese + hardware requierment and previous work and on top of it the NAI leak...
all of that made the experience not worth trying back then. what model to finetune ? will there be a controlnet that i care about ? how about gen speed ? where is NAI for SD2.0 ?
many are willing to pay the entreprise taxe to help their own finetune but as of now... they are waiting answers. (some won't)
it would take at least 6 month in my book or maybe a year... (or another company) regardless SD3 will be the last so...
your guess ?
Because the first SDXL images were perfect >_>
Patience, Iago.
License...
5:48 I needed that laugh!
I think the 1.5 + merged models version is the less buggy one (imo).
I still hope ToonCrafter releases a decent auto-coloring tool with guided sketches. Then I could focus solely on animation and leave the automatic color and shadows to the AI. Seems like it won't happen soon though.
I would really enjoy seeing a ComfyUI tutorial installing the SD3 model (and any new models that have popped up in the last few days), and maybe going over some common workflows and nodes within ComfyUI. All the ComfyUI tutorials I find just point to premade workflows, and I would love a break-down of what nodes you can add or use to improve generations!
Generate slimey green wall. Turn it into grass. This is the exact level of cursed spaghetti I expect to exist in the background of future tech.
I wonder if generating "wall with realistic grass tapestry" instead of "slimy green wall" would work? It would be funny if the bodies suddenly become distorted by "touching grass" lol.
The biggest problem now is sd3.0 medium can't not train without A100 GPU, it not friendly for community to easy to train it with local machine
A100?, people train SDXL on a rtx 3090, and SD3 is smaller then SDXL, it can be trained with just a rtx 3090 no problem
@@volcanowater right now, the only thing i can see may training base on sd3 is lora, and it take around 27gb vram to train a 512 batch img, testing dreambooth train ckpt take more than 40gb vram
Picked the smallest problem and called it the biggest lmao.
People absolutely don't get it
The model was released as a very powerful development platform for everyone to use.
We, the community, just need to put some training effort into it, and we're gonna have an uncensored DALL-E 3 (or better maybe) in no time
Thank you! This is the type of videos I would like to see more on this channel!
The real issue is not the model itself, but the license offered by SAI.
The doubt it has created in the community doesn't encourage anyone to train LORA or checkpoints. Until SAI clarifies its position, no one will want to waste computation time on SD3, at the same level as SDXL.
As a result, while we could already be seeing improvements through the community, SAI has cut itself off from one of its assets: the users who advance its product.
I absolutely REFUSE to support this level of censorship. Oh screw this model. I'm not going to sit here and pretend like I dont. Yes I totally use AI to generate "fun" adult content. No way in hell I'm using this model.
There are thousands of images of people lying down on the internet. It's crazy to me how they wouldn't include that in their dataset since it's going to be a fairly common thing people will try to get SD to generate. Dall-E 3 does it quite well.
Nah, someone is going to wipe stability ai off the map very soon. They did this to themselves.
To surpass other models like Midjourney it won't be enough to just generate pretty females. It needs to be good at:
1. Understanding a wide variety of artistic styles
2. Blending different concepts to create novel concepts
3. Replicating various types of lighting and camera effects
4. Adding a certain twang to the sauce
I have a sneaking suspicion that the future of text to image generation probably isn't in single companies with single models. I think that standard Diffusion architecture models have kind of topped out and hit a variety of limitations, and the future of the medium will likely be in some sort of parallelizable architecture with contextual understanding (perhaps some form of image tokenization and a naive transformer?) which is trained in a federated, open and distributed manner by groups of people with similar interests on their PCs, as opposed to being trained in a centralized manner as SD originally was.
if you sencor all anatomy and anatomically correct poses from your data set you this is what you get. if you go to art school you learn anatomy by drawing body parts in different poses. and that also include models without clothing. you can't learn anything without using the real stuff. if you sensor all the skin showing you fail hard in making a realistic model.
Exactly! To create realistic depictions of humans you need to know "how they work". Anatomy lessons for artists also include sections on the inner anatomy, such as the skeletal structure, muscles, fascia and tendons, fat deposits, and so on.
While AI doesn't "draw" in the same way as humans, I have seen "partially rendered" SD images (e.g. at step 10 out of 20) where the person in the preview had much less clothes than in the final image. So the model did "draw" the anatomy, and then added the clothes in later steps.
Just to save your time, data and sanity, don't bother with the base model. Just wait for the fine tune one.
I mostly agree with what hes saying about eventually the community "fixes" SD with fine tuning....but...Does anyone here use SD2.1? Hard censorship is a death for AI.
I really hope that people have kept their SDXL finetune and lora training datasets around because as far as I can tell they can use those exact same datasets for SD3.
oooooh yes ... they did ! CAREFUL THESES GUYS LOVE TO PROVE THEMSELF RIGHT :D
Rip MJ he said, yet SD3 is horrible XD
If you live in brazil the price is a problem.
It's apparently really good at transporter accidents!
zero integrity , SD3 is a joke and its terrible licensing means that no one is interested in finetuning it to fix the garbage . your words need so many caveats in order to compare it favorable only to base models . acting like juggernaut and pony variants dont exist , even 1.5 finetunes are leagues better , but at least sd3 can render text a feature no one cares about .
Juggernaut is overrated as fuck, it's minimal finetune on a less than 3000 image dataset. There's tons of other finetunes on CivitAI that blow it out of the water.
@@ShogoKawada123 really only 3k ? thats surprising if true . i find juggernaut is generically good for cinematic realism rendering of items . certainly depends on what kind of style or concepts hoping to render . some checkpoints are overtuned or lacking concepts but great at their niche . consistent_factor_euclid sd1.5 for example is very overtuned and cant do textures however renders materials of items in a very interesting way . usually end up mixing a few together for the specific "style" of the target render . do you have specific examples of sd1.5 or sdxl checkpoints that you consider to be so much better then juggernaut?
as someone who is very active in community and know the insides of stability
about sd3
- currently the model you have (medium) is very undertrained (it was supposed to be beta model but they had to release because community was just asking weights)
- safety team over censored the model on anatomy on levels of sd2 (so words like laying etc wont work but a women on grass and long descs will work)
- the model is only 2b (smaller than 2.6b of sdxl), it can understand concepts and everything better but it was not trained enough
- the model is a clear upgrade and people are forgetting about SAI models, sdxl was bad with their all previous ones aswell
- stop comparing it with dalle,mid,ideogram the model is only 2b it cannot compare, the real model we are after is 8b one
My AI Overlord: Laying down aside (experienced AI artists know of this issue from 1.5 already and that it is "lying_down"), the license is bad news since it acts as a chain and collar for anyone who wants to use this model. The money isn't the issue unless somehow you are collecting 20 from everyone down the line, the act of license should deter most people from switching to Pony in the first place, since any work done on SD3 can be pulled if SD retracts the license (for say, making NSFW work for example!)
the license of the model is a big NOP
Nice video, but why does the title say "RIP MIDJOURNEY"? Can you do the same things with SD3 like with midjourney but better? Like create professional photos for linkedin etc. from selfies? Or generate icons for apps?
i think there is a very deep divide between what we consider as good aesthetics in the sd ai-art community cause it baffles me how peple think this sd1.5 aesthetics=great aesthetics. i dont know how y'all would think that when midjourney and dalle is whopping our asses in aesthetics arena. The aesthetics in this is verrry poor, the api aesthetics however is absolutely incredible idk why you didnt get that, wth is wrong with stability.
I was with you right up till, wait till the community gets their hands on it. Which is absolutely true... the problem is I don't think people will bother. The license is so restrictive why would anyone waste their own money training finetunes for this.
Gemini is better than SD3 so sad.
Information is interconnected; you can't just skim away NSFW content and then expect everything else to stay intact. Just look at what happened when OpenAI tried to censor ChatGPT.
SD3 is a joke. Body horror, censorship, and a license on par with Adobe's latest shenanigans. They've completely lost their marbles.
I would love to see a comfy UI sd3 tutorial. I want to start using comfyUI but i dont know where to start. I get lot of error.
Same
Olivio Sarikas and Sebastian Kamph both have Comfy tutorials on their channels.
I'd recommend using Stability Matrix as an installer, and install Comfy and and Stable Swarm. Swarm is a web interface that uses ComfyUI as its backend that makes it look and work a bit more like other UIs so there is less of a learning curve, and you can still access the workflow if you want or need to.
Stability Matrix really simplifies the installation. And also comes with a model browser and helps centralize all you SD related stuff.
@@nagrom777 I didn't bother with Stability Matrix but I did go through StableSwarm as I can't stand Comfy's nodes. It's just too much hassle to deal with all the errors and the manager only goes so far.
Well, if the dev of pony doesn't make a finetune, I am sure someone else or a group of people will try to find a way to make a good finetune.
Please show us how to fine tune sd3
I love your content in general, but is this sponsored or something? This model is unusable in current state
I'm pretty sure StabilityAI just cencored the Model to hard, the word "Laying" itself is cursed... including it in your prompt just screws up the whole render. Probably because they are afraid of it being used for Pr*n. So, all that we need to do is wait until the CivitAI pr*n peddlers get into the Model generation and they'll fix it pretty soon.
SD3 is truly very strong and once the Prompting system gets implemented in community generated models everything will get better!
I truly wonder if clip_g, clip_l and T5 can be tied into SDXL model workflows because that alone would be an amazing win!
Why would you be using word "Laying" - when you want some person on ground? It is like "lie down" or "lying in bed". If you do laying (putting something down), it is like "lay bricks" AFAIK.
I'm mixed on it's quality, just like it's output. It's good at a lot of things (once you get the right sampler/scheduler) but does appear to suffer most with humans and does nothing for hands. Text is better but still not good enough IMO. As you say, none of the base models have been great. Things only get better once the community models appear so I'll mess around with it while waiting for them to pop up.
Thank you for summarizing the recent discussions about SD3. Yeah I remember SDXL coming out and it was totally horrible, everyone saying it's utterly useless. Then NAI showed their tune and people saw potential with it. Now we have tunes like Animagine, Pony Diffusion and Autism Mix. I'm also optimistic about it's future potential but it could take while until we really figured out how to finetune it effeciently.
Aitrepeneur, I appeal to your mastery in AI text to image models. Please, make a video for those who are having problems installing this model in Comfy.ui. Thank you for your good work. I follow your videos from the beginning. You make it easy, man! Thank you!!!! 😁
SD3 have weird prompt following that average SD user don't get yet, they are used to simple prompt that finetuned version have.
but many of us are smarter than this : if pony showed us a thing is that weight can be overwritten. (both in SD2.1 and SDXL)
any conplain turn into motive. people love to be right.
no one fully grasp what SD2.0 situation were as a whole. SD3 start like SD2, it's rough, intolerable for normies... and also in a way it's safer for end user and commercial use alike.
see the generated image on the CivitAI page. it have the power.
could have been beter of course and always ! but get this : if it was, stupid ppl (especially anti ai extremists) would do weird prompt trick to generate illegal stuff to try and take down stability claim about safety and copyright.
now we are complaining with a "dumb" model. but i won't stay this way.
the main issue is the thing between the keyboard and the chair
AI without the option for NSFW is just boring....basically 80% of my private AI usage is NSFW related :D
Man SD3 is NOTHING compared to midjourney. Midjourney is always gonna be the king of alot of things especially artistic photorealism
this video reeks of copium. the model is trash and the license is trash.
Finally a positive video thank you so much for making this
A Comfy-UI tutorial looks so good..
plenty of yt already
pls make a video on how to install it and maybe pls explain the difference of models and how to setup comfyui cause it's the only webui that can run it and personally i'm having a lot of trouble at understanding it
I said this to someone else, but:
I'd recommend using Stability Matrix as an installer, and install Comfy and and Stable Swarm. Swarm is a web interface that uses ComfyUI as its backend that makes it look and work a bit more like other UIs so there is less of a learning curve, and you can still access the workflow if you want or need to.
@@nagrom777What would be the easiest UI to generate business photos from casual photos locally? (for free)
SD3 is a contortionist model with feel like it is 1.5 model that is censored like 2.1 model but trained on higher resolution like xl
Nobody will use your product not even another companies if your product is defected
Wooohoo this model is so unbelievably bad! Just gonna stay with SD 1.5
If you make tutorial for installing SD3, please mention about compatibility with AMD or Intel gpus.
Hey, it's been a while since you've covered LLMs. I'm curious if there's been any major improvements on fine-tuned LLMs that are tailored for translation. Looking for Japanese to English. Bonus points if there's a way to run an LLM overlay that can read text from a window such as a video game and live translate. Keep up the awesome work!
ugh.
i want a model to generate textures (just flat textures i can use in games) and pixel art sprites. does something like that exist?
I've seen SD3 generates pixel-conforming art as opposed to other models that generates pixels of different sizes, dunno about sprites, but just see for yourself.
You don't need to subscribe to SAI for the rest of your life. Just sub during the period you're releasing something commercially.
AI models are the output of an algorithm, they lack human authorship and have no copyright protections. You cannot impose a license on them or the also un copyrightable images they produce.
Model A, which has no intellectual property right protections creates image B which also has no legal protections. Anyone can do with either whatever they wish and there is nothing SD can do about it.
The only possible way to protect these things is as trade secrets, which necessitates them being run on closed hardware and never releasing the weights. Even then the images generated are defacto public domain.
Well... it's the medium, it has 6 billion(?) parameters less than the model people were using through the api.
Horrible! Skipping SD3 completely.
Yes for the comfy tutorial from basic installation, system requirement and everything pls!
i'm optimistic that our dude will make a fair takes. there is much power hidden bellow the broken anatomy.
just don't ask for an human just yet and SOMEHOW your anatomy is (mostly) fixed
No chance. It’s terrible at people. They were so scared of inappropriate images they made it worse at humans that old models.
all that comes to mind is PHENOMENOL
I don't understand why people are complaining about the censoring and the quality of generations?
The model is censored for obvious reasons, same like all LLMs that are released and the community will always uncensore them, so this is expected.
Second the quality of generated images are way better the any base previous model when the checkpoint size is smaller than SDXL.
And stability AI train their models to specifically be easy to fine tune, this is why SDXL was so good, one of the goal was to be very flexible.
Witch they did also for SD3.
Yes the license fee kind of a pain but I understand that they need money and the model is completely free for personal use, you win some you lose some.
Can't wait for tools to fine tune it!
expected? No one is willing to finetune this payin $20 a month plus the other limitations in the license, hell look at what happen to cascade!
SDXL was good? It was pure garbage lol
Even fine-tuned pony based models are still worse than 1.5 🤦♂️
@@bews Yes, SDXL is way better then 1.5 you clearly just not being using it enough.
I suggest you try to properly experiment with the finetuned SDXL models.
You are probably using 1.5 for very specific narrow use cases.
The only problem with SD3 is the license, this sucks and most likely will hurt SD3 in the long run.
@@zYGote04 well I did search for a good SDXL model like 6 months ago and couldn't find anything - all of them were garbage. Even those pony-based models had that specific bad looking style (my personal opinion) attached to them that was making them unusable.
But I did search again rn and looks like this problem was already solved in another pony based model that everyone is using now.
@@bews I'm sure we gonna see similar thing with SD3, for the first several months the fined tuned models will suck until people will figure out the best way to tune them and then we gonna see sharp spike quality.
The architecture of SD3 is way superior, and I was able to get outputs similar to the finetuned SDXL models form the SD3 base model with proper prompting.
The only problem I see is with the licensing ,which I hope stabilityAI will change, that might discourage people to invest time and finetune the model.
could you tell us how many GPU it take to generate a simple 1024x1024 image ?
At least 8GB VRAM is recommended
Until we can train Textual Inversion and Lora's, locally, with less than 16bg Vram, everyone is just going to stick to SD 1.5.
SD XL was mid.
Probably censored it to the ground. Celebrity faces absent as well, so no surprises here.
Why would anyone pay this company anything, to be honest. Their model wouldn't produce anything without high quality artistic photos, 3D model images and paintings which they pretty much scraped (by using LAION 5B dataset earlier, and whatever else now). I'd rather pay artists for their images to be used in a proper open source generative AI art dataset / model than I would pay anything to this company lol.
What are you talking about? So far everyone hates SD3 and isn't going to bother working with it, due to the Unity-mode licensing scandal.
Y'all get back to watching the video
make a comfyui tutorial so that i can comfortably use it!
why comfy, is is way easier to use in Swarm
Where is the tutorial on how to installand run SD3?
Please, do a series of tuts about ComfyUI!!!
Managed to get it better hours after the release with merging SDXL clips into SD3 basically.. It's still not as good as SDXL right now
Only way I'll use it is pirated. That license is a deal breaker. No software ever invented is worth those kinds of restrictions.
What about performance? how does it compare to 1.5, 2 or XL models? Does this use the 2 step approach like XL models?
single step, faster generation than 1.5 or xl. slower than turbo.
License is a total no go. My company is slightly over 1M annual revenue but uses AI image creation only in rare occasions at all and now we would have to subscribe to an enterprise plan or hire a contractor that is a small company just so they can use it? Solution is easy: we would never ever use sth. like that. Now they are just like dall-e and midjourney but... worse.
Sd3 is nice, till you read the license :D