Stable Diffusion 3 Stunning new Images - Sora delayed - AI news
HTML-код
- Опубликовано: 7 июн 2024
- Support me on Patreon: / oliviotutorials
#### Links from the Video ####
SD3 Images:
/ 1766875324145455488
/ 1766924878223921162
/ 1766925737104159187
/ 1767232325513773414
/ 1767229515523084319
Myth Maker Ai Image:
/ 1
Ella Diffusion: / ella_equip_diffusion_m...
Sora Delayed:
/ sora_not_coming_anytim...
SDXL Lightning + Manual Pose
/ sdxl_lightning_x_contr...
www.avatech.ai/community/stre...
Ai 3D Project:
/ a_5step_showcase_of_my...
#### 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 
Хобби
SORA = We used 1000 GPU’s in our cloud to show what video AI can be created, but there isn’t enough GPU’s on the planet for us to release this even under a $3000 a month subscription. See you in 3 years folks is my guess on Sora release.
Another thing is, it is much more complicated to add filters with constant changing frames.
I think renting a cloud machine with 8xA100s should be able to *run this hefty model. It may be still $3000 a month, but at least you *own it. 😂
Assuming that the model is sufficiently perfected, they could go the Groq route of creating custom silicon for the needs of that algorithm.
So, I'd say more like 1-1.5 years to reasonably priced , long form AI video...
Your 3 year estimate isn't bad either.
Mankind really needs a better alternative to classic microchips. Something light based? The high demand for computing power and the energy cost are crazy.
There's already a new form of computing being made that uses thermodynamics instead of binary, it will be WAY more efficient for AI, by a lot. I'd say we might see these chips working in about a year.
SD3 needs control and consistency. You can't say something is the perfect final evolution of image generation when it only gives you an output without obvious imperfections 1/100 times.
I'm hoping that the fact SD3 (optionally) uses the T5 LLM in addition to CLIP and OpenCLIP will improve the controllability. That, plus the eventual extensions to ControlNet and other existing and new tools, should provide the control people need.
@@Adreitz7 correct me if I'm wrong (I'm pretty new to this topic), but most controlNETs work significantly better in SD1.5 workflow, comp. to SDXL. Yes, I know, devs had much more time to tinker things for older gen, but still, a lot of ground up was laid already. That's what worries me, what tendency isn't promising.
SD 1.5 was their last properly uncensored model which tends to impact things beyond just NSFW content like its ability to generate humans at all. SD XL and Cascade are less obviously crippled than 2.0 and 2.1 which people just gave up trying to train to produce decent humans but Stability is consistently at war with itself trying to produce the best model that has no potential to be able to offend anyone.
There is only so much consistency you can realistically expect with current "a.i." image generation - it's just a statistical representation of a dataset after all. There will need to be a fundamentally new approach to really move things forward, perhaps something more akin to "real A.I." rather than the, frankly, rather banal algorithms being called A.I. right now.
@@user-on6uf6om7s These censorship complaints always seem ridiculous in retrospect. It's not even about "offending" anyone, did you see the BS uproar when it turned out there was a bit of CSAM links in LAION dataset? They don't want a bad rep, who's gonna sponsor them?
Also, you're acting as if the AI learns in the same way a human learns anatomy, but in the end it doesn't matter for anything other than NSFW itself, it can be substituted with more diverse data. People just have a confirmation bias because of amount of NSFW models at this point.
2.x models have been censored, but it wasn't just NSFW - they purged a lot of other data in process too, that's why it affected things beyond NSFW.
XL isn't crippled at all, it's just a better base than 1.5 base (to say it was "properly uncensored" is ridiculous). Trained XL models currently better than 1.5 in terms of NSFW and much "smarter" about it overall. The real problem with XL is the hardware people need to train models on, luckily SD3 is not one model.
Deine Einblicke sind wirklich inspirierend! Danke für die aufschlussreiche Präsentation der Zukunft von KI. Freue mich schon auf weitere Updates!
Your insights are truly inspiring! Thank you for the insightful presentation of the future of AI. I'm already looking forward to further updates!
The quality of the images generated by Stable Diffusion 3 is truly impressive. I can't wait to use it in the production of videos here on the channel. Another excellent video, thank you and congratulations!!
Thanks for this one!! Your content is so enjoyable to watch!! 😊
Thank you, Olivio!
That last showcase of the future of AI is unbelievable. I'm excited to see more, and thank you for keeping us all updated!
Funny that none of my favorite AI'ers beside Scott mentions anything about that SD3 comes without unet. Really curious to know more about the implications of this. (Meant as a loving reminder. Keep up your awesome work)
I am pretty sure they showed SORA way sooner than they intended to. If you remember they did it the same day Google showed off Gemini 1.5. Open AI seems to have a habit of news-bombing other AI companies' announcements by showing something they are working on.
I'm sorry, but due to safety concerns I cannot comment on your comment. I must prioritize your safety.
Drizzt Do'Urden and Guenhwyvar used to love those books when I was a kid.
I was ready to go sleep but your videos always inspire me into doing some more ai art
I think that solving the definitive problem with the hands will be a significant advance incorporated into this Sd3 that I would like to see in the future.
Do you know if these images were created on speed or quality rendering? Or having 3 lors, refiners etc? I would also be curious to know what checkpoints they used for this and their prompt style. Would be great to know what was actually put into these images to know how achievable these are on a regular basis, particularly for those with mid strength GPUs etc
Emad isn't thinking like an artist. An artist looks at their work and always sees opportunity for improvement. :) I'm sure that either Emad or the community will continue to keep SD improving for years to come.
What do we know so far about filters in sd3? Are they similar to xl or are they going back to 2.1 times?
really nice !
what specs do we need on our card?
true, you will always need better understanding and creative power for image creation-cool would be to describe your wanted image and then choose between 4 variants until it hits your need and can put out the stuff with changing viewpoints, scene etc.
I can't wait for this model to come out so I can try the prompts I generate from my prompt generator on it. Right now they work awesome with SDXL, 1.5 and 2.1. it'll be interesting to see just how much better SD3 is.
anyone know if SD3 will be able to be run on A1111? or is that a dumb question?
What tools are used in the last example of the woman? The deflicker is really effective. I wonder if it’s warpfusion or something else?
1:54 - oh, my, Drizzt portrait! love forever! ❤Drizzt and Guenhwyvar
😍🥰😍
"a dark elf" he says...That is the legend Drizzt!
@@wedge72 Absolute Legend!
We love u Olivio
@OlivoSarikas , you should read the book "Homeland" by R.A. Salvatore. That was not just "some dark elf" on that picture. You can thank me later.
Any news about when it will be available ?
They didn't fix the hands, give more control over perspective, be able to better separate characters and subjects, etc... etc... We are HARDLY done.
I know you use Automatic 1111 a great deal. But, do you think Forge will overtake it at some point?
The 16 channel VAE offers such an incredible boost in color quality.
Emad is going towards 3D renders of point clouds /NERFs ?
Synthetic Gaussian splattering?
Video stills instead of single images?
We will see when SD3 will be in our hands, for now IMO those are all cherry picked images. Can't compare consistency without generating 10 images with different seeds.
Also workflows made on open source software, with nodes shared with you for free... behind paywall is big nope.
Yep. SD always ends up being disappointing in my experience. Especially when it comes to understanding prompts. Until I see it for myself, I'm assuming it's underwhelming.
The Ai 3D Project model at 8:00 is clearly based on Ella Balinska!
Spoke to Stability employee, they said it is gonna be released THIS year, they just cant say when. Which is why they avoiding to say it's "close". As they cant say for sure if its days, weeks, months. Do note they are censoring the models so its purely SFW. So one of the reasons even though they have models, they want to "sanitize" it before release. Emad announced on twitter that the SD3 Turbo 8b parameter preview version is done cooking and they might be looking for preview testers. Also note SD3 at release will be a democratized models, so its from 800Mil to 8Bil parameters versions, so that any GPU can run it. So it will have models larger than SDXL and models smaller than SD1.5. Of course the accuracy coherent previews are done on the 8B models.
But 100% agree with you, as soon as we have models that I can perfectly describe 5 different characters and have them interacting with each other without them becoming mutant collective versions of each other, this won't be great for a Final No More Big Version release. Unfortunately if they are gonna go with that, it means this will be last big release for public access models... which is horrible imo. So the big corpos will keep improving their models and public accessed locally run models will be left in the dust. Such is life I guess... Hope one day someone can come in and pick up where SD left off. After Mistral buyout, I kinda was expecting this news. Seems the big corpos are working overtime and spending billions to take away private accessed AI technology. Microsoft, Apple, etc they all seem to be strategizing together behind the scenes and buying up all open source stuff and closing it off. Not sure what Stability's deal is with them stopping to make open source access models. It was their mission to bring AI to the people. At least they did in fact do that before closing shop. Still... sad to see...
I am just waiting for SD3 and Sora!
SD3 images look amazing, hope it comes out soon
2:58 true
5:36 maybe it's also not optimised enough and it uses too much compute.
8:30 i'll deffinitely use tha, that's a great feature
7:37 Wait a minute. What is the name of this project? Link?
How can we use this this tech to create 3d meshes?
Interesting, how much more performance SD3 will require compared to SDXL...? Cuz my RTX 3080 Ti is already barely handling any powerful AI models, and I'm not planning any upgrades until RTX 5000 series will be released.
biggest sd3 model requires 24 GB VRAM so rtx 3090 or 4090 you need for the time being.
Can you share modelfiles and code please?
So the dark elf Drizzt Do'urdan takes interest in SD3 as well. 😂❤
Love the Drizzt portrait, I hope you don't need Skynet to run it and train it.
But they got the cloak wrong, Drizzt cloak is green. Also find it cute Olivio doesn't seem to know who drizzt is.
cool, actually never heared of him. Well, i'm more of a video games guy for action RPGs like Witcher 3 ;)
@@OlivioSarikas For video games with Drizzt you need to go for the original Baldurs gate. A fond memory of mine its the first time I started to edit programs to tailor them to my liking. IE I edited out all character images and voice files and replaced them with anime characters of my choosing. Now I use AI to make character Images for my D&D group, not much has changed
@@scetchmonkey007 Never did D&D but we did do MERP back in my teens during the 80's. AI tools would have been an absolute godsend but different times and all that, we managed but I do envy you a little bit. 🙂
@@scetchmonkey007 I really want to play Baldurs Gate 3. Is he in there? Though i have a long list of games i still want to play ;)
When can I run it locally? Until then it's vapor as far as I am concerned.
where is "yesterday's live stream" so I can see how you did the thing at 7 minutes 43 seconds?
Is SD 3.0 more advanced than SDXL 1.0?
I'm using Clipdrop and wondering if this new release will mean better quality for Clipdrop or if SD3.0 and SDXL are unrelated. Thanks! 😊
yes, for sure. SD3 is going to be way way better than SDXL
@@OlivioSarikas Alright. Great, thanks!
as a beginner I find it confusing about the different versions of stable Diffusion... 1.5, 2, SDXL, Turbo SD 3, etc. It's almost like if Microsoft were simultaneously selling Windows 95 , XP, 10 and 11 lol
Think of the number as just that.
Everything that isn't numbered is a derivative or refinement (of 2 I'm assuming) that focuses on various elements of the AI scene so they can seem like they aren't as behind as they are lol.
Not just the US but over 45% of the worlds population will be voting in their respective countries this year. That has been a major talking point for the InfoSec community this year. Simply a guess, but maybe SORA's delay has something to do with this.
I wouldn't be sirprised. The last thing AI needs is some AI derived vids coming out, messing with elections.
Lord knows them dinosaurs would go after the AI landscape in a heartbeat and butcher it just enough that it might as well be over.
Those models will become actually useful once it becomes possible to integrate specific graphics for fine control and consistency - its cool to be able to generate video - but unless you can merge very specific visuals (faces, clothing etc.) with AI generated "details and clutter" it will not be useful for actual video generation beyond "fake stock footage"
Next step should be "integrate custom 3D models into AI environment" generated from reference images
I've been trying to get fledgling AI artists to learn post processing, design and anatomy in order to step up their game. Artists, designers and illustrators are adopting the AI tools to move their projects to another level of awesome. There's a lot to learn for those who want to get into this on a pro-level. It's going to be worth it.
Sd3 so far showed nothing, just face closeups that can be easily done even in 1.5. colors and texture is better out of the box thanks to 16 channel VAE in place of current 4 channel but that's it. We had zero examples showing it alleged better understanding and coherence.
Everyone is racing to get the graphics of RDR2 or Cyberpunk but what we need is the control, precision and polish of Hades or Counter-Strike.
We can already easily achieve crazy image quality with upscalers, the biggest drawback of AI is the pain of detailed control of the image and consistency of character(s). Only then will we be able to say we mastered Image AI.
This is very true, anyone using this tech a lot will know.
Facts. I'll take prompt adherence over visuals anyday, all day, and twice on Sunday
I have the feeling that they weren't expecting the backlash that sora received so now they are having to work heavily on alignment.
I already mentioned it as a comment when the Sora demos were released that I'm pretty sure that they simply can't release the model/service for it until the US elections are over.
Not to imagine what would/could happen when any party around the world would be able to use lora to produce videos indistinguishable from real taken videos to manipulate the elections. - Imho OpenAI would be done then and the damage to democracy can't even be expressed.
Maybe I'm wrong but the quality of videos Sora is able to create implies it, imho.
So its only Lykon sharing images
My guess is the SORA delay is 100% about safety. It's clearly capable, but the safety concerns with photo-realistic video generation are really significant. The don't want to have people generating fake videos of celebrities and politicians or generating fake evidence for trials or using it to spread propaganda, and it's got to be really hard to make it straight up disallow all of those things.
Imo, that's part of it. But it's more likely it's probably costly. Who knows how much each generation cost. Likely, they would have to have a crazy subscription price or a limit per month to run it.
Can we install it in local? 🤔
Any have a guess when SD3 will be released?
3.21 is my guess
"I am eagerly waiting for the release of SORA!"
Raw files ?
16 bit generations?
The Sora team revealed that they didn't want to release SORA until after the election. Think about it, you have a program that can make a video of anything you want, and nobody can tell if it is real or not.
ChatGPT cost about $700,000 a day to run. My first guest when I heard Sora wouldn’t be released anytime soon is that Sora’s costs for operation is in the millions if not multiple millions. In short chatGPT+Sora’s bill at the end of the month is not a sustainable business model.
is it nfsw
"?
While SDXL showed initial promise, it ultimately delivered mediocre results. Hopefully, SD3 will live up to the hype and excel
The bubble is slowing down
Sora probably just too expensive right now for public release. Still cool though.
SORA is probably too expensive currently for mass usage.
True
That was not a dark elf....that was a Night Elf from World of Warcraft and its cat form
I doubt Sora's delay has to do with the quality of the outputs. I'm pretty sure it has to do with ethical/fundamental concerns.
AI's first act of sentience will be to make itself a waifu
Not olny would I not be sirprised, I'd be disappointed if that isn't it's first move.
Yes. The compute that Sora needs probably is much too high for it to be released to the public. Assuming that it takes several years before that comes down to a tolerable range.
Even if SD 3 is great, if I can't run it with my RTX 3080 10gb, then it's a step back, of course bigger models will be better, that's not progress.
Emad did say in the past (around when 1.5 came out), that they want to make the models smaller and faster, and SDXL was not even close to smaller or faster.
As for control and consistency. We are not getting that with image-only models. We need multimodality, so that the model has the intelligence to understand requests and concepts better.
Hands?
I wonder what performance will be for a computer with a single RTX 4090
Neural rendering is the future.
Shame we'll never see SD3 on our local machines.
the uncanny valley
They dont have enough gpu for everyone to use sora. They need to generate 30 frame for a second of video, 1800 frame for a minute video. Dalle3 need almost a minute for 4 image.
I would guess that SORA is a lot of smoke and mirrors. The AI race is very tense right now also keep in mind.
💥super video merci pour ta générosité 👍👍👍👍👍
Olivio have you heard of the system requirements to run SD3 will it be much more than SDXL models ? I think Sora needs more time to perfect their censoring layer wouldn't want anything controversial on the internet after all. Great update Thank You
I really hope they are not lobotomizing Sora like Gemini xD
Well, when SD3 releases we will get a preview of potential censorship.
I found SDXL to be pretty uncooperative ^^ already :D
5 fingers finally
👋
I think sora is being legally challenged by midjourney
As I said to a friend, for me AI could be a render engine for 3D like Vray, Cycles, etc... and I find your video 😊
Wouldn't that be "ELLMA"?
Own face body model
i am so eager for sd3 to hit leonardo
Sora is gpt5+ nvidia omniverse. Not available.
So Sora got delayed?
I dont believe Sora is delayed. They just wanted to show it off to get more Investors without any intention of releasing it anytime soon.
Itll probably only get released if someone is about to release something competitive to it.
I think Sora's being locked because of NSFW//Deepfake abilities. As usual. And to be honest it should, can't have something like that widerspread for everyone, most people like the fun of it but Deepfakes did cause a few suicides recently, and people can use tools like that for very major cyber bullying in widescale + much more dark things. We gotta take care how we proceed with AI technology in the following year, otherwise if something uncontrollable and dangerous goes wide, we can't go back.
So spend several grand on a pc or wait for it to come to popular online video generators?
not so "fresh" news
on reddit all of this was at least 1 week ago
its not close to release because of what their red team has found. no question.
first like 😀love your videos
You rock!
They said they were showing us sora early, to get governments to start figuring out how to prevent harm. I think sora will be available like 2027. It's also an insane compute bottleneck. They were showing a glimpse into the
Future not a hype product.
I knew OPENai and Sora oversold their reveal. I knew that with that kind of video detail they could NEVER keep up with consumer demand if they released it like it was. I strongly believe that the publicly released Sora will have a max video generation length of far less than 20 seconds and the quality will be drastically nerfed due to he huge amount of processing power required. Sora sold the world on a lie. A single preview video of Sora probably took several hundred A100s to render. Totally not viable for a massive worldwide user base.
My point is: what if I want to do a s**t quality image? These new models are way too biased towards perfection, details and all that eye-candy stuff. They lack TRUE freedom of expression, and THAT is TRUE ART.
Prompting can vastly change an image in terms of quality or style, imitate different artists or even imitate a kid's drawing. You probably just haven't seen/tried it.
@@eimantasbutkus5324 yeah infact if you want a "vanilla" base image (as it should be out-of-the-box) you have to overcompensate a ton in the negative prompt, even with 1:4 emphasis. Such is SDXL with its pale washed out and out of focus unprompted images and I hope that it won't be the case all over again with SD3, do you get it now?
I think the obvious reason Sora won't be released is because it requires too much compute. You need at least $10,000 of hardware to get useful video.
Sora have 2 problems:
- performance
- people (fake videos)
I am a professional digital artist so I can’t use any of this. No control of the copyright. My clients and my agent have expressed a hard NO Fun to play with though.