it should be VERY well noted this is just a **concept** nothing has actually been objectivly created yet from it or if they have they have yet to publically say it.
Great if true. After all the hype and promises from the SD3 fiasco, I rather be skeptical. I will believe it when I see it and hope it can be locally installed.
Just a little tip... I'd put a "spill suppressor" on your green screen key to get rid of the green haze around the edges of the cutout. I'm mainly and after effects guy, so I use "advanced Spill suppressor set to "ultra." Not sure if premiere has the same effect.
To be honest, I am quite fed up with all the installation / configuration / combination hassle with ComfyUI. Yes, it's powerful, but so many extensions stop working after some time or are so complex that I can't remember a workflow after a couple of weeks. We need a NEW, easier way to do all this. And this model looks VERY promising!
always exciting to see new stuff coming up. That being said, if it needs to run a LLM on top of the Image gen, I fear my trusty 2080 won't do the trick anymore, maybe 4090s and the likes. Also I'm a bit skeptical overall, because while it's great to have everything unified and easy to use, it means less tweaking, less personalization etc. So unless the model can be fine tuned like SD and Flux, we'll be losing quite a lot of possibilities. Also Control net: While I'd be happy to just tell the AI what I want, sometimes putting it in text isn't the easiest solution and a proper control net would do the trick better possibly. In any case I hope we'll see the results soon and I'm looking forward to it
It doesn't have to be so extra special in my books, imagine XL/Pony with a better text encoder for fluent language... Could be available pretty soon instead of far away with a lot of challenges. Flux is great, but the compute it requires is far ahead of consumer gpu market.
Interesting and it certainly sounds promising but as usual, it doesn't mean much until it's actually out and usable. Cheers for bringing it to attention though.
We are in pre-history of image generation. Prompt gives random results, and we can only generate again. It is great to get something from nothing, but to get something we want is still tedious or impossible.
The concept sounds good, but it still has to exist, run locally on consumer hardware and follow a prompt at least as faithfully as Flux for me to get excited about it.
I want an image generator with an iterative natural language interface: "give me an elf on the grass" Ok "Pointier ears" Ok "Now he is smoking a bong" Ok Sort of like iterative inpaint
Great video, i am hyped. Are you planning on releasing another video about your life as an AI Artist? Loved the "A Personal Journey between Photography and Life" video.
They are removing control net, it sound more like they are just adding it into the distro version so in fact it is a much fatter code because there is no opt out to load certain models and extensions.
First, I love that you called it a FLUX killer instead of the 1000th MJ-Killer 😂Second, according to the papers, the Model will be smaller than SD3 (3.8B params vs 4.7B params), so I would guess it should be running on any machine that can run SD3, or are the Params not the deciding factor in this matter? I have to admit I am not that technically versed... :(
Lets wait before there is more concrete information and progress to see. I don't see this taking off out of nothing, they are in an early phase, and even if they succeed in unifying there are a lot of question marks regarding requirements. The bar they have for this project is so high that I don't think this will hit the consumer market. Flux is already almost out of reach for a big chunk of consumer hardware.
This sounds like something I would really like to try out. Does it give you for example put a sheathed katana on a characters back? This is something image generators can rarely do.
The "chain of thought" concept makes me think of the last improvement in chatGPT-o1. But I guess it's different here, OmniGen is not gonna be like a SegMoe model with many "experts" in their own field brainstorming about your request.. right ? ^^ My thought is that having everything separated, the controlnets the loras etc might be difficult to maintain, but maybe that's why everyone can run it. Now, one big AI model that can do everything... How much that's gonna weight, I'm really worried that my 3090 recently purchased get already deprecated, after only two months of use x)
Hi Olivio, I've noticed a bit of an oddity on your RUclips channel. There seems to be a "disproportion" of sorts. On one side, there's a hefty number of views and likes, and on the other, a barrage of negative comments. So, here's my burning question: Are you pulling some digital wizardry to game RUclips, or is this just a quirky glitch in the matrix? Can we anticipate a flood of positivity washing over your comments section soon? Fingers crossed for an algorithmic miracle! ✌
Interesting to know ... but I think no one cares until it actually goes out for real ! Remember all the hype for SD3 ? now it's in the toilet and the community flushed-it like it was shit
If this runs locally, I'm all over it.
It does run, but it takes like 40s per image on a a100
I love your colourful channel. Cheers me up. 😊
it should be VERY well noted this is just a **concept** nothing has actually been objectivly created yet from it or if they have they have yet to publically say it.
Well they have a few showcases highlighting issues in their paper. So prototypes are surely already made.
Yes it has
Great if true. After all the hype and promises from the SD3 fiasco, I rather be skeptical. I will believe it when I see it and hope it can be locally installed.
Just a little tip... I'd put a "spill suppressor" on your green screen key to get rid of the green haze around the edges of the cutout. I'm mainly and after effects guy, so I use "advanced Spill suppressor set to "ultra." Not sure if premiere has the same effect.
this is done in OBS in realtime. I don't think it has that. But i could refine the settings a bit
And memory is the key to making a sequence of shots in a film. Wow. Cant wait enough for this.
To be honest, I am quite fed up with all the installation / configuration / combination hassle with ComfyUI. Yes, it's powerful, but so many extensions stop working after some time or are so complex that I can't remember a workflow after a couple of weeks. We need a NEW, easier way to do all this. And this model looks VERY promising!
this is fascinating and yes definitely the way things are heading
always exciting to see new stuff coming up. That being said, if it needs to run a LLM on top of the Image gen, I fear my trusty 2080 won't do the trick anymore, maybe 4090s and the likes.
Also I'm a bit skeptical overall, because while it's great to have everything unified and easy to use, it means less tweaking, less personalization etc. So unless the model can be fine tuned like SD and Flux, we'll be losing quite a lot of possibilities.
Also Control net: While I'd be happy to just tell the AI what I want, sometimes putting it in text isn't the easiest solution and a proper control net would do the trick better possibly.
In any case I hope we'll see the results soon and I'm looking forward to it
I mean, the 2080 was never going to be relevant forever. No tech is, nor should be expected to.
OmniGen seems like it is on to something special. Language is a better unifier than extra layers.
It doesn't have to be so extra special in my books, imagine XL/Pony with a better text encoder for fluent language... Could be available pretty soon instead of far away with a lot of challenges. Flux is great, but the compute it requires is far ahead of consumer gpu market.
Great video, thx man!
Wow! That sounds promising!
Interesting and it certainly sounds promising but as usual, it doesn't mean much until it's actually out and usable. Cheers for bringing it to attention though.
Once it's released, it will need 4 RTX4090 to run locally.
seems like they are taking the next step, seems impressive
The model is out waiting for the real review
We are in pre-history of image generation. Prompt gives random results, and we can only generate again. It is great to get something from nothing, but to get something we want is still tedious or impossible.
if this is prehistory - please go to time of windows paint only .. ok? Delusional comment
That's true, theres a lot of work still with image generation even if we are focusing into video.
The concept sounds good, but it still has to exist, run locally on consumer hardware and follow a prompt at least as faithfully as Flux for me to get excited about it.
I want an image generator with an iterative natural language interface: "give me an elf on the grass"
Ok
"Pointier ears"
Ok
"Now he is smoking a bong"
Ok
Sort of like iterative inpaint
Patiently waiting for that model that can generate my film scripts without needing to mess with files, folders, and settings. So exhausting.
Great video, i am hyped. Are you planning on releasing another video about your life as an AI Artist? Loved the "A Personal Journey between Photography and Life" video.
i'm thinking about it. I actually wanted to make more videos like it, but didn't release any yet
Sounds very interesting
They are removing control net, it sound more like they are just adding it into the distro version so in fact it is a much fatter code because there is no opt out to load certain models and extensions.
Hi Olivio! Your videos are excellent.
Can you make new video about OmniGen for ComfyUI with details about workflow for ComfyUI?
great, lets see, how good it will be , I paid for Kling but now usinig MiniMax,
First, I love that you called it a FLUX killer instead of the 1000th MJ-Killer 😂Second, according to the papers, the Model will be smaller than SD3 (3.8B params vs 4.7B params), so I would guess it should be running on any machine that can run SD3, or are the Params not the deciding factor in this matter? I have to admit I am not that technically versed... :(
its build on sdxl model so dont get THAT excited, image quality wont be better than flux
@@solomslls from what i understand is that they used the sdxl vae in testing, but that's not the end goal
Awesome news... Like Musk's flight to Mars..... Looking forward to the coming
Intelligent Simplicity as the new Normal - bad times for Comfy-UI🙂
Lets wait before there is more concrete information and progress to see. I don't see this taking off out of nothing, they are in an early phase, and even if they succeed in unifying there are a lot of question marks regarding requirements. The bar they have for this project is so high that I don't think this will hit the consumer market. Flux is already almost out of reach for a big chunk of consumer hardware.
WOW, big if true! 😁
Isn’t it what open AI already showed in their demonstrations like a half a year ago?
This sounds like something I would really like to try out. Does it give you for example put a sheathed katana on a characters back? This is something image generators can rarely do.
Depends if we will could run it locally.
Fingers Crossed it runs locally.
The "chain of thought" concept makes me think of the last improvement in chatGPT-o1. But I guess it's different here, OmniGen is not gonna be like a SegMoe model with many "experts" in their own field brainstorming about your request.. right ? ^^
My thought is that having everything separated, the controlnets the loras etc might be difficult to maintain, but maybe that's why everyone can run it. Now, one big AI model that can do everything... How much that's gonna weight, I'm really worried that my 3090 recently purchased get already deprecated, after only two months of use x)
this look super fast training model ,so you upload 2 images, and you have the lora
Hi Olivio, I've noticed a bit of an oddity on your RUclips channel. There seems to be a "disproportion" of sorts. On one side, there's a hefty number of views and likes, and on the other, a barrage of negative comments. So, here's my burning question: Are you pulling some digital wizardry to game RUclips, or is this just a quirky glitch in the matrix? Can we anticipate a flood of positivity washing over your comments section soon? Fingers crossed for an algorithmic miracle! ✌
i need to start to ban the salty people again
This sounds like an excellent move towards a better workflow. Will it do scenes and prople ot locked in at prople only??
Only me hearing OmniChan? :P
VRAM Requirements: 40GB RAM
actually these models are siblings while one trained in generating images the other trained in generating text, so why not try to train in both data
You can't play around with source code! You need the model weights for that.
So, you need an Excel sheet to keep track, but not anymore with OmniGen. That means that the new model is worth one Excel sheet. OK.
Interesting to know ... but I think no one cares until it actually goes out for real ! Remember all the hype for SD3 ? now it's in the toilet and the community flushed-it like it was shit
I still use google.
MJ killer?!?!?!
the whole "GEN" space will change .. what we need is another pandemic lockdown 2028 perhaps 😆
It is too slow
No...
One question: when?! :D
Half of the comment section: Now? No? Buuhh!
lol
bullshit