This AI image generator does EVERYTHING
ฝัง
- เผยแพร่เมื่อ 20 ก.ย. 2024
- OmniGen can create & edit images with natural language. No more controlnet, loras, inpaint.
#ainews #ai #agi #singularity
TurboType helps you type faster with keyboard shortcuts. Use it for FREE:
www.turbotype....
OmniGen: arxiv.org/pdf/...
Newsletter: aisearch.subst...
Find AI tools & jobs: ai-search.io/
Support: ko-fi.com/aise...
Here's my equipment, in case you're wondering:
Dell Precision 5690: www.dell.com/e...
GPU: Nvidia RTX 5000 Ada nvda.ws/3zfqGqS
Mouse/Keyboard: ALOGIC Echelon bit.ly/alogic-...
Mic: Shure SM7B amzn.to/3DErjt1
Audio interface: Scarlett Solo amzn.to/3qELMeu
TurboType helps you type faster with keyboard shortcuts. Use it for FREE:
www.turbotype.app/
👋 hi
Why are doomed. It has been a good one guys 🫡
You can keep reviewing papers while advertising TurboType, or you can be more informative. It’s your choice.
This is perfect for someone who want to create games, visual novels, comics and manga without spending tons of time learning and practicing the tools, I'd be excited to give it a try when it's out
likewise!
Sound great but people will be shocked and reject it when they find out it was created by AI.
@@wwk279nowadays everything is generated using AI!
@@wwk279 overtime i think people will get used to Ai genereted content as it gets better and better
What if this paper is itself generated by AI 😂 @@truelies5431
Imagine next year when this level of control is available for AI video.
Imagine the year after next. OMG it is going to be so great
I can not wait
@@robertmartens7839what about year after "after next one" what's gonna happen then?
Next year would be too optimistic. AI can barely generate 5 second videos and takes forever to create them. My bet is 3 years if we are optimistic.
@@elon-69-musk really right!
Some of the early image gens had bad prompt adherence and you just glossed over that part. For example, a blonde woman was asked for and a brunette (dirty blonde at best) was generated. She also had no clothes compared to the "minimally clothed" that was asked for, background was not magenta. You said the prompt was "indeed what we get"
I agree, this looks like a weak version of ComfyUI using ControlNet, which AI Search is aware of exists. Good and specific images take effort and time. There are multiple tools where you can just use one image of a character and you can generate images with their face without issues. ControlNet is king right now.
ok dude we all saw that. don't be picky.
@@cyberprompt It's not about being picky. It's not a useful tool if it doesn't generate what you prompt for. This is a big issue for people working with this stuff.
The point of this technology is not really prompt adherence. It will be made OpenSource, so the abilities detailed in the video will (probably very soon, it all goes so fast) be attached to generation tools with better prompt adherence like Ideogram for example.
It also messed up the the second image in the iron man prompt.
That's a monkey like samurai character , but in the output made him very human looking.
Its just a paper dude. Its not real until we can use it.
The paper shows that it's becoming real. Because if they can do it ijn a paper, then it can be done without a paper.
Yep, this video is nothing
@@Corteum "PAPER SHOWS" doesn't show shit, where can I use it?
@@ignoreme1141 Yeah but most of these things begin wth papers. right? that's how it works
Sadly, that's the truth. There haven't been one or two duds that promise great things, and deliver none.
This is a staggering upgrade to AI image generation. We are on year 2 of the OpenAI public era. I can not imagine what comes next, let alone, the next year!
exponential growth!
I've been telling you all for 4 years now
@@TheThetruthmaster1 be honest your prediction just happened to come true lol
I feel this is a dream I've been having for years about to come true. I can only hope this will come out with a free limited version for those of us who crave creation but truly can't afford anything...
i hope so too!
You died a very long time ago old friend... Read Codex of the Celestial Dream 🙏
@@NakedSageAstrology Please be careful what you say to people regarding death. I'm struggling with depression and almost killed myself recently. It has the potential to create the opposite effect you're aiming for.
@@iloveyoutoohuman it's ok human the robots will takeover soon
@@iloveyoutoohuman stay strong!
OmniGen's ability to simplify complex image generation is a huge step forward. Seeing it in action shows just how intuitive AI tools can become for creative professionals!
I believe this is what the omni-modality of GPT-4o has to be. The ability to chat with a model that inherently can create images.
Now that we have this, it really makes me wonder why the GPT-4o we're getting is still nerfed.
Now that you mention it, I had totally forgotten about 4o's multimodal features. In theory, 4o should be able to do what this paper claims as well. It's strange that it still resorts to DALLE for images instead of using it's native image capabilities. Maybe they still need to figure out the guardrails
what's your daddy do for living? he's professional design graphic....
teacher: (silent and smile with burrowing her eyebrow)
When SD first came out, I went all in, first to just understand the concepts, then learn all the extras. Now I'm waiting. There have been so many advances and competition for what is "best" I don't want to burn myself out while the winner is decided. Or max out my hard drive with unnecessary tools and models.
I feel the same! There's just so much going on
it is exhausting
This is how JavaScript web developers feel when there's always a new trend
So ChatGPT but images.
basically
it's Chat SD!
This will be extremely useful for everything from forensics to TH-cam thumbnails. I can't wait for this to come out!
Doesn't show me an image of where is your current location, it sucks
???
open source devs need to step up, i'm tired of chinese research papers and stuff like this, but not an actual tool i can actually use for free and locally
Release date?
Consider this your cheat sheet for applying the video's advice:
1. Explore OmniGen as a user-friendly alternative to traditional AI image editing tools.
2. Communicate your desired image edits to OmniGen using natural language prompts.
3. Experiment with OmniGen's ability to add objects, change colors, adjust poses, and generate depth maps.
4. Understand the power of OmniGen's unified architecture, combining a VAE, LLM, and diffusion model.
5. Leverage OmniGen's emergent abilities for multi-step editing and in-context learning.
6. Start with simple edits and progressively increase complexity to maximize your results with OmniGen.
7. Be aware of OmniGen's limitations, such as prompt sensitivity and occasional struggles with hands and fingers, and unfamiliar image types.
8. Stay informed about the evolving capabilities and future developments of OmniGen.
wow... this is definitely would be my bucket list for testing...
can't wait for it to be released!
The proof is gonna be in the pudding (or not...)
I need this thing. This is beyond amazing.
transformers are one of the coolest technologies to have been invented recently - and to think thankfully google released the attention is all you need paper in 2017. Its quite a general architecture
Sounds really cool. Looking forward to it's eventual release.
i can't wait to try it out!
Mark my words, people will use this to ask the AI to strip naked people by providing a photo.
Twitter disinformation just levelled up
This is what I've been waiting for! I think other image generators will start having problems soon.
There was actually something like this back in SD1.5 era. I've been trying to find it again, but I believe it had the word llama in it. It was injected into specific layers like LoRa and controlnet, but it was an LLM that could comprehend all these things in the image and let you command it using natural language instead of prompting.
I probably wouldn't use these as the quality looks poor compared to a well trained LoRa on Flux (esp the wu kong example which probably wasn't already in the base model like bill gates would be), and you have much more control over composition with something like latent coupling, but hopefully future models use these techniques
Oh interesting. Thanks for sharing!
Llava was the captioner
@@SearchingForSounds that's a vision llm , but this was an A1111 extension that was primarily an embedding into an AI art model, the llm was just controlling it. also long before we had multimodal llms
I don't understand why but you always manage to amaze me and make me think ‘as if something like this already exists it's so incredibly fascinating!’ your videos are always the best and I love it when you always show something new better than the old which always beats the old by a lot I LOVE THAT ABOUT YOUR VIDEOS! Youre my fav ai youtuber!
Wow, thank you!
Have you tried using this app? Because to me it looks like a lot of these examples are fakes.
Amazing man , I think your way of explaining and enjoyable storyteller 🎉
it won't be long before you can specify anything. we already have video. next is to create a persona for companionship. it's always that direction.
I can say or fact that currnt AI doesn't suck.
Very intriguing. Although given the current state of AI development and previous heartbreaks, I'm practicing a healthy dose of skepticism. Hopefully when we get some hands on experience with it, the model holds true to the article's claims.
there are as many outrageous yet still true claims as grifts in the AI space which 1 what makes the grifts believable and 2 what makes AI so special, the fact that apparent scams are real technology
Apparently this generator did you too.
Holy #! Can't wait to get my hands on this once it's released.
Bro just went "aww using control net is so hard and annoying, it adds extra steps" then went: "so this AI has control net prepackaged so you can do the exact same thing but it's this AI so it's better"
I look forward to conversational AI merging with these models so we can just chat back and forth with the generators and tweak our images together into a final image. I'm sad the conversational Pi chatbot was abandoned.
Hopefully OpenAI and Apple release their models to the masses.
Pi just got a huge upgrade and he is faster and smarter than ever 😊😊
i will suggest my group to check out you im impressed
If adobe finds out about this one, they're getting shut down, it's adobe's motto after all, end the competition by buying them
❤ love the progress
Please raise the loudness of audio when you edit the video.
Thank you for your excellent programs as always.
You're welcome!
How you always stay updated 😂😂 , good job man , i apriciate you hard work ❤❤❤❤❤❤❤
Thanks!
Wait I was looking at mechanical turk jobs a little bit ago, and one of the jobs was to do what this image bot does - identify things like 'where can i wash my hands' or 'what can hold water'. Interesting.
this might be peak ai image generating
one ring to dominate them all!
😃
They are simply using a language model to call various tools like the IPAdapter for that ironman example or depth, canny, and open pose models.
An excellent work, but nothing extraordinary. You can do all of that already in A1111 or Comfy
No, it's an LLM that's specifically trained to respond with images based on input images, similar to how text-based LLMs like gpt4o respond with unique text based on your input. It's not a collection of different tools working together behind the scenes. It's an LLM that accepts both images and text as input, but unlike gpt4o which only outputs text, this one outputs images.
Nuts! Imagine the possibilities
Release it an make it open without restrictions and I am hyped
This will be good for architectural renderings. Say goodbye to overpriced rendering studios.
I'm pretty sure this is Adobe's vision for Photoshop Ai
But they are so far off being this capable it's laughable. This new Ai is going to make so many peoples tasks so much easier
Surfing the vibe wave all night, every night 🤙
As long as it is unrestricted.
a trully intelligent AI image generator :3
Seems too good to be true
Seems too good to be true. Wouldn't be the first time a chinese paper claimed it can do things it can't
This is the similar methods used by udio and suno, they learn from reference audio plus text description
Unfortunately many papers just papers and we have nothing to use years after it
looks like it understands the general physics/objects in the image and can construct the prompt then just mixing it up like stable diffusion, some martial arts or gymnastic pics would be nice. Strange body positions or interactions are a huge problem with the image creators from beginning the year. Or just human object interactions, like a female woodworker climbs a tree with a belt full of working tools in the early morning hours, watched by a curious crow sitting on a nearby branch.
Thanks for your share! in personally mind i don't think this model will become dominator in image generator because it seems can't supply ability for our job in fine-grained like ControlNet or IPAdapter, particularly in consistent images workflows.
Trying to replace LoRAs with one-image thing is certainly convenient if you just want to create low-effort memes to share with friends. But you are inherently giving the AI less data and details, and will almost always result in a lower quality result. This is still an impressive model and I appreciate that they are trying to package everything into a user-friendly interface with a fast workflow, but "There's no longer any need to train LoRAs" is hyperbole bordering on misinformation.
looking forward to this, if the model actually does come out...free
There's a great tradition of unbelievable papers promising things that never matetialize. This "review" is like preordering a game based on a cinematic. It's a great way to get scammed.
So when this ai becomes available will you make a video explaining how to use it?
definitely!
We will never see this be open source
In future we can program our babes
that's what i'm waiting for
thank you! any guess when this will come out?
would be interesting to see how it handles video generation.... asking it to generate the next frame, do this 1'000 times to get a 40 second video 🙂
It might be the end of flux when this is released.😁
I just wanted to take a moment to express my gratitude for the amazing video you created on AI Your dedication and creativity truly shine through, and it’s clear how much effort you put into making such valuable content. I know how hard it is to start and generate subscriptions. I watched the full video, because it was interesting. Please keep us updated on this Generator. As a new content creator what would you say is the best image to video generator to start to use.? I use Kling but it takes way too long, to be honest i dont really find it great
Thank you again for your incredible work. Keep up the fantastic content!
Thanks! For image to video, Kling is your best bet. Other options include Luma (worse quality) and Runway (expensive af)
@@theAIsearch Thank you i will keep watching you a new subscriber
Need this in comfyui asap...
Awsome dude, but Q: how the hell did you find this repo??? it's not really famous or anything???
I don't really remember - probably on X. There's so much crazy news everyday it's hard to keep up
Poor Bob! Stop giving ideas to his enemies!
Every day we get 1 step closer to the Holodeck. 😁
Yeah I just hope the final Holodeck wont be hampered by "ethic" and "moral" standards. It's the perfect playground for humanitys darker aspects, allowing us to separate it into a virtual space where it does no real harm.
it would be great if it generates directly PSDs with separate layers. A Designer then can work on the details, make changes easily. Adobe, are u listening?
Мы давно ждём такую штуку. + сеть должна понимать направление и положение в пространстве.
Me: Thank you Lord for the gifts we are going to receive
Also me: no one man should have all that power
So we can create a virtual influencer and make them pose differently by just giving 2 images one the influencer and make them pose in different way by giving another image reference
Damn social media will be flooded with ai influencer
That sounds possible
I don't know if you noticed. But there is already a ton.
@@ahtoshkaa there exists but this tool will make it even more easier so it’s gonna be even more worse
really cool ! That game changer, Thanks for looking up !
Amaizing work
Thanks!
lemme guess this is equal to llama 3 405b model in file size :D
Yet again we stride from precision tools to general. You will never be able to specify pose and character with words alone. This model is a toy.
When do y’all think we will have LEV, FDVR, AGI, ASI, maybe UBI and FALC
AGI within 3 years. UBI depends on the stupid govt
Dear god man i do like your videos but there is _SO_ much fluff. I dont need to hear you reading out every image prompt agonisingly word for word right off the screen, while _not even showing the corresponding image at the same time_ ! Man this is like pulling teeth, there's so much unnecessary repetition of the words you just sayid, barely rephrased, or "commentary" on an image where you basically just repeat the prompt again. Please for the love of god, stop describing and start giving actual meaningful insight. If you dont have any, _just show the picture_ and say nothing. Even on 2x speed this is unbearable 😞
Would you recommend me to buy RTX 4060 TI In open source AI applications? Or 3060? Because of the 12gb of vram❤
if it is true ... then so much easy to make consistent character
Shouldn't we be able to use an AI for ComfyUI? This way we should be able to use all the different tools with a single prompt even without OmniGen. The workflow json file could be AI generated as well and this should be no rocket science. Am I right? What do you think? I could start researching about this but maybe there already is something I don't know about.
cool idea. it sounds possible, since it's just json generation
This is fire
😃
I wonder if it will be possible to run it on a consumer level GPU
Well the 5000 series is going to launch this year so I mean, probably on such a beast the model should ruin
Now this is a game changer!!
Let's hope my 8g Vram potato PC can run it
That's not magenta
Thanks
why is it so recent?
would this model be "better "if it had a bigger parameter count???, what of they increase the number of images which it was trained on and problems as the hands or fingers, even the text are resolved???
yes, increasing the dataset would help a lot. the dataset they used was 10 times smaller than what SD used. imagine how good this would be if it was trained on more data
@@theAIsearch do you inow of the small parameter count would help it to run consumer level hardware and old GPUs?
Is this actually real? I guess we will see, if they actually release this model...
Impressive images but lots of red flags.
How did the inclusion of Jack Ma get past the local censors? That is not very politically correct these days.
Note that Beijing Academy of Artificial Intelligence is not an academic institution, but a company that has repeatedly released Shanzhai versions of large models.
I don't believe it until I test it 😜
Anybody knows of a model that would let us upload a pic of a person smiling and take the smile away so they're not smiling anymore, everything else remaining the same?
Face App for android / iOS
How to use it? What's the link?
GREAT HOW TO ACTUALLY USE IT?????
What about custom cartoon characters
in theory, a reference image is all it needs
While this is cool, you are extremely hyperbolic, specially about this replacing LoRAs.