AI-Imager Midjourney v5 Stuns With Photorealistic Images (arstechnica.com) 56
Midjourney announced version 5 of its commercial AI image-synthesis service, which can produce photorealistic images at a quality level that some AI art fans are calling creepy and "too perfect." From a report: Midjourney v5 is available now as an alpha test for customers who subscribe to the Midjourney service, which is available through Discord. "MJ v5 currently feels to me like finally getting glasses after ignoring bad eyesight for a little bit too long," said Julie Wieland, a graphic designer who often shares her Midjourney creations on Twitter. "Suddenly you see everything in 4k, it feels weirdly overwhelming but also amazing."
Wieland shared some of her Midjourney v5 generations with Ars Technica [images shared in the linked story], and they certainly show a progression in image detail since Midjourney first arrived in March 2022. Version 3 debuted in August, and version 4 debuted in November. Midjourney works similarly to image synthesizers like Stable Diffusion and DALL-E in that it generates images based on text descriptions called "prompts" using an AI model trained on millions of works of human-made art. Recently, Midjourney was at the heart of a copyright controversy regarding a comic book that used earlier versions of the service.
Wieland shared some of her Midjourney v5 generations with Ars Technica [images shared in the linked story], and they certainly show a progression in image detail since Midjourney first arrived in March 2022. Version 3 debuted in August, and version 4 debuted in November. Midjourney works similarly to image synthesizers like Stable Diffusion and DALL-E in that it generates images based on text descriptions called "prompts" using an AI model trained on millions of works of human-made art. Recently, Midjourney was at the heart of a copyright controversy regarding a comic book that used earlier versions of the service.
Free alternative (Score:2)
URPM and other models hosted on civit.ai can achieve these levels of realism, or almost as good.
Re: (Score:3)
Yeah, I've been getting better stuff than this with even stock SD models for months. And now we have ControlNet to allow for complete control over scenes.
There's supposed to be a high-parameter model coming out soon from Stability, which will be interesting. Can't run on less than 24GB VRAM. It's said to be able to spell without even requiring ControlNet. Can't wait to try it!
Re: (Score:2)
To me that "barbarian beside a TV set" looked like a plastic doll. The others were sufficiently good that I wouldn't have questioned them.
Re: (Score:2)
Re: (Score:2)
Oh wow, you weren't kidding about URPM. Those are remarkable results even without face fixing.
Re: (Score:2)
I only wish I had an upvote for this comment.
Time to pack bags and head to the wilderness... (Score:5, Insightful)
As an ageing greybeard, I'm not sure I can cope with this barrage of Machine learning advancement anymore.
I've tried to stay up to date, but I actively fear what is happening in tech now.
We're already in a place where it is close to impossible to have any trust in what we are told, unless we witness it with our own eyes.
Arguably, we've always been there and it's just a case of picking a version of truth via multiple sources that are similar - assuming you bother to DYOR.
We've had hyper-realistic images of humans for years - heck, since the 1960's and the super realism art movement, but that required extreme skill.
Now, anyone with access to software, can generate photographs of humans so convincing, they look real - even on second, third or fourth look.
Video will be next. Again, with video, the ability to doctor video has been around pretty much since its inception - if you were skilled.
Soon, anyone will be able to do it via a text description.
Music, art, homework, coding - you name it - all rapidly heading toward the fingertips of anyone, to create instantly.
There will be no truth. There will be no privacy. There will be no skills left for humans to excel at - what's the point, when it can be generated in seconds?
Head to the wilderness, just switch it all off. The wilderness of your own mind, away from all of this - throw out the damn computers.
I'm just glad I've managed to hit my 50's with gainful employment as a skilled worker - I think it highly unlikely I will have a job in my industry a decade from now.
Time to plan an exit...
Re: (Score:3)
Even so, still worried. When it comes down to it, how 'realistic' or well-produced something is weighs heavily in h
Re: (Score:2)
Be happy that for now, AI generative tools suck at video coherence.
Be unhappy that lots of people (myself included) are working to change that situation.
no concerns (Score:1)
Re: (Score:2)
Welcome to the Free AI movement :) On the image front, StableDiffusion is the champion, hands down. Free, designed to run on consumer hardware, minimal license restrictions (can be basically summed up as "do no evil, and don't sue us"), and a huge open source community designing around it.
On LLMs, it's a more mixed bag. LLaMA is clearly the best "open" model out there, definitely a rival to the closed GPT series, but while it's released and you can run it (and even custom train it), there are license res
Re: Time to pack bags and head to the wilderness.. (Score:1)
Re: (Score:2)
The challenge is that doesn't work so well because most of the most important imagery would explicitly *not* be endorsed by the subject matter. So random citizen snaps a video of an important person knocking over someone in a wheelchair, and the best you can hope for is that you can prove that the random citizen vouches for the picture. It's not like the important person is going to sign it.
Now you could have camera equipment sign the pictures to prove raw, but you can have an analog hole to start having
Re: (Score:2)
Re: (Score:2)
There will be no truth. There will be no privacy. There will be no skills left for humans to excel at - what's the point, when it can be generated in seconds?
Easy there, geriatric doomsayer. For people who do graphic arts or art in general to make money, this will be a boon. Less time spent for more output. Learning to get the exact thing you're imaging from the AI output will be the new art class. That's not the end of the world, that's just the future.
Also, nobody is forcing people to use this. If you do art, and you enjoy it, keep doing art. I do 3D art and I actually LOVE Midjourney because it's free concept art or inspiration for me to work from without
Re: (Score:2)
Easy there, geriatric doomsayer. For people who do graphic arts or art in general to make money, this will be a boon. Less time spent for more output. Learning to get the exact thing you're imaging from the AI output will be the new art class. That's not the end of the world, that's just the future.
And in that future, your clients will still be willing to pay you the same rate for generative art that they are paying you now? I'm not so sure about that.
Re: (Score:3)
For people who do graphic arts or art in general to make money, this will be a boon.
Will it? I've just tinkered with it a little bit, and my poor, design-challenged brain got somethings generated that are really snappy.
The problem is, most design jobs don't need "great" - they need "good enough". And all that work just vanishes.
Re: (Score:2)
Well hopefully you're right and this phase is largely transitory. After all, the industrial revolution did enable the masses to join the artisan class. Though it was also a bit of a mess for workers rights, political corruption, and corporate monopolization. Sure sounds familiar...
However, there are some major concerns I still have in the long-term. Namely the potential for "outsourcing" and the gradual brain-drain of the creative industry. I can't help but look at what happened to the manufacturing industr
Re: (Score:3)
We're already in a place where it is close to impossible to have any trust in what we are told, unless we witness it with our own eyes.
we have been in that place forever. everything the media (any media) has been throwing at you during your half a century of life has been carefully calculated interest and bias, if not blunt propaganda.
this ai revolution will only change how we go about it, these are new and different tools and will reshape the labor landscape and consumer habits, they will allow to expand the limits of art and media, they are already starting to produce hilarious upheaval in courts by ip trolls of every nature, but the per
Re: (Score:3)
Re: (Score:1)
Ted Kaczynski had the same thoughts, but in 1995.
Re: (Score:3, Interesting)
Hit your 50s have you? I've got a couple of decades on you whippersnapper. (I suppose people from your generation would say 'grasshopper' rather than 'whippersnapper' or is Kung Fu with David Carradine too old for you?) I've been worrying about the future since the 1950s when we were learning about fallout shelters in grade school.
First of all, nobody knows the future. We could be headed for the mother of all disasters, or we may learn how to deal with this stuff. Partly it's a matter of luck, and I th
Towards a New Way of Thinking/Sanity (Score:2)
"I'm just hoping that at some point people will get tired of crazy and want a return to sanity. But that's a hope, not a prediction."
Maybe it won't exactly be a return but a transcendence? Even as from one perspective Hunter/gatherers were "The Original Affluent Society" (as Marshall Sahlins suggests) so there is an aspect of a return too.
My thoughts on all that:
https://pdfernhout.net/recogni... [pdfernhout.net]
"Despite probably having more computing power per square foot than any other place in the world, they seem not to
Re: (Score:2)
Doomspeak. That is all I hear.
But literally: "Now, anyone with access to software, can generate photographs of humans so convincing, they look real - even on second, third or fourth look." Good for this. Now everyone can be an artist. No longer is the art pool limited to just those with the time and money to become an "artist". Anybody can do it. Good. Awesome even, as now the choices we will have should explode. Where there were only one or two you might like there will now be thousands. This is Go
Re: (Score:2)
The problem is that when you make it easy to cheat and be lazy, everyone will be a cheater. You'll have no choice if you want to compete. In the end, however, you devalue everything and end up with a low-trust society with massive wealth disparity, which has been a steadily increasing problem since the 80's and is only continuing to accelerate.
It certainly will be a doomsday scenario unless we, as a species, can evolve to a new level of society where we don't need to "compete" just to put food on the tabl
Web interface (Score:2)
The interface for midjourney seems really weird, through discord and in a mess of channels. For 60 dollar a month you can pay them not to post the creations you asked the bot for in a public channel...
Maybe they can ask GPT-4 to build them a proper web interface.
Where is the problem? (Score:2, Troll)
It is clear that Artificial Ignorance is rather limited. "Too perfect" is a serious defect in a picture. So what?
Re: (Score:2)
FWIW, "too perfect" is rather easy to fix. You can even choose what kind of noise you want to add.
Re: (Score:2)
As far as I understand the art of photography (not very well), the imperfections are actually what marks a master. "Noise" does not cut it.
Re: (Score:2)
Well, it depends on what they mean by too perfect. Certainly I'd call some of the images more realistic than is plausible (better contrast between foreground and background) and a certain amount of noise could fix that. It would slightly soften the edges and decrease the contrast. The expert touch would be deciding what kind of noise and just how much, and where to apply it.
OTOH, if the goal were to make the foreground image stand out more, you'd just want to apply the noise to the background. Or you co
Re: (Score:2)
. What you should be thinking of here is not how an expert photographer would proceed, but rather how an expert painter would proceed, since that what this kind of post-processing most resembles.
Agreed. And that is something AI cannot do. At best it can copy and then it will not be a real match for the picture in question.
Hmmm, all this AI (Score:3)
There's some merit to saying they're "too realistic" for my uses since I'm making concept art for 3D models and very surreal and abstract stuff tends to be better for me, but for on-demand stock footage that's exactly matching your theme, colors, subject - this could be amazing.
Re: Hmmm, all this AI (Score:1)
I was in the same place as you.. Avoided chat but had tonnes of fun with images. Then I got to use the ai enhanced Bing search and it's fantastic. It carries a conversation quite well, so you can ask follow-up questions.
I have used it to plan my garden (when should I plant xyx seed) , to identify programming solutions (even going as far as, can you modify this code to make use of the new api level) and while it's not perfect, it's quite fantastic. It asks questions to narrow the answer (ie, where do you li
ChatGPT's comments on this (Score:5, Interesting)
please write a critical review of midjourney v5 image generating AI, in the form of an iambic pentameter poem
Midjourney V5, an AI for art,
That generates images with a start,
Its pixels dance and colors blend,
A visual world, it does extend.
But in the realm of art and form,
Is this machine enough to charm?
Can it surpass the human mind,
And create art that's truly refined?
While the AI can surely try,
And its creations catch the eye,
Its limitations still remain,
For art requires a human brain.
A machine lacks soul and emotion,
And can't imbue its work with devotion,
It lacks the passion and the fire,
That elevates art to inspire.
Midjourney V5, a step ahead,
But can't replace what's human bred,
Its images may be a delight,
But art still needs a human's sight.
Re: (Score:3)
I know it's not significant in the grand scheme of things, but this impressed me the most:
"But in the realm of art and form,
Is this machine enough to charm?"
That willingness to give slightly on the rhyming as a sacrifice to maintain the cohesion is notable.
Re: (Score:2)
Nice catch.
Re: (Score:3)
Honestly that poem has more soul, emotion, and demonstrated intelligence in its art than most words 99.9% of Slashdot can come up it.
I'm sold. When the uprising starts I'll fight for the machines. They seem to be more human than the rest of us.
Re: (Score:2)
Honestly that poem has more soul, emotion, and demonstrated intelligence in its art than most words 99.9% of Slashdot can come up it.
I'm sold. When the uprising starts I'll fight for the machines. They seem to be more human than the rest of us.
I don't disagree with you on this.
And, if I'm honest, this is the first tech I have some apprehension about.
Re: (Score:2)
Max Headroom - "Trojan Sheep" (Score:1)
Beware of Trojan Sheep [tvtropes.org], they may not be what they appear to be.
[the bad guy] refuses to budge but the team successfully creates a fake vid
This was aired back in 1987.
Yepp. And this is just the start. (Score:3)
This is ZX81 level ML/AI stuff. (ZX81 being the first viable cheap kids homecomputer back in 1981). And given that tech innovated by tech doesn't grow linear but logarithmically until it hits some hard physical barrier for the site of artificial brains, we are bound to see some areas crossing the famous singularity within the next few years.
Imagine midway 7, ChatGPT9 and some image and voice recording software looking at the world 100000 new images and sounds per seconds from all over the place and being orders of magnitude smarter than any human within a few days.
We're just about there. Sh*t is about to get veneers crazy.
Re: (Score:2)
There are some fundamental problems to solve.
How can you train it to mull on a problem through thought experiment? (Without a human holding its hand in a query session) How do you train it to apply things like method of induction or divide and conquer techniques for things which can't just be solved in one pass with some output token feedback? Deep thought doesn't really suit any existing training algorithms.
They know how to make it larger, they are improving fine tuning ... but that can only get them so fa
Re: (Score:2)
You're focusing too narrowly. Yesterday (or the day before) there was a story about an F16 (I think it was) that took off, played dog-fight, and landed under AI control. Amazon wants to automate their warehouses to avoid labor problems. And MS has announced coupling ChatGPT to robot bodies.
Clearly current ChatGPT isn't quite good enough for the proposed uses. Also it clearly doesn't need too much more work. (It may not be the occipital cortex of the robot, but it can be the analog of the cerebellum...e
Re: (Score:2)
And given that tech innovated by tech doesn't grow linear but logarithmically
Eh, you probably mean exponentially. But I'm not sure what you mean by "tech innovated by tech". As far as I can see new tech is still created by humans. I haven't (yet) seen AI coming up with innovations or new discoveries, so we don't have a positive feedback loop where a machine builds a better machine who builds an even better machine, and so on until we end up with something humans wouldn't even be able to conceive, much less understand.
An important factor in the growth of machines/AIs is however inher
Re: (Score:2)
i know not what tools gpt5 will be written with, but gpt6 will be scratched into the earth with pointed sticks
Re:Yepp. And this is just the start. (Score:4, Insightful)
Imagine midway 7, ChatGPT9 and some image and voice recording software looking at the world 100000 new images and sounds per seconds from all over the place and being orders of magnitude smarter than any human within a few days.
The sophistication of these models is a bit of an illusion. They are recapitulating their inputs, and for a huge corpus of training data it can start to feel uncanny, but they don't possess any real reasoning capacity.
Try asking ChatGPT a moderately simple arithmetic problem, like 'What is the result of 2345*329?' It can't do it. It's seen and so been trained on millions of arithmetic problems but it has no capacity to do math. It will probably stick a 5 in the ones place because it's seen that in many ..5 *...9 problems, 5 is the last digit of the answer, but it can't figure out why.* But if you ask it something people post on the internet a lot like what 2+2 is, it can tell you that. It's a google search with some extra composition.
My bet is we will see some very cool developments over the next few years and then we'll just get used to having some smarter algorithms same as we're already used to the existing clever algorithms. Magic to someone from twenty years ago, but doesn't continue to impress as you go from awe at the novelty to shaking your fist that its inherent limitations are preventing you from getting work done.
*This example will be a bit ephemeral as I don't doubt someone will plugin a separate piece to offload math calculations soon. But the observation will be the same.
Are we scared of the right thing? (Score:5, Insightful)
I see a lot of people being scared of this being abused to incriminate people that are innocent.
Do you know what I think is way more scary? Bad people now being able to do their thing with impunity now because it doesn't matter anymore if someone takes a picture of it... the mere existence of this AI lets them claim deepfake.
Granted, we're not quite there yet. AI images and photoshop can still be identified in forensics but for how long?
Re: (Score:2)
We've had skilled photo faking (even before digital images) for a while. AI won't know how to replicate the hardware artifacts inherent to a particular lens and camera sensor, or many details of the person and scene and lighting, and will bear its own subtle markings in the generated images. Cameras could start including steganographic encoding or otherwise cryptographically sign their images. Courts will be plenty happy to have some real human testifying that 'yes, this image came from physical camera X
Re: (Score:2)
So why couldn't a sufficiently-prepared AI include the same metadata, and make it specific to the created image?
After all the camera and software are doing exactly that; it's not human-inserted data.
Amount of control (Score:2)
I've played with a few different companies' text to image systems, trying to render some specific images (think classic book cover illustrations or similar) and find I just can't get enough control to get output like I want. In some cases it's fine to have it produce something that is 95% out of your control, and when you throw that online on its own without context or giving an idea of what you were attempting to create, it may still look amazing and impressive. However it isn't what you really wanted. Ta
Re: (Score:2)
Well, it's your lucky day, because now with controlnet and stable diffusion distribution you can get really astoundingly fine control. You need to use inpainting and/or outpainting to get very specific results, though. That's mandatory. Controlnet is an absolute miracle, it lets you use sample images to control poses, colors, sizes, etc. You can use it with inpainting to solve the hand problem. Hopefully someone will get that worked out better in SD in the future. Clearly MidJourney has got it pretty well n
Re: (Score:2)
There's a new thing just released that is amazing: Controlnet. It has various processing algorithms you can load that give you a LOT of control over the image generation process.
https://www.youtube.com/watch?... [youtube.com]