What do you think? Has DALL-E 2 been DETHRONED? *My DISCORD Just got Revamped! Check it for the LATEST AI news! **discord.com/invite/bQgcbjs2Sg* ▼ Link(s) From Today’s Video: ✩ Gilbatree's Video (ft. me!) ruclips.net/video/EadcJIz-E48/видео.html&ab_channel=Glibatree ✩ Midjourney: www.midjourney.com/home/ ✩ DALL-E 2: labs.openai.com/
Three notes to add about MidJourney: 1) This is currently alpha-testing of V4 2) You can use --chaos and then a number from 1-100 to introduce more diversity in the results 3) You can use --q 2 to let MidJourney spend a little more time to potentially increase quality
dall-e 2 is definetly dethroned, although dall-e 2 is better at photographs, its still not very coherent. Midjourney is super good at digital art style with a lot of creativity, so i'd say midjourney is the best right now
Midjourney is amazing. I don’t think the detail is there when it comes to humans and certain other things though. I think of it kind of like my love for Nintendo, versus when PlayStation first came out.
I still prefer dall e 2 when making humanoid animals since Midjourney isn't that very good at making humanoid things. But mj is my second favorite now and hopefully in future version they can now male humanoid animals like what dall e 2 can
MJ was better before at times! And is the photo better? Seriously?)) And show me your photos, maybe they are on the websites of photographers, and so the photographers would not have guessed. I already have a lot of these made with the help of MJ, moreover, quite a few already have the status of a photo of the day. You need to know how to use. Of course, I finalized it in Photoshop, correcting small jambs, but there are just small ones.
There is a big problem with these type of comparisons: the same prompt does not necessarily lead to the same latent vectors in the two different software. Which can be intended in your comparison or unintended, let me explain. So on the one hand, a better comparison would be to figure out what extra text should be added to the DALLE2 prompt (for example "highly detailed, trending on artstation" etc) to generate the exact same style that Midjourney uses by default, then generate the pictures and compare them. It covers the use case where people are using the two software after reading some promp-guides. For example at the warlock example you could see that DALLE2 indeed can generate better than clipart quality (or not, but that would be the point of the comparison). It is completely possible that Midjourney generates its artistic style because either it adds "artistic style" to every prompt, or because the model is trained in a way that the text encoder makes the latent vector skewed towards the "artistic style" direction by default. You could test it by trying to generate explicitly non-artistic style renderings, and comparing those. On the other hand, the vast majority of users might not have knowledge about those prompt-guide tricks, and for them the default results will look better, and that comparison was covered in this video. A completely different thing: it seems that you have the problems with bigfoot, because the word "bigfoot" is not an entity in the text model (in Stable Diffusion it is CLIP). So the diffusion model generates "big foot", and now you have it... Nevertheless, I liked this video, it was informative, thanks for the effort. :)
Yeah, the Midjourney style was definitely artificially biased towards artistic images by the developers, while Dall-E 2 seems to be mostly unbiased in terms of style. Well, except in terms of OpenAI randomly adding "diverse" people (blacks, Asians) when you try to do create a person, and except for the fact they try to remove celebrities etc... The point is, Dall-E 2 manipulates the content, while Midjourney also manipulates the style. Which makes Midjourney look more impressive at first glance, but apparently prevents any photorealism.
we have now got to a point in the ai art generating space where dall e 2 is getting left behind by midjourney and stable diffusion, and only a few months ago, dall e mini felt like gold to us
@@evalover8622 true, but rn we havent heard anything about dall e 3, so i wouldnt expect it until mid next year, and by then sd and mj wouldve probably caught up to dall e 3 when it comes out cuz they r updating at a rapid pace
@@FRareDom just look at the difference between dalle and dalle 2, dalle 3 will be at the retina level, meaning at the average viewing distance at the average resolution, it will be very hard to distinguish from other photos
My favorite part of the new V4 algorithm is not the images themselves, but how much MidJourneys vocabulary expanded! I can finally create images in the style of Arcane, or concept art from Valorant
v4 is only in alpha right now, they're still working on more big improvements (better upscaling, different aspect ratios, overall image quality etc). The current state is already fantastic, but there's even more to come. You can nudge Midjourney into photorealism btw, when you're using the right prompts.
There is quite a big difference between midjourney v3 and v4 models. Last one creates exact more accurately what you asked in prompt. So, i think it's a killer of all existing txt2img generative AI at the current moment.
It’s definitely better than anything else out there right now, it’s capable of understanding and depicting ideas I couldn’t get before. It truly feels like there’s a lot less limits
Thanks for the review! I don't use DALL-E or MJ often, so this helps me stay current. I feel like MJ is best if you want something consistent, but it's also limited by that built-in painterly aesthetic. DALL-E has self-imposed limits that cripple txt2img, but the img2img, inpainting, and outpainting work great on an existing image. Still, I'm very happy I can run Stable Diffusion on my own PC. Over 88k generations and counting.
@@remylebeau2357 That has absolutely nothing to do with their idiotic decision to be on Discord alone. They could easily set up a normal web page, like any other business.
Just a helpful hint. You commented on the sameness of the "character concept, old mage warlock" images. But if you re-rolled, your would get a different family of images (for example, when I just tried it, mine had pointy hats, and three of them carried staffs), and if you use the --chaos parameter, you can get wider differences within the group of four (I tried it again with --chaos 90, and I got elven wizards, bareheaded ones, ones reading tomes, although, yeah, they were all still white-bearded white guys) --your friendly neighborhood MJ guide
Midjourney clearly uses invisible dynamic prompting. Ie; they are are auto-filling your prompts for you, but don't show you the things they added. It's very cool but i think they should either A. Give you the option to turn that on or off and generate very basic looking results or B. at least show you the secret sauce they are adding on for you to make things come out in their signature MJ style. I rate SD above them all tbh. Nothing beats the control, the power, the flexibility, i get that it's not very user friendly, but being able to train your models and create hypernetworks and have all this control over 'how' the AI interprets your prompt and so many other features, there's nothing else like it. SD is the Linux of the txt2img world, MJ is the Windows, and DALLE is the Mac. With the latter two, what you sacrifice in power and control, they make up for in user friendliness. So if you don't want the AI to spoonfeed your prompts for you and you want to have more independence in how you communicate with it and how it understands you, then you need to go for that technical proficiency and learn all the cool tricks of the trade. SD is made for that.
Sorry to have to say, but you are incorrect. It all simply comes down to the "brain" that the Ai gets implanted with. So basically the image database that are used to train the Ai, coupled with the algorithym that is paired to that dataset that dictates exactly how MJ is expected to interpret prompts. Simple as that. Each one of these psuedo-Ai's has it's own "personality" or "style" or lack there of in Dall-E's case. The most obvious and easy to spot being disco diffusion. No matter what prompts you feed into disco diffusion, it will always be easily identifiable as Disco diffusion simply because it has the heaviest hand in adding its signature style to everything it does. How do I know? I am a programmer and prompt engineer who has been working with Ai models for the past 4 years.
@@remylebeau2357 I understand how models and training datasets work, but that doesn't detract from my assertion at all, both can be true and i believe that to be the case. They are not mutually exclusive concepts.
it’s funny that Matt got more excited about the avocado chairs, but I thought the wizard and dinette images showed the greatest superiority of Midjourney over DALL-E 2!
Great video! very helpful and informative. Might be worth writing the prompts that you used in your description for people to copy and compare results :)
the AI models are changing for the better so quickly, I can imagine that in 10 years from now we will be able to ask an AI to generate a whole movie, using simple voice commands "I want an action film with actors X,Y,Z, they must do A,B & C" lol
I literally just bought a $4,000 computer because I am so impressed by this new AI stuff. I can't even grasp how amazing this is. I am in a non-stop state of shock. No, I am not rich.
I don’t blame you! Currently, I can do low resolution stable diffusion images on my six year old Alienware Area 51, but I am thinking of buying a new computer just so I can do AI art on it!
@@dialecticalmonist3405 How did you spend $4000 and get a low end CPU/GPU? Or is that your old rig? If you spent 4k on a 3600x / 1060 then you got absolutely scammed.
You can usually tweak the Midjourney output to be more photographic-like by spamming camera-related keywords. If the prompt doesn't specify anything related, it tends to default to painting-like. As an example, to your Bigfoot prompt, I was able to turn it into a realistic photograph by adding "selfie photograph, 35mm, kodak gold 200, 1980 Polaroid" in front of the prompt.
Midjourney absolutely blew DALL-E 2 in every prompt. No doubt about that. Basically in all the images it produced, none of the characters fell into the uncanny valley, all of them had correct faces, limbs, everything. DALL-E 2 seriously struggles in maintaining this coherency inside the image itself, when compared to MJ. Also, the diversity generated by the prompt, outside of the chaos parameter in MJ, partially happens because MJ doesn’t alter your prompt like DALL-E 2 does, by appending words like woman, black, and other ethnic descriptions automatically. So the prompts actually being compared were not even the same.
I don't use dall e, so I didn't realize it did that. Honestly, is it too much to ask to just leave my prompt alone and make what I asked? I agree with you that Midjourney won hands down.
Thanks for the detailed comparison. I’m not sure it’s really an ‘Apples to Apples’ comparison, so I wouldn’t go so far as to call it a dethroning. Dali has a bias towards photographic realism and in the case of the armchairs, they were much more useful designs for an avocado-style chair that people would actually want to sit on. Midjourney is as heart an illustrator and certainly got a massive jump in coherency now. But I wouldn’t want sit on those chairs at all, however avocadoey they might be, and if I were a chair designer, the Dali version would be more helpful. The same goes for the shitzu dogs, both look good to my eyes and Dali is realistic whereas Midjourney is incredibly creative. What will probably dethrone Dali is the release of Google’s Imagen and/or Dali 3, which will probably boost coherency and details to another level. This photographic realism is presumably one reason why they block depictions of real people, whereas Midjourney is in the land of the playfully creative and happy to be there. But yes when it came to things Dali can’t do (the waitress or Warlock) Midjourney certainly smoked it. But even so, I’d say, without really replacing what Dali would do, if it could.
You can increase the variation in MJ by modifying the chaos figure. To get more photographic results in MJ you have to use different prompt words, like calling out a specific camera, etc.
As an artist I’m both nervous and excited about this AI art stuff. Excited because of the things I will be able to create that I wasn’t able to before and nervous because I’m scared of being completely replaced by AI.
You've probably gotten this before, but I just realized... the way you speak kinda reminds me of Carl Sagan. You also look like you could be his grandson or something. Cool!
@@MetigArt How? i tried it but it said i need to wait 24hrs to deactivate my account, I'm so broke and i'm tired of creating hundreds of accounts just to get another 25 free image
I liked Midjourney more dan Dall-e 2 before but it seems to be distancing itself more and more midjourney is a lot better for me, I get way better results
The frog logos dall-e generated are definitely better. The midjourney ones would be unusable as logos due to their complexity, unless you want to look very unprofessional.
Isn't the AI "look" basically defined by what images it's trained on? Seems like Midjourney is trained on Artstation/DeviantArt, while Dall-E on Google Images (mostly photos). Most of Dall-E looks like someone scrapbooked together a bunch of Google Image photos.
for stable diffusion you have to train a model to get such results or write a big ass prompt essay about the thing with a big ass negative prompt.But you get tons of flexibility. If somethings not working img2img it,. 😅
I don't know what to think. It has amazing potential to make our dreams true, but it hurts real artists really bad. Makes you think how many jobs ai will replace in the future...
We should not comparing different AI with the same prompts. It's like comparing the female portrait created by the Stable Diffusion 1.5 with Waifu Diffusion 1.3. I can guarantee you that 99% is Waifu Diffusion win, because the model has heavy style in it. I think he should at least include different styles and artists in the prompts.
DALL-E 2 was dethroned many months ago when Midjourney came out. They haven't really made any serious gains, the images are still very poor quality, the in-painting maybe looks great to use but from a commercial point of view I'd never use DALL-E 2, I have lots of credits, never used them. Even Stable Diffusion is way better at quality and coherence, that is free to use also. Midjourney is night and day compared to the others, for $30 a month it has everything I need so far, I'll stick with it for the foreseeable.
I still refuse to pay for any of these AI Art generators except for one. That one you don’t cover or haven’t covered yet. It’s light years beyond all other AI tools in my opinion. Only time will tell though regarding which one is become the G.O.A.T.of AI Art tools.
That upscaled avocado chair should definitely calm down or it's going to get a NSFW warning lol. It's not that I don't like Midjourney, I'm an early member on their Discord and still have a few image credits, but guys... Is vignetting++ and stylized look mandatory ? :( Though DALL-E 2 never really evolved so it was the king of the game back in the days, but now I very rarelly use it personnally, and when I do, it's to get init images for Stable Diffusion img2img '^^
17:00 it clearly misunderstood the prompt - if they're playing video games, they're doing it on their iphones, so it didn't understand what an "iphone photo" was. Is that a good prompt for other engines? I've never heard anybody describe a photo as looking like an "iphone photo" - how would you tell? If the model hasn't learned that, that's probably a good thing?
The prompt was intended to be extremely difficult for the AI models. I’m sure you could get a better image for the same idea, but I want to see how well the difference ai interpret the prompt. It clearly shows their differences
Is this getting to be like a regular Google search? Can you type in your (anyone’s) name? Instead of saying “Google yourself!” could you say “Prompt yourself!”.? Or is it just famous people? Maybe you could see if your pictures come up, as you have pictures on RUclips and such..Maybe we are getting exceptionally close to making our Avatar from pictures of us gathered on the internet..You could try : using prompts with your name, prompts with your friends names, prompts who you think might have the most pictures on the internet, maybe try prompts with big RUclipsrs like Mr. Beast and such..Is it just actors?
You can't just search for your name. The AI only knows what has been included in the training data, which obv. isn't gonna be pictures of random people with their names. Celebrities are in sufficient amounts of media that they can appear pretty well. But if you look at DreamBooth with Stable Diffusion, you can now do a fine tune of the SD model with inputting images of something you want the AI to recognize, like your cat, or yourself, and then the AI will be able to generate images of that thing using a keyword, so you could generate images of yourself walking on the Moon or turning into the Hulk etc. You only need like a dozen or more images (tho more is better).
Been increasingly disappointed with Dalle 2 over the past few months. The quality of the results seem to have deteriorated greatly, with really mushed incoherent results, and if anything, they are getting worse. Meanwhile, midjourney v4 is imho, the best model I have ever used. Not sure what openAI are doing, but they really need to up their game. If this trend continues, they'll be dead in the water within 6 months. Imagen / parti and Nvidia bringing out their own models, plus ongoing tremendous progress in midjourney and SD, is some mighty competition.
The only ai worth buying is novel ai because of its really good anime art generation, and midjourney is basically free infinitely because you can easily make alt accounts to use rofl so midjourney is the best for normal images/realism and novel ai is the best for both text generation and anime/other stuff that is also anime if yknos what I mean generation
I'm also using NAI's model lately on Opus tier. It is fantastic at Anime, to the point where it's almost hard to get a bad result. The only problem I have with it is that it seems almost overfitted to Anime because it's hard to make anything that isn't cute girls (or maybe that's just me, lol).
Why is AI art exist? What is AI artists? Somebody who input a bunch of prompt is the same as we commission Artist to do the art, right? How is it a tool when it's done everything for us from design to shading?
You can keep changing stuff over and over till you find exactly what you're visioning, does it matter the way i create it? Maybe it's through my hands by using a brush and paint on a canvas, maybe it's through a camera, maybe it's through a mouse in MS Paint or maybe it's through these new technologies, if "normal" artists can call a banana taped to a wall, a random splash of paint, or even a completely white canvas art then i can call an image generated from a word art too... Sure drawing or painting is clearly more skillful and calling someone that just wrote "painting" on some AI website an artist is a stretch but you can definitely do more than just type some words with these tools also not every imagery has to be defined as art, maybe someone just quickly need a picture for a school presentation, a meme, a background for a pfp, etc.
@@_._._._._._._._._._._._._. you are pitching an idea to a machine that makes art for you, sometimes it is not even a specific idea, do you have any sense of satisfaction of creating art this way?
Bhahaha goes to show you know nothing about ML or how it actually works. It needs dataset and it actually try to trace back to the original piece. That’s why in the style of getting you better results, it just looks different because of math. Tried to really simplified it for you - but yea, they also not suppose to use the database as LAION stated it’s for research purposes so stop the for profit bs it doesn’t work this way.
What do you think? Has DALL-E 2 been DETHRONED? *My DISCORD Just got Revamped! Check it for the LATEST AI news! **discord.com/invite/bQgcbjs2Sg* ▼ Link(s) From Today’s Video:
✩ Gilbatree's Video (ft. me!) ruclips.net/video/EadcJIz-E48/видео.html&ab_channel=Glibatree
✩ Midjourney: www.midjourney.com/home/
✩ DALL-E 2: labs.openai.com/
Three notes to add about MidJourney:
1) This is currently alpha-testing of V4
2) You can use --chaos and then a number from 1-100 to introduce more diversity in the results
3) You can use --q 2 to let MidJourney spend a little more time to potentially increase quality
dall-e 2 is definetly dethroned, although dall-e 2 is better at photographs, its still not very coherent. Midjourney is super good at digital art style with a lot of creativity, so i'd say midjourney is the best right now
Midjourney is amazing. I don’t think the detail is there when it comes to humans and certain other things though. I think of it kind of like my love for Nintendo, versus when PlayStation first came out.
I haven't compared them directly with photographs but for MJ you can use " prompt, photograph by (famous photographer), (camera) (lens type)"
I still prefer dall e 2 when making humanoid animals since Midjourney isn't that very good at making humanoid things. But mj is my second favorite now and hopefully in future version they can now male humanoid animals like what dall e 2 can
NAI's model is best for anime from what I've seen. Although I haven't used MidJourney and I am shocked by how good these results were.
MJ was better before at times!
And is the photo better? Seriously?)) And show me your photos, maybe they are on the websites of photographers, and so the photographers would not have guessed. I already have a lot of these made with the help of MJ, moreover, quite a few already have the status of a photo of the day. You need to know how to use. Of course, I finalized it in Photoshop, correcting small jambs, but there are just small ones.
There is a big problem with these type of comparisons: the same prompt does not necessarily lead to the same latent vectors in the two different software. Which can be intended in your comparison or unintended, let me explain.
So on the one hand, a better comparison would be to figure out what extra text should be added to the DALLE2 prompt (for example "highly detailed, trending on artstation" etc) to generate the exact same style that Midjourney uses by default, then generate the pictures and compare them. It covers the use case where people are using the two software after reading some promp-guides. For example at the warlock example you could see that DALLE2 indeed can generate better than clipart quality (or not, but that would be the point of the comparison). It is completely possible that Midjourney generates its artistic style because either it adds "artistic style" to every prompt, or because the model is trained in a way that the text encoder makes the latent vector skewed towards the "artistic style" direction by default. You could test it by trying to generate explicitly non-artistic style renderings, and comparing those.
On the other hand, the vast majority of users might not have knowledge about those prompt-guide tricks, and for them the default results will look better, and that comparison was covered in this video.
A completely different thing: it seems that you have the problems with bigfoot, because the word "bigfoot" is not an entity in the text model (in Stable Diffusion it is CLIP). So the diffusion model generates "big foot", and now you have it...
Nevertheless, I liked this video, it was informative, thanks for the effort. :)
Yeah, the Midjourney style was definitely artificially biased towards artistic images by the developers, while Dall-E 2 seems to be mostly unbiased in terms of style. Well, except in terms of OpenAI randomly adding "diverse" people (blacks, Asians) when you try to do create a person, and except for the fact they try to remove celebrities etc...
The point is, Dall-E 2 manipulates the content, while Midjourney also manipulates the style. Which makes Midjourney look more impressive at first glance, but apparently prevents any photorealism.
Great analysis, so what's better to you?
If you want more variety in the pictures for MJ V4 use the chaos parameter. V4 is incredible!
i prefer using M670, it helps with the consistency of homogeneous areas with many different colors and makes my images stand out amazingly
we have now got to a point in the ai art generating space where dall e 2 is getting left behind by midjourney and stable diffusion, and only a few months ago, dall e mini felt like gold to us
Dalle 3 is going to change the game
Eh its because sb and mj keeps getting updated and i think openai is more focused to creating dall e 3 more than updating dall e 2
@@evalover8622 true, but rn we havent heard anything about dall e 3, so i wouldnt expect it until mid next year, and by then sd and mj wouldve probably caught up to dall e 3 when it comes out cuz they r updating at a rapid pace
@@FRareDom just look at the difference between dalle and dalle 2, dalle 3 will be at the retina level, meaning at the average viewing distance at the average resolution, it will be very hard to distinguish from other photos
For me Midjourney has won hands down on everything I've done. I just love the inventive and creative art-leaning style it has
I still find that they are all are bad at making vehicles especially helicopters.
@@orlock20 True. And their one big weakness is hands imo
My favorite part of the new V4 algorithm is not the images themselves, but how much MidJourneys vocabulary expanded! I can finally create images in the style of Arcane, or concept art from Valorant
Ofcourse, from valorant.
You’re the AI info/graphical stud, man. Thank you for keeping us up to date on this stuff!
v4 is only in alpha right now, they're still working on more big improvements (better upscaling, different aspect ratios, overall image quality etc). The current state is already fantastic, but there's even more to come.
You can nudge Midjourney into photorealism btw, when you're using the right prompts.
There is quite a big difference between midjourney v3 and v4 models. Last one creates exact more accurately what you asked in prompt. So, i think it's a killer of all existing txt2img generative AI at the current moment.
It’s definitely better than anything else out there right now, it’s capable of understanding and depicting ideas I couldn’t get before. It truly feels like there’s a lot less limits
Thanks for the review! I don't use DALL-E or MJ often, so this helps me stay current. I feel like MJ is best if you want something consistent, but it's also limited by that built-in painterly aesthetic. DALL-E has self-imposed limits that cripple txt2img, but the img2img, inpainting, and outpainting work great on an existing image. Still, I'm very happy I can run Stable Diffusion on my own PC. Over 88k generations and counting.
Version 4 is a big improvement. The dragons in it are amazing now.
Great analysis, I actually love the comparisons you made with the new model! And thanks for the very kind shout out!
Thanks man! And thanks for having me in your video! 👌
Oh hey Glibatree! This model is awesome
@@carolusmagnus9757 Heyyo, fun to see you in the wild
Now if only MJ would run outside Discord.
It makes absolutely no sense
Its because they unfortunately have not made their software/code opensource like the others have.
Yeah, I had to join discord and learn all the ins and outs of it just in order to run Midjourney.
Ah, it seriously runs in Discord? That is pretty gross lol.
@@remylebeau2357 That has absolutely nothing to do with their idiotic decision to be on Discord alone. They could easily set up a normal web page, like any other business.
I tried the new V4 right now and it Blew up all my expectations 😮 ❤️
Just a helpful hint. You commented on the sameness of the "character concept, old mage warlock" images. But if you re-rolled, your would get a different family of images (for example, when I just tried it, mine had pointy hats, and three of them carried staffs), and if you use the --chaos parameter, you can get wider differences within the group of four (I tried it again with --chaos 90, and I got elven wizards, bareheaded ones, ones reading tomes, although, yeah, they were all still white-bearded white guys) --your friendly neighborhood MJ guide
This is actually facts, I overlooked this. Thank you for the insight
I love your video style..It really helps to separate the upcoming tech and you are making it clear…. Thanks!!
Woow, this is an incredible progress. This blows everything in past away.
Thanks for your Comparison
Midjourney clearly uses invisible dynamic prompting. Ie; they are are auto-filling your prompts for you, but don't show you the things they added. It's very cool but i think they should either A. Give you the option to turn that on or off and generate very basic looking results or B. at least show you the secret sauce they are adding on for you to make things come out in their signature MJ style.
I rate SD above them all tbh. Nothing beats the control, the power, the flexibility, i get that it's not very user friendly, but being able to train your models and create hypernetworks and have all this control over 'how' the AI interprets your prompt and so many other features, there's nothing else like it. SD is the Linux of the txt2img world, MJ is the Windows, and DALLE is the Mac. With the latter two, what you sacrifice in power and control, they make up for in user friendliness. So if you don't want the AI to spoonfeed your prompts for you and you want to have more independence in how you communicate with it and how it understands you, then you need to go for that technical proficiency and learn all the cool tricks of the trade. SD is made for that.
Sorry to have to say, but you are incorrect. It all simply comes down to the "brain" that the Ai gets implanted with. So basically the image database that are used to train the Ai, coupled with the algorithym that is paired to that dataset that dictates exactly how MJ is expected to interpret prompts. Simple as that. Each one of these psuedo-Ai's has it's own "personality" or "style" or lack there of in Dall-E's case. The most obvious and easy to spot being disco diffusion. No matter what prompts you feed into disco diffusion, it will always be easily identifiable as Disco diffusion simply because it has the heaviest hand in adding its signature style to everything it does. How do I know? I am a programmer and prompt engineer who has been working with Ai models for the past 4 years.
@@remylebeau2357 I understand how models and training datasets work, but that doesn't detract from my assertion at all, both can be true and i believe that to be the case. They are not mutually exclusive concepts.
it’s funny that Matt got more excited about the avocado chairs, but I thought the wizard and dinette images showed the greatest superiority of Midjourney over DALL-E 2!
Great video! very helpful and informative. Might be worth writing the prompts that you used in your description for people to copy and compare results :)
the AI models are changing for the better so quickly, I can imagine that in 10 years from now we will be able to ask an AI to generate a whole movie, using simple voice commands "I want an action film with actors X,Y,Z, they must do A,B & C" lol
I literally just bought a $4,000 computer because I am so impressed by this new AI stuff.
I can't even grasp how amazing this is. I am in a non-stop state of shock.
No, I am not rich.
I don’t blame you! Currently, I can do low resolution stable diffusion images on my six year old Alienware Area 51, but I am thinking of buying a new computer just so I can do AI art on it!
@@AmazingArends
May I ask what your setup is? Because I can't seem to do anything with my current setup (3600x + gtx 1060)
@@dialecticalmonist3405 How did you spend $4000 and get a low end CPU/GPU? Or is that your old rig? If you spent 4k on a 3600x / 1060 then you got absolutely scammed.
@@larion2336
The 3600x + gtx1060 is my CURRENT computer. Lol.
I ordered a 13900k + 4090 from Corsair.
@@dialecticalmonist3405 Ahh, I see. Thank god for that haha.
This is extremely helpful information. Thank you for your research and exploration into this emerging creative medium!
Dall-e got an "outpainting" functionnality that is very performant and can be use with MJ images to extend them.
You can usually tweak the Midjourney output to be more photographic-like by spamming camera-related keywords. If the prompt doesn't specify anything related, it tends to default to painting-like.
As an example, to your Bigfoot prompt, I was able to turn it into a realistic photograph by adding "selfie photograph, 35mm, kodak gold 200, 1980 Polaroid" in front of the prompt.
Absolutely incredible stuff, really. Thanks for your great videos; I've learned a lot!
This is exciting now, but I'm really interested in how good these AI engines are going to be 5 years from now.
Midjourney absolutely blew DALL-E 2 in every prompt. No doubt about that.
Basically in all the images it produced, none of the characters fell into the uncanny valley, all of them had correct faces, limbs, everything. DALL-E 2 seriously struggles in maintaining this coherency inside the image itself, when compared to MJ.
Also, the diversity generated by the prompt, outside of the chaos parameter in MJ, partially happens because MJ doesn’t alter your prompt like DALL-E 2 does, by appending words like woman, black, and other ethnic descriptions automatically. So the prompts actually being compared were not even the same.
I don't use dall e, so I didn't realize it did that. Honestly, is it too much to ask to just leave my prompt alone and make what I asked?
I agree with you that Midjourney won hands down.
Playing with it now, pretty incredible. I honestly wished more Discorders would have voted to hold off until we have the various aspect ratios though
Thanks for the detailed comparison. I’m not sure it’s really an ‘Apples to Apples’ comparison, so I wouldn’t go so far as to call it a dethroning. Dali has a bias towards photographic realism and in the case of the armchairs, they were much more useful designs for an avocado-style chair that people would actually want to sit on.
Midjourney is as heart an illustrator and certainly got a massive jump in coherency now. But I wouldn’t want sit on those chairs at all, however avocadoey they might be, and if I were a chair designer, the Dali version would be more helpful. The same goes for the shitzu dogs, both look good to my eyes and Dali is realistic whereas Midjourney is incredibly creative.
What will probably dethrone Dali is the release of Google’s Imagen and/or Dali 3, which will probably boost coherency and details to another level. This photographic realism is presumably one reason why they block depictions of real people, whereas Midjourney is in the land of the playfully creative and happy to be there.
But yes when it came to things Dali can’t do (the waitress or Warlock) Midjourney certainly smoked it. But even so, I’d say, without really replacing what Dali would do, if it could.
You can increase the variation in MJ by modifying the chaos figure.
To get more photographic results in MJ you have to use different prompt words, like calling out a specific camera, etc.
This tech is so insane... and we're only at ground zero. In our lifetimes we'll see user prompted AI virtual worlds. Unreal.
A lot of Dall-e pics seem to have a corporate toy look compared with the more artistic expression that Midjourney outputs..😳
The moment when dogggy becomes pirate doggy
As an artist I’m both nervous and excited about this AI art stuff. Excited because of the things I will be able to create that I wasn’t able to before and nervous because I’m scared of being completely replaced by AI.
I’m goad your atleast honest with yourself many artists are in denial about this tech
BETTER. I'm excited, watching this immediately
Probably unintentional, but the abrupt cuts kept making me laugh 🦔
Blame my stupid editing software
@@MattVidPro Nahh that's fantastic editing software*
O....M...G... the artistic coherency! VIVA LA MIDJOURNEY!
Overall, I think Midjourney v4 is better than Dall-E 2.
I used to never care about MidJourney, now it may be the best.
You've probably gotten this before, but I just realized... the way you speak kinda reminds me of Carl Sagan. You also look like you could be his grandson or something. Cool!
I have never heard this before 😂
Great news, btw in the last couple of videos, it sounds like Descript has been cutting the end of your sentences
Yeah not sure what the issue is… it sounds good when editing but once it’s rendered it seems like it cuts them off more
the difference in creativity and artistic skills is just crushing in MJ's favor.
16:15 you are wondering why there are a lot of feet and then saying BIG-FOOT several times after that ;D
"Something happened to grandma" lol
It's really better, but it only lacks a loophole to work for free :)
Deactivate account> create new account> repeat
@@MetigArt How? i tried it but it said i need to wait 24hrs to deactivate my account, I'm so broke and i'm tired of creating hundreds of accounts just to get another 25 free image
@@i.washere That's the only way, I'm afraid. Ya just gotta wait for the cooldowns. :/
to clarify, v4 is in Alpha and has not been released, its available for a limited time for testing.
Prompts "Bigfoot", wonders why there are many feet in the pictures...
Next time please test stable diffusion as well.
I liked Midjourney more dan Dall-e 2 before
but it seems to be distancing itself more and more
midjourney is a lot better for me, I get way better results
Okay yeah Midjourney is blowing Dalle-2 out of the water. This is nuts
Great video, thank you.
The frog logos dall-e generated are definitely better. The midjourney ones would be unusable as logos due to their complexity, unless you want to look very unprofessional.
I was interpreting it based on how close it was to the prompt
You look like the baseball kid from Meet the Robinsons if he actually grew up to be a normal guy.
Not the first time I’ve heard that
MidJourney in some situations makes Dal-E look like a WISH product..
Isn't the AI "look" basically defined by what images it's trained on? Seems like Midjourney is trained on Artstation/DeviantArt, while Dall-E on Google Images (mostly photos). Most of Dall-E looks like someone scrapbooked together a bunch of Google Image photos.
for stable diffusion you have to train a model to get such results or write a big ass prompt essay about the thing with a big ass negative prompt.But you get tons of flexibility. If somethings not working img2img it,. 😅
The time it takes to walk 3 kilometers has got to be all kinds of depressing.
I don't know what to think. It has amazing potential to make our dreams true, but it hurts real artists really bad. Makes you think how many jobs ai will replace in the future...
Is actually better even Photographic style, u just have to put some Phtography prompts
Amazing!❤
Just cause dall e came first doesn't mean they will be the better man
We should not comparing different AI with the same prompts. It's like comparing the female portrait created by the Stable Diffusion 1.5 with Waifu Diffusion 1.3. I can guarantee you that 99% is Waifu Diffusion win, because the model has heavy style in it. I think he should at least include different styles and artists in the prompts.
I need to start learning a.i. damn
Isn’t MD based on Stable Diffusion these days?
DALL-E 2 was dethroned many months ago when Midjourney came out. They haven't really made any serious gains, the images are still very poor quality, the in-painting maybe looks great to use but from a commercial point of view I'd never use DALL-E 2, I have lots of credits, never used them. Even Stable Diffusion is way better at quality and coherence, that is free to use also. Midjourney is night and day compared to the others, for $30 a month it has everything I need so far, I'll stick with it for the foreseeable.
fascinated about what will kill us all...
You know if they have changed their artists or I mean added more with this version? Midjourney I mean.
is there any way to get mid journey free or cheap...?
That amazing...
I still refuse to pay for any of these AI Art generators except for one. That one you don’t cover or haven’t covered yet. It’s light years beyond all other AI tools in my opinion. Only time will tell though regarding which one is become the G.O.A.T.of AI Art tools.
And what is that?
It’s my best kept secret 🤐
I prefer the realism in Dall-e because the "painting" look on these are kinda weird..
What gallery are you accessing for that midjourney render library?
Its right there in the discord where you have to go to use MJ to begin with.
I love Midjourney V4 compared to Dall-E 2 but why do the images have to be so dark
That upscaled avocado chair should definitely calm down or it's going to get a NSFW warning lol.
It's not that I don't like Midjourney, I'm an early member on their Discord and still have a few image credits, but guys... Is vignetting++ and stylized look mandatory ? :(
Though DALL-E 2 never really evolved so it was the king of the game back in the days, but now I very rarelly use it personnally, and when I do, it's to get init images for Stable Diffusion img2img '^^
Thanks.
17:00 it clearly misunderstood the prompt - if they're playing video games, they're doing it on their iphones, so it didn't understand what an "iphone photo" was. Is that a good prompt for other engines? I've never heard anybody describe a photo as looking like an "iphone photo" - how would you tell? If the model hasn't learned that, that's probably a good thing?
The prompt was intended to be extremely difficult for the AI models. I’m sure you could get a better image for the same idea, but I want to see how well the difference ai interpret the prompt. It clearly shows their differences
midjourny will never be at the top for me until they add inpainting feature
Is this getting to be like a regular Google search? Can you type in your (anyone’s) name? Instead of saying “Google yourself!” could you say “Prompt yourself!”.? Or is it just famous people? Maybe you could see if your pictures come up, as you have pictures on RUclips and such..Maybe we are getting exceptionally close to making our Avatar from pictures of us gathered on the internet..You could try : using prompts with your name, prompts with your friends names, prompts who you think might have the most pictures on the internet, maybe try prompts with big RUclipsrs like Mr. Beast and such..Is it just actors?
You can't just search for your name. The AI only knows what has been included in the training data, which obv. isn't gonna be pictures of random people with their names. Celebrities are in sufficient amounts of media that they can appear pretty well. But if you look at DreamBooth with Stable Diffusion, you can now do a fine tune of the SD model with inputting images of something you want the AI to recognize, like your cat, or yourself, and then the AI will be able to generate images of that thing using a keyword, so you could generate images of yourself walking on the Moon or turning into the Hulk etc. You only need like a dozen or more images (tho more is better).
awesome
Midjourney 4 is superior to Dall-E hands down.
Plot twist this entire channel is ai generated
i hate this LOL I am so excited and midjourney is not working :(
Been increasingly disappointed with Dalle 2 over the past few months. The quality of the results seem to have deteriorated greatly, with really mushed incoherent results, and if anything, they are getting worse. Meanwhile, midjourney v4 is imho, the best model I have ever used. Not sure what openAI are doing, but they really need to up their game. If this trend continues, they'll be dead in the water within 6 months. Imagen / parti and Nvidia bringing out their own models, plus ongoing tremendous progress in midjourney and SD, is some mighty competition.
Can't the a.i. create itself?
The only ai worth buying is novel ai because of its really good anime art generation, and midjourney is basically free infinitely because you can easily make alt accounts to use rofl so midjourney is the best for normal images/realism and novel ai is the best for both text generation and anime/other stuff that is also anime if yknos what I mean generation
I'm also using NAI's model lately on Opus tier. It is fantastic at Anime, to the point where it's almost hard to get a bad result. The only problem I have with it is that it seems almost overfitted to Anime because it's hard to make anything that isn't cute girls (or maybe that's just me, lol).
Why doesn't DALL-E allow Walter White? He's not a real person.
If you wanted iPhone selfie results than include the exact Pixel dimensions, --ar 19.5:9.
Why is AI art exist? What is AI artists? Somebody who input a bunch of prompt is the same as we commission Artist to do the art, right? How is it a tool when it's done everything for us from design to shading?
You can keep changing stuff over and over till you find exactly what you're visioning, does it matter the way i create it? Maybe it's through my hands by using a brush and paint on a canvas, maybe it's through a camera, maybe it's through a mouse in MS Paint or maybe it's through these new technologies, if "normal" artists can call a banana taped to a wall, a random splash of paint, or even a completely white canvas art then i can call an image generated from a word art too...
Sure drawing or painting is clearly more skillful and calling someone that just wrote "painting" on some AI website an artist is a stretch but you can definitely do more than just type some words with these tools also not every imagery has to be defined as art, maybe someone just quickly need a picture for a school presentation, a meme, a background for a pfp, etc.
@@_._._._._._._._._._._._._. you are pitching an idea to a machine that makes art for you, sometimes it is not even a specific idea, do you have any sense of satisfaction of creating art this way?
Only true chads have ten-thousand accounts in midjourney
looks great, oh... almost forgot 4th
dalle-2 looks like junk now.
👋
Wow. AI is scary, but it is the way. Hope we aren’t destroyed by it in the end, but how cool.
w
What creativity and style? It doesn’t come up with its own, its stealing the identity of artists
It doesn't steal anything, it learns
Bhahaha goes to show you know nothing about ML or how it actually works.
It needs dataset and it actually try to trace back to the original piece. That’s why in the style of getting you better results, it just looks different because of math. Tried to really simplified it for you - but yea, they also not suppose to use the database as LAION stated it’s for research purposes so stop the for profit bs it doesn’t work this way.
@@natv6294 ok bye
@@FabrizioCia He's right tho
And those artists "stole" (were inspired by) the work of artists who came before them. What exactly is the difference here?
4st
It's good. Really good.
3st
xd