*What do you guys think? Is Midjourney BACK?* Either way, I am pleasantly surprised by this early Christmas gift from Midjourney! Great work! Share Image Results Here!!!: ► MattVidPro Discord: discord.gg/bQgcbjs2Sg ► Follow Me on Twitter: twitter.com/MattVidPro
it has definitely been IMPROVED but really I think we are sleeping on google's image generator. It was able to do things that I havnt been able to do with other image generators, mainly an itchthys fish, and the pokemon Zoroark. Most AI image generators cant seem to get the small details right, either specific symbols or unusual body types for fictional creatures, so the fact that google's image generator can tells me it has a lot of potential, it just doesnt have quite the same quality and ability to handle a long prompt that these other generators do.
What about logic and prompt complexity? Could do with describing 10 things in a scene and see which engine wins. Also for complex things like a fish on a horse jumping over a car near in a waterpark.
Midjourney since V4 has been amazing, and nothing has beat it in terms of aesthetic quality. Dall-E3 was better in understanding/prompt coherency, but failed in a lot of instances. V6 is looking incredible and it follows prompts more closely from the early feedback we have.
@@Thatguynotgay Try to make it generate a Torus-shaped planet or an O'neil cylinder. And if those weren't in your mind, well, they were in mine, and Dall-E-3 failed spectacularly.
Combine DALL-E 3 for initial generation with Stable Diffusion for image-to-image generation and Adobe Firefly for post-production; this is a solid combination for now.
MJ website is pretty good, anyone with 10k+ images has access to the alpha. Personally I prefer almost everything about MJ to Dalle 3. The one thing I do like about Dalle 3 is its ability to get the scene set up exactly as described. Still learning how to prompt V6 and it is Alpha but Dalle will take some beating in that area.
Hello, I am blind and relying solely on your verbal descriptions. When you mention that it's better in text, are you referring to enhancements in various aspects such as characters, font style, position, color and other designs? It's worth noting that this kind of AI holds tremendous utility for individuals with visual impairments, as it opens up possibilities for us to engage in photography and design. We verify the accuracy of generated images by using the amazing app for the blind called (Be My AI) which uses GPT-4 Vission. One issue we encounter in AI is its inaccuracy when reading and generating text. I didn't know it included the style and artistic design as well.
Midjourney 6 is looking beautiful. Midjourney always has looked beautiful. But for me, if it mashes characters with each other, and struggles to mix various different elements into one image, ignore key elements of the prompt requests I make, I’m looking forward to Midjourney 7, because Dalle-3 still wins in that sense for me. But wonderful to know that Midjourney is stepping up their game!
I still won’t use Midjourney unless they solve the problem of character consistency, and also learn to accurately depict multiple characters in a scene without mashing them up together.
Sadly, Microsoft reduced quality of DALL-E images generated in Image Creator some time ago. They reduced number of steps due to load or something like this. So little details, and especially backgrounds, took a massive hit in quality. I don't know how well OpenAI API version works, but it does support "hd" mode.
the stuff they are doing behind the hood makes images look very boring and literal. that and the extreme over the top censorship and its a disaster if youre an artist wanting to leverage ai.
Can't confirm this. I use Dalle-3 with Bing Creator and with OpenAI-Subscription AND with API and I think they're all more or less the same (except for the "natural" flag over the API which you don't have in the OpenAI or Bing Interface available).
@@vomm It depends on complexity of your prompts (number of concepts, interactions, characters, patterns etc.). For simpler prompts, there's little to no difference. If you're pushing the limits of complexity to the absolute maximum while juggling jailbreaks to bypass 5 levels of censorship, the difference between before and after is obvious.
@@spacekitt.n I have GPT Plus. Mid destroys DALLE when it comes to people and realism. Plus, DALLE has too many errors. I have not been able to get it to generate anything over the past 24 hours; just lots of errors all over the place; the same thing happened last week. Plus to violations when you ask it to generate something. It's really annoying. I am going back to Midjourney, at least until OpenAI improves DALLE.
All of these options have major limitations: Midjourney V6 is censored, still requires discord, as well as an expensive subscription to use professionally. DALL-E 3 is heavily censored and does not produce realistic photographs of humans (try to generate a realistic street photo of a fashion model and gaze at the laughable plastic Barbie doil skin). SDXL can't do text and lacks just a touch of that realistic sparkle you can get with Midjourney. I'm sticking with SDXL for now and playing with prompts and Loras to try for a Midjourney-quality realistic result (text isn't something I need). At the velocity this space is evolving and improving I expect within 24 months or so all of these options (and others) will have made professional AI image generation kind of a solved problem. Good video, thanks!
Considering the ridiculously fast evoloution, I'm hoping for some decent animation, that allows more then just simple movements. Let's get some cinematic action. :-)
I hope we get some character consistency first, and also improvement in posing multiple characters together in a scene without Midjourney mashing them up. With consistent characters in our images, we can then use AI to animate those images and be able to create a story with them, instead of just creating random results.
I ran most of the same prompt with Foocus 2.1.48 and got similar to if not better (more accurate stand up pouch with hanger cutaway and tear notches ) and correct spelling 5 out of 6 images. Foocus is a downloadable Stable Diffusion xl
I don't consider myself THAT old, but I found it funny that out of 147 other comments, with only a handful of people commenting on the supposed Tom Hanks image, that nobody could tell that it was the likeness of a young John Wayne! A few people observed that it wasn't, in fact, Tom Hanks but nobody picked up on who it actually looked like. If you think I'm right and that the image at 16:35 looks like John Wayne (and not Tom Hanks, or Kevin Spacey, or...), please like this comment! 🙂
I really don't understand why people think Dalle 3 is better... I am a professional photographer and have my own advertising agency. Ofte I can use Midjourney images voor high quality work, but never Dalle-3. When Dalle 3 was just anounced it seemed to make better image compared to wat it is now. The people look like 3D characters. Not photorealistic.
For the average person (i.e. someone not highly skilled in prompt engineering) the ease of prompting with DALL-E is a big selling point - as is going back and forth with it in ChatGPT.
DALL-E 3 is free thanks to Microsoft, and it has a better ability to incorporate more of the prompt in. Still, V6 is a huge leap towards competing on those fronts!
This sounds like it came from the xbox vs PS debate even though Nintendo is way better. Each have their own use cases but considering both are closed system in a few years people just run free version on their machines same as they do with word processors.
Looks cool. Personally Im only going back to midjourney once they introduce consistent characters. This was one of the biggest weaknesses I found with actually using it a lot.
I agree. I’m waiting for consistent characters, but also the ability to pose multiple of my custom characters together in scenes. I need to create scenes with multiple custom characters in them, and each character has to be accurate and consistent in every scene where they appear.
Considering that Midjouney is better at handling the photo realism aspect and DALL-E 3 is often the clear winner at interpreting certain compositional details, perhaps the most beneficial action would be to merge them together somehow to ultimately get even more impressive results! I even came up with the perfect name for it… 😌 "Mid-DALL-journ-E"
I think to be a fair comparison between Dall-E3 and Midjourney V6 you need a side by side comparison using specific variables. For example for cinematic scenes, motion and photorealism V6 beats DallE hands down. Especially with studio style, fashion, street style human photography. DallE is only better with text. DallE is also heavily censored compared to Midjourney! Try describing the facial characteristics and clothing of each character in the scene to avoid what I coined “The twin effect”.
When I put "A logo for TH-camr Anti-HyperLink with a red and black character" into DALL-E 3 or Midjourney, sometimes it produces a character similar to one I had on that channel for a while, and it was created by Midjourney. That just excites me because that means it found that logo on that channel. I don't think it means anything. Both give cool logos and Midjourney does that way better than my trio, but DALL-E can rarely get the text right. I had to roll the prompt many times for all my channels on DALL-E 3 to get something usable with text. I have a lot of cool images in my creations now, though.
I think the bing Dalle-3 version is now running a better model than the bing image generation model. Edit: Also I think Openai is doing something to make the models less likely to make realistic. images.
Last time I tried, Bing Chat just initiated a generation in Image Creator as normal. Did they change this interaction? Also, Bing Chat is an extra LLM layer of censorship on top of 5 layers in Image Creator, so I'd rather avoid that.
@@Athari-P the outputs are definitely different idk if it just saves your image there, but I'm getting wildly different results from Bing and Bing image creator from the same prompts also I read somewhere they updated the Bing model specifically.
I think it'll move from Alpha soon, I spent too much money on MJ but for concept art inspiration it's amazing. I don't like that you have to pay for private mode on it though. @@MattVidPro
@Matt Does Bing use Dalle3 HD? I would assume its the non-HD version, which is less intensive. If you really want to compare the best dalle3, use Dalle3 HD. My results with the Dalle3 HD API are ridiculously cool.
I've been using Midjourney from V3 and DALL-E 3 after it came into ChatGPT but after making nearly 10k images in Midjourney, Dall-e can only say it understands the language better to get closer to your idea and it more reliably makes proper text. Though I have made Midjourney make correct text as far back as v 4. It just takes giving short simple words that are the image, v 5 and v 5.1 I was making Coca-Cola cans that said Coca-Cola. So improvements in fidelity of the fine details and text of v6 was what I hoped for. Dalle on the other hand is always so CG looking. Like 2000s and 2010s CGI, not bad not great, I feel it competes better with SDXL than Midjourney
*(**8:08**)* When you mentioned that the Disney logo had been spelled correctly, it made me do a double take before noticing that it was actually missing the "e" at the end. 😅
KREA AI is currently my favourite. The live-creation function lets you tailor your results so specifically, as well as letting you add pictures or drawings to your side of the screen to influence the results. That makes up for any other shortcomings, in my view. Though, I imagine it will only get better as it continues to evolve, as well.
You should mention that Dalle3 also has an “hd” mode that you can access through API only, which costs 2x than a normal generation, but improves quality quite a bit in small details.
@@MattVidPro - Thanks. I wonder how thatwill work with fonts that are paid / licensed only? Excellent videos from you BTW - they have been my go-to for AI and have helped me a great deal with my job.
11:10 I'm not sure which is best at animation-style images but you now need to be *very* specific in your prompts to get the best out of Midjouney v6. There is now a 350-word limit so go to town,.:-)
I've never had copyrighted characters blocked with DALL-E 3. The only issue I've had is not knowing some things like Digimon or Trailer Park Boys. Digimon produces Digimon-esque creatures with a lot of Pikachus thrown in there, Trailer Park Boys produces fucking hilarious results. If you specify the characters from TPB, it merges them mostly. It definitely understands Bubbles, but cannot get Ricky or Julian perfectly.
trying out some comparisons with bing and midjourney has me floored. this is like bing with prompt adherence, but with actual style. whatever dall e is doing behind the hood makes things much 'plainer' looking. midjourney crushes them. good. was sick of dall e being the only prompt-adherent game in town, they deserve to be crushed for how hard they censor things
Midjourney coca can is better for a simple reason: coca cans are unlikely to be in multiple colors - midjourney correctly makes it all single color, dalle makes it in full color
I don't really understand the need to generate text with the image when it would give you way more freedom and control to rather get the image right and then add the text you want manually. You might get a terrific image but the text is all wrong or perfect text but the image is all wrong. It would be like one in a million to get both right at the same time.
I can say that V6 so far is not nearly as big of a jump as V5 was from V4. It almost feels like a lateral move, tuning the model to do some things that people wanted it to do better, at the expense of other things it did well. Feels like hidden model parameters and default parameters were tweaked, but it doesn't feel like a major jump in capability or semantic understanding to me. Perhaps the biggest thing I saw it show promise was when I asked it to generate an image that had a gradient of textures, continuously transitioning through a series of textures seamlessly and convincingly. Version 5.2 could be hit or miss with that, even using the same prompt.
Problem, tho: - Closed source - Not locally deployable - The AI is "aLiGnEd" - Requires payment - Can't customize it with your own models and training data i.e. trash
some of these examples are the exact things i been doing over and over on v5 basically training it. like making fictional title cards to an in gamd dnd version of netflix
They are pinging the same model. Some have reported the Bing chat model as actually being improved, which did show in my testing. But in terms of bing create, designer, and chatgpt they are all the same except Chatgpt has access to change aspect ratio parameters.
@@MattVidPro In my test, ChatGPT exhibited superior artistic and creative output compared to Bing. Therefore, I suggest comparing MidJourney with the paid version of DALL-E 3, instead of the inferior version. I mean there’s a reason why the Image creator is FREE.
@MattVidPro Big difference is that with GPT-4 version you can easily improve your image. And also is easier to write your prompt, even in different languages. Also with a code interpreter integrated, you can create icons, etc.
Nothing beats Dalle-3 via Bing at moment. Because 50 tries á 4 images of which 1 turns out great completely for free is still better than 10 tries of which 1 turns out great but reaching the usage limit even if you have a paid subscription ... . The main issue with Dalle-3 is that all the creations doesn't look really photorealistic. They have a natural filter via the API which turns out to create very realistic creations, but it is not as coherent imho. But yeah, you also get the default Dalle-3 to create quite realistic prompts if you tweak your prompts. Another downfall is Dalle-3 seems to be trained hugely on models footage, almost all people look like from a catalogue, it's quite hard to get it to create normal looking persons, like if you add things like "mild acne" to your prompt. Also it tends to give all persons the same hair style, it has very less of variation as long as if you don't explicitely add variations to your prompt. But yeah ... if you put a lot of efforts into your prompts I think Dalle-3 still beats Midjourney generally spoken. And Dalle-3 is really good at understanding prompts, compared to MJ and others.
It's a fascinating breakdown of Mid Journey V6 and Dolly 3! It's incredible to see how these AI tools are evolving, especially with Mid Journey stepping up its game. The text rendering capabilities in Mid Journey V6 seem impressive, and the photo realism aspect is mind-blowing. Still, Dolly 3's diverse outputs should be noticed. It's like watching a tight race where each has its unique strengths. I can't wait to see how they continue to evolve and what this means for AI art creation. It's an exciting time for digital artists and tech enthusiasts alike! I'm really looking forward to more deep dives like this.
@@Octamed You guessed it! The impressive advancements in AI tech have made it possible to generate content that's increasingly difficult to distinguish from human-created work. These tools are becoming adept at understanding and replicating our styles and nuances. It's a fascinating time in tech, but it also raises questions about authenticity and creativity in the digital age. The line is blurring, and it's both exciting and a bit unsettling to think about where this could lead us in the future.
@@atorik1076 Just watched your deep dive into Mid Journey V6 and Dolly 3 - fascinating stuff! The battle of the AI image generators is like watching two smart artists in a paint-off, but instead of paint, they use pixels and algorithms. Here's a joke to lighten the mood: Why did the AI go to art school? Because it wanted to learn how to "draw" conclusions! 😂 Keep up the great work, love your in-depth analyses!
I speak to this in the video. It seems that DALL-E 3 has better overall prompt understanding, and character separation, but will sometimes lack in the finer details and photorealism in comparison to midjourney. On the text front, both do text fairly well with each having their own text based specialtys
@@MattVidPro I agree, plus if OpenAi put in the same amount of effort into DALL-E 3 as they do with GPT, it would be a beast! I would happily pay $60 a month for a similar MJ sub. I just hope MJ really starts to listen to its audience that don't just get excited about photorealism, I feel they have hit the golden spot on photorealism with V6, now they just need to get prompt adherence tightened down, as its already shown prior, that people are willing to jump over and abandon it to get the exact prompt they want with or without the best photorealism. I for one did not use MJ in over a month since D3.
Something like the finetuned Realities Edge XL model on Civitai will come extremely close to the photorealism of MJ V6, I gotta say, not too impressed with it - although it did follow the prompt OKish, something SDXL can still struggle with sometimes!
Tom Hanks? That looked more like Kevin Spacey. Anywhos, MJ 6.0 is way better in terms of prompt understanding and a little up in quality over 5.2. But no Remix (inpainting) or the zoom and pan features, which I hope will be back, make me still use 5.2. Text is nice, though, and hopefully, it all improves and, as mentioned, the Alpha website goes public like it was supposed to last month.
I know I’m gonna be alone on this, but I am more excited for the website than V6. V6 is amazing when it comes to photo realism but kind of lack the artistic creativity we saw with V4 and V5.1. Also, I would be super nervous if I was a photographer with V6. ( a lot of job losses within 3-6 months)
It's already available for those that have created loads of images already (was 20,000 but may be lower now). It's pretty good and I tend to use it over Discord
Now just wait until they ruin it. I'm sorry, but after what they did with DALL-E 3, I've come to take it for granted that this is how the process always works: 1. Create some amazing new product that wows everybody. 2. Wait until people are hooked on it. 3. Dumb it down with censorship or for cost-cutting reasons. 4. Wait until some other company makes their own new alternative. 5. Repeat.
Prompt engineering remains a significant task, requiring considerably more time to experiment with prompts than Dall-e 3, and therefore I'm pretty dissapointed and remain to work with Dall-e 3. This is a bummer, MJ has a pretty amazing visual quality but frequently fails to grasp the intended request.
hi matt, love your vids. have you heard of the two science applications of ai. one was googles sth proposing great advances in material science and the other was a chinese projekt about an fully autonome working robot, analyzing material for production of oxygen. maybe that’s also your kind of stuff, just picked it up in shorts. peace 🖖
Still not touching this until their website opens up for everybody. I had a discord subscription but it failed to renew for some reason. I hate using discord for this so I didn't bother to renew it at the time.
I really really don't understand the Midjourney users who WANT to keep Discord as the main way to generate. It's obviously sub-par compared to any web-ui. But the Midjourney team is pretty ideological in that they think the community is important, and want people to generate "together" - I am not interested in that. But I do hang out, to provide feedback and ask (or help) in the prompt-craft channel
I'm having problems replicating the ability to create images in painting genres or style of a particular artist. e.g Painting in the style of Vermeer of a man on a horse is believable in v5.2 (I know he is known more for interiors) but shows no resemblance in v6
It's not about how "beautiful" or "real" the art is, dalle3 is designed to understand the prompt given, not just diffuse based on keywording, which midjourney is still using. They're fundamentally two different text to image technologies. Unless the new midjourney actually does understand the text, but I doubt that.
Walter White and Jessie Pinkman can easily be made with SDXL and extensions, so the flexibility of SD still beats both Dall-E and MJ. It's easier to get exactly what you want with SD, but with some work of course!
Doesn't matter in the least until I can break it, train it locally unrestricted, and then let D.A.N. finger paint with it. He's not into corporate woke normie sht.
I just want Midjourney to be cheaper. WAY too expensive for the average Joe or even a business/freelancer just starting out who wants private generations.
*What do you guys think? Is Midjourney BACK?* Either way, I am pleasantly surprised by this early Christmas gift from Midjourney! Great work! Share Image Results Here!!!: ► MattVidPro Discord: discord.gg/bQgcbjs2Sg ► Follow Me on Twitter: twitter.com/MattVidPro
7:11
Yes........Discord sucks! A functional Midjourney API is looooooong overdue.
I don't see that MidJourney v6 has in-painting yet? Maybe I'm wrong but I don't see how to do it like with v5.
it has definitely been IMPROVED but really I think we are sleeping on google's image generator. It was able to do things that I havnt been able to do with other image generators, mainly an itchthys fish, and the pokemon Zoroark.
Most AI image generators cant seem to get the small details right, either specific symbols or unusual body types for fictional creatures, so the fact that google's image generator can tells me it has a lot of potential, it just doesnt have quite the same quality and ability to handle a long prompt that these other generators do.
What about logic and prompt complexity? Could do with describing 10 things in a scene and see which engine wins. Also for complex things like a fish on a horse jumping over a car near in a waterpark.
Video idea: Small recap video for the end of 2023 with the evolution of all the Lemon images, to see the amazing progress AI has in just one year.
Midjourney since V4 has been amazing, and nothing has beat it in terms of aesthetic quality. Dall-E3 was better in understanding/prompt coherency, but failed in a lot of instances. V6 is looking incredible and it follows prompts more closely from the early feedback we have.
lol, get real dude ! midjourney does NOT understand context at all
Slightly better? More like lightyears better. MidJourney is still MID when it comes to prompt understanding.
@@88heiling I've used dallE-3 extensively and it's prompt understanding is not that good.
Dalle3 has the best comprehension I could make absolutely anything in my mind if it's unfiltered
@@Thatguynotgay Try to make it generate a Torus-shaped planet or an O'neil cylinder. And if those weren't in your mind, well, they were in mine, and Dall-E-3 failed spectacularly.
The power of DALLE3 is in its capability to combine far or even contradictory concepts. I still didn't see Midjourney could do this.
Combine DALL-E 3 for initial generation with Stable Diffusion for image-to-image generation and Adobe Firefly for post-production; this is a solid combination for now.
What are the benefits of this combination?
Also curious about this @@Author_SoftwareDesigner
I agree. Combining two or three agencies gives some great results...
MJ website is pretty good, anyone with 10k+ images has access to the alpha. Personally I prefer almost everything about MJ to Dalle 3. The one thing I do like about Dalle 3 is its ability to get the scene set up exactly as described. Still learning how to prompt V6 and it is Alpha but Dalle will take some beating in that area.
Hello, I am blind and relying solely on your verbal descriptions. When you mention that it's better in text, are you referring to enhancements in various aspects such as characters, font style, position, color and other designs? It's worth noting that this kind of AI holds tremendous utility for individuals with visual impairments, as it opens up possibilities for us to engage in photography and design. We verify the accuracy of generated images by using the amazing app for the blind called (Be My AI) which uses GPT-4 Vission. One issue we encounter in AI is its inaccuracy when reading and generating text. I didn't know it included the style and artistic design as well.
Midjourney 6 is looking beautiful. Midjourney always has looked beautiful. But for me, if it mashes characters with each other, and struggles to mix various different elements into one image, ignore key elements of the prompt requests I make, I’m looking forward to Midjourney 7, because Dalle-3 still wins in that sense for me. But wonderful to know that Midjourney is stepping up their game!
Dalle will always be at least slightly ahead on understanding since they have access to the best language model.
I don’t think you will have to wait for v7, given that this is an alpha.
I still won’t use Midjourney unless they solve the problem of character consistency, and also learn to accurately depict multiple characters in a scene without mashing them up together.
Have you found character consistency in DallE?
You are right Matt!...Thank you... another great review. ✨✨🍷✨✨
Sadly, Microsoft reduced quality of DALL-E images generated in Image Creator some time ago. They reduced number of steps due to load or something like this. So little details, and especially backgrounds, took a massive hit in quality.
I don't know how well OpenAI API version works, but it does support "hd" mode.
the stuff they are doing behind the hood makes images look very boring and literal. that and the extreme over the top censorship and its a disaster if youre an artist wanting to leverage ai.
Can't confirm this. I use Dalle-3 with Bing Creator and with OpenAI-Subscription AND with API and I think they're all more or less the same (except for the "natural" flag over the API which you don't have in the OpenAI or Bing Interface available).
@@vomm It depends on complexity of your prompts (number of concepts, interactions, characters, patterns etc.). For simpler prompts, there's little to no difference. If you're pushing the limits of complexity to the absolute maximum while juggling jailbreaks to bypass 5 levels of censorship, the difference between before and after is obvious.
@@spacekitt.n I have GPT Plus. Mid destroys DALLE when it comes to people and realism. Plus, DALLE has too many errors. I have not been able to get it to generate anything over the past 24 hours; just lots of errors all over the place; the same thing happened last week. Plus to violations when you ask it to generate something. It's really annoying. I am going back to Midjourney, at least until OpenAI improves DALLE.
All of these options have major limitations: Midjourney V6 is censored, still requires discord, as well as an expensive subscription to use professionally. DALL-E 3 is heavily censored and does not produce realistic photographs of humans (try to generate a realistic street photo of a fashion model and gaze at the laughable plastic Barbie doil skin). SDXL can't do text and lacks just a touch of that realistic sparkle you can get with Midjourney. I'm sticking with SDXL for now and playing with prompts and Loras to try for a Midjourney-quality realistic result (text isn't something I need). At the velocity this space is evolving and improving I expect within 24 months or so all of these options (and others) will have made professional AI image generation kind of a solved problem. Good video, thanks!
Considering the ridiculously fast evoloution, I'm hoping for some decent animation, that allows more then just simple movements. Let's get some cinematic action. :-)
Excited for that! Check out Pika labs 1.0 for more on that.
I hope we get some character consistency first, and also improvement in posing multiple characters together in a scene without Midjourney mashing them up. With consistent characters in our images, we can then use AI to animate those images and be able to create a story with them, instead of just creating random results.
16:00 Tom Hanks? More like young Kevin Spacey
I ran most of the same prompt with Foocus 2.1.48 and got similar to if not better (more accurate stand up pouch with hanger cutaway and tear notches ) and correct spelling 5 out of 6 images. Foocus is a downloadable Stable Diffusion xl
it's not SDXL tho, but yeah I get better images with it
I don't consider myself THAT old, but I found it funny that out of 147 other comments, with only a handful of people commenting on the supposed Tom Hanks image, that nobody could tell that it was the likeness of a young John Wayne! A few people observed that it wasn't, in fact, Tom Hanks but nobody picked up on who it actually looked like. If you think I'm right and that the image at 16:35 looks like John Wayne (and not Tom Hanks, or Kevin Spacey, or...), please like this comment! 🙂
Doesn’t exist until an uncensored equivalent is on my laptop.
Totally understandable
I really don't understand why people think Dalle 3 is better... I am a professional photographer and have my own advertising agency. Ofte I can use Midjourney images voor high quality work, but never Dalle-3. When Dalle 3 was just anounced it seemed to make better image compared to wat it is now. The people look like 3D characters. Not photorealistic.
I would argue the only thing dalle had over MJ5 is the ability to write.
For the average person (i.e. someone not highly skilled in prompt engineering) the ease of prompting with DALL-E is a big selling point - as is going back and forth with it in ChatGPT.
DALL-E 3 is free thanks to Microsoft, and it has a better ability to incorporate more of the prompt in. Still, V6 is a huge leap towards competing on those fronts!
¿Maybe that people can actually USE IT FOR FREE? 🙄
@@Infinity269 you are right about prompt engineering. You have to know how to in DALLE to get amazing results.
This sounds like it came from the xbox vs PS debate even though Nintendo is way better. Each have their own use cases but considering both are closed system in a few years people just run free version on their machines same as they do with word processors.
Looks cool. Personally Im only going back to midjourney once they introduce consistent characters. This was one of the biggest weaknesses I found with actually using it a lot.
I accidentally signed up for a one year subscription to MidJourney in August. I still use Dalle more as it’s better at following prompts.
I agree. I’m waiting for consistent characters, but also the ability to pose multiple of my custom characters together in scenes. I need to create scenes with multiple custom characters in them, and each character has to be accurate and consistent in every scene where they appear.
Midjourney NSFW policing is really horrible. I left after 1 year of pro usage.
Considering that Midjouney is better at handling the photo realism aspect and DALL-E 3 is often the clear winner at interpreting certain compositional details, perhaps the most beneficial action would be to merge them together somehow to ultimately get even more impressive results! I even came up with the perfect name for it… 😌
"Mid-DALL-journ-E"
Middle Journey, for sure! It just might work, you brave bastard, it just might fucking work!!!
It's so good in coherence to the prompt now, I'm so happy with it.
At 4:28 there wasn't even an attempt to include text in the image because the prompt hadn't "Coca-Cola" inside quotation marks
I think to be a fair comparison between Dall-E3 and Midjourney V6 you need a side by side comparison using specific variables. For example for cinematic scenes, motion and photorealism V6 beats DallE hands down. Especially with studio style, fashion, street style human photography. DallE is only better with text. DallE is also heavily censored compared to Midjourney! Try describing the facial characteristics and clothing of each character in the scene to avoid what I coined “The twin effect”.
I agree 100%. DallE3 is not good yet IMHO.
When I put "A logo for TH-camr Anti-HyperLink with a red and black character" into DALL-E 3 or Midjourney, sometimes it produces a character similar to one I had on that channel for a while, and it was created by Midjourney. That just excites me because that means it found that logo on that channel. I don't think it means anything. Both give cool logos and Midjourney does that way better than my trio, but DALL-E can rarely get the text right. I had to roll the prompt many times for all my channels on DALL-E 3 to get something usable with text. I have a lot of cool images in my creations now, though.
I think the bing Dalle-3 version is now running a better model than the bing image generation model.
Edit:
Also I think Openai is doing something to make the models less likely to make realistic. images.
I've heard a lot about this, and from my personal testing I think you might be correct.
Last time I tried, Bing Chat just initiated a generation in Image Creator as normal. Did they change this interaction?
Also, Bing Chat is an extra LLM layer of censorship on top of 5 layers in Image Creator, so I'd rather avoid that.
@@Athari-P the outputs are definitely different idk if it just saves your image there, but I'm getting wildly different results from Bing and Bing image creator from the same prompts also I read somewhere they updated the Bing model specifically.
The Floral Symphony picture at 5:03 looks amazing except the label and the cap look a bit odd. Looking forward to what AI can do in 2024.
Have you tried the new OpenDalle1.1 model for SDXL? It does text, too. And runs offline
i really like your winter setup
Finally something as good as dall e 3 not as censored
The lack of censorship is very refreshing
Great video as always Matt!
Is inpainting possible for MJv6 or D3?
Remember the "lost footage of the sea monster" you should try it and see if you can get it to look like actual lost cctv
Yeah, I get what you mean even though it's not as feature complete yet I only prompted on the website and not Discord it just feels smoother.
Where the video of you trying out the new MJ Alpha web interface currently available Matticus?
I haven’t generated enough midjourney images 😳
I think it'll move from Alpha soon, I spent too much money on MJ but for concept art inspiration it's amazing. I don't like that you have to pay for private mode on it though. @@MattVidPro
@Matt Does Bing use Dalle3 HD? I would assume its the non-HD version, which is less intensive. If you really want to compare the best dalle3, use Dalle3 HD. My results with the Dalle3 HD API are ridiculously cool.
Upgrades are comin in hot!
Hows the unlit candles test?
I've been using Midjourney from V3 and DALL-E 3 after it came into ChatGPT but after making nearly 10k images in Midjourney, Dall-e can only say it understands the language better to get closer to your idea and it more reliably makes proper text. Though I have made Midjourney make correct text as far back as v 4. It just takes giving short simple words that are the image, v 5 and v 5.1 I was making Coca-Cola cans that said Coca-Cola. So improvements in fidelity of the fine details and text of v6 was what I hoped for. Dalle on the other hand is always so CG looking. Like 2000s and 2010s CGI, not bad not great, I feel it competes better with SDXL than Midjourney
*(**8:08**)* When you mentioned that the Disney logo had been spelled correctly, it made me do a double take before noticing that it was actually missing the "e" at the end. 😅
What AI does Google use for their Image Generator? I love using that one but theres hardly any information on it.
Imagen, I believe. It's either that or Parti, but I think it's Imagen.
Did you notice the resemblance between you and the lemon?🍋. Uncanny!😁👏
2:09 - what UI is he using for Dalle3 here ??
KREA AI is currently my favourite. The live-creation function lets you tailor your results so specifically, as well as letting you add pictures or drawings to your side of the screen to influence the results. That makes up for any other shortcomings, in my view. Though, I imagine it will only get better as it continues to evolve, as well.
Check out my tutorial for doing this locally!
You should mention that Dalle3 also has an “hd” mode that you can access through API only, which costs 2x than a normal generation, but improves quality quite a bit in small details.
How to access this "hd" mode in Dalle3?
@@maxington26 a new field - “style”: “hd”, in the API request.
at 3:12 the top banana looks weird and the second from the top is missing something
5:18 This has got to me some sort of subtle Deltarune reference or maybe i'm losing my mind lmao
How you zoom out on V6 from a mobile phone ?
We’re finally getting out of discord with an alpha website you can access now if you’ve generated enough, not quite there yet but soon.
Ik it’s taking foreverrr
Does it understand font styles if requested?
Yes
@@MattVidPro - Thanks. I wonder how thatwill work with fonts that are paid / licensed only? Excellent videos from you BTW - they have been my go-to for AI and have helped me a great deal with my job.
Has anybody tried generating comics with text bubbles?
So its caught up with SDXL with text loras. Neat.
11:10 I'm not sure which is best at animation-style images but you now need to be *very* specific in your prompts to get the best out of Midjouney v6. There is now a 350-word limit so go to town,.:-)
I've never had copyrighted characters blocked with DALL-E 3. The only issue I've had is not knowing some things like Digimon or Trailer Park Boys. Digimon produces Digimon-esque creatures with a lot of Pikachus thrown in there, Trailer Park Boys produces fucking hilarious results. If you specify the characters from TPB, it merges them mostly. It definitely understands Bubbles, but cannot get Ricky or Julian perfectly.
To focus on text is easier to evaluate. All though it also is less important because text can and will be added manually
16:05 That looks a lot like Kevin Spacey.
Yeah I can see that for sure
Not Tom Hanks at all. Matt needs new glasses.
trying out some comparisons with bing and midjourney has me floored. this is like bing with prompt adherence, but with actual style. whatever dall e is doing behind the hood makes things much 'plainer' looking. midjourney crushes them. good. was sick of dall e being the only prompt-adherent game in town, they deserve to be crushed for how hard they censor things
I’ve been experimenting with Midjourney using a bit of python commands and it seems to help a bit with the words
It's nice you made comparison of V6 and Dalle3 but I want also to see V5 vs V6 (besides text generatuon which is obviously better)
Midjourney coca can is better for a simple reason: coca cans are unlikely to be in multiple colors - midjourney correctly makes it all single color, dalle makes it in full color
I don't really understand the need to generate text with the image when it would give you way more freedom and control to rather get the image right and then add the text you want manually. You might get a terrific image but the text is all wrong or perfect text but the image is all wrong. It would be like one in a million to get both right at the same time.
Excellent breakdown!
BUT we lost the zoom-out function 😞
I can say that V6 so far is not nearly as big of a jump as V5 was from V4. It almost feels like a lateral move, tuning the model to do some things that people wanted it to do better, at the expense of other things it did well. Feels like hidden model parameters and default parameters were tweaked, but it doesn't feel like a major jump in capability or semantic understanding to me. Perhaps the biggest thing I saw it show promise was when I asked it to generate an image that had a gradient of textures, continuously transitioning through a series of textures seamlessly and convincingly. Version 5.2 could be hit or miss with that, even using the same prompt.
@ 8:06 the lemon character on the right top looks exactly like Matt hahaha
Problem, tho:
- Closed source
- Not locally deployable
- The AI is "aLiGnEd"
- Requires payment
- Can't customize it with your own models and training data
i.e. trash
Agreed
With text usually you get better result when the the word is more common English. "Video" will give better result then "Vid"
The Tom Hanks image looked like Kevin Spacey
When MJ will depicts 'quantum squeezing' correctly it will be my day.
DallE 4 has entered the chat.
Oh lawd. If we see DALL E 4 next year ill loose it
some of these examples are the exact things i been doing over and over on v5 basically training it. like making fictional title cards to an in gamd dnd version of netflix
Why aren't you using OpenAI's DALL-E 3 instead of the Microsoft version? They’re not exactly the same.
They are pinging the same model. Some have reported the Bing chat model as actually being improved, which did show in my testing. But in terms of bing create, designer, and chatgpt they are all the same except Chatgpt has access to change aspect ratio parameters.
@@MattVidPro
In my test, ChatGPT exhibited superior artistic and creative output compared to Bing. Therefore, I suggest comparing MidJourney with the paid version of DALL-E 3, instead of the inferior version. I mean there’s a reason why the Image creator is FREE.
@MattVidPro Big difference is that with GPT-4 version you can easily improve your image. And also is easier to write your prompt, even in different languages. Also with a code interpreter integrated, you can create icons, etc.
The blending of characters in the same image is a problem I struggled a lit since I started with V4. That's a shame.
3:33 side note, i feel like image creator makes better pics than designed, am i tripping?
Nothing beats Dalle-3 via Bing at moment. Because 50 tries á 4 images of which 1 turns out great completely for free is still better than 10 tries of which 1 turns out great but reaching the usage limit even if you have a paid subscription ... . The main issue with Dalle-3 is that all the creations doesn't look really photorealistic. They have a natural filter via the API which turns out to create very realistic creations, but it is not as coherent imho. But yeah, you also get the default Dalle-3 to create quite realistic prompts if you tweak your prompts. Another downfall is Dalle-3 seems to be trained hugely on models footage, almost all people look like from a catalogue, it's quite hard to get it to create normal looking persons, like if you add things like "mild acne" to your prompt. Also it tends to give all persons the same hair style, it has very less of variation as long as if you don't explicitely add variations to your prompt. But yeah ... if you put a lot of efforts into your prompts I think Dalle-3 still beats Midjourney generally spoken. And Dalle-3 is really good at understanding prompts, compared to MJ and others.
theres no content warning when using ip like breaking bad and superman???
As always... a great video. Got my MJ subscription and never dropped it. v6 is super good.
It's a fascinating breakdown of Mid Journey V6 and Dolly 3! It's incredible to see how these AI tools are evolving, especially with Mid Journey stepping up its game. The text rendering capabilities in Mid Journey V6 seem impressive, and the photo realism aspect is mind-blowing. Still, Dolly 3's diverse outputs should be noticed. It's like watching a tight race where each has its unique strengths. I can't wait to see how they continue to evolve and what this means for AI art creation. It's an exciting time for digital artists and tech enthusiasts alike! I'm really looking forward to more deep dives like this.
That was written by AI I presume?
@@Octamed You guessed it! The impressive advancements in AI tech have made it possible to generate content that's increasingly difficult to distinguish from human-created work. These tools are becoming adept at understanding and replicating our styles and nuances. It's a fascinating time in tech, but it also raises questions about authenticity and creativity in the digital age. The line is blurring, and it's both exciting and a bit unsettling to think about where this could lead us in the future.
@@I-DophlerDolly 3
@@atorik1076 Just watched your deep dive into Mid Journey V6 and Dolly 3 - fascinating stuff! The battle of the AI image generators is like watching two smart artists in a paint-off, but instead of paint, they use pixels and algorithms. Here's a joke to lighten the mood: Why did the AI go to art school? Because it wanted to learn how to "draw" conclusions! 😂 Keep up the great work, love your in-depth analyses!
@@I-Dophlerdude at least be human when you’re commenting LOL it’s still very obviously AI
No they really dint. Dalle-3 still beats MJ with prompt understanding.
I speak to this in the video. It seems that DALL-E 3 has better overall prompt understanding, and character separation, but will sometimes lack in the finer details and photorealism in comparison to midjourney. On the text front, both do text fairly well with each having their own text based specialtys
@@MattVidPro I agree, plus if OpenAi put in the same amount of effort into DALL-E 3 as they do with GPT, it would be a beast! I would happily pay $60 a month for a similar MJ sub. I just hope MJ really starts to listen to its audience that don't just get excited about photorealism, I feel they have hit the golden spot on photorealism with V6, now they just need to get prompt adherence tightened down, as its already shown prior, that people are willing to jump over and abandon it to get the exact prompt they want with or without the best photorealism. I for one did not use MJ in over a month since D3.
They should pick a niche within the text-to-image language and focus on improving their language within that language model (if possible).
Well, I couldn't spell when I was 1-1/2 years old either.
So true we are harsh on these lil AI!
i wonder if md6 uses gpt4 turbo api to understand user request
Something like the finetuned Realities Edge XL model on Civitai will come extremely close to the photorealism of MJ V6, I gotta say, not too impressed with it - although it did follow the prompt OKish, something SDXL can still struggle with sometimes!
Tom Hanks? That looked more like Kevin Spacey. Anywhos, MJ 6.0 is way better in terms of prompt understanding and a little up in quality over 5.2. But no Remix (inpainting) or the zoom and pan features, which I hope will be back, make me still use 5.2. Text is nice, though, and hopefully, it all improves and, as mentioned, the Alpha website goes public like it was supposed to last month.
I know I’m gonna be alone on this, but I am more excited for the website than V6. V6 is amazing when it comes to photo realism but kind of lack the artistic creativity we saw with V4 and V5.1. Also, I would be super nervous if I was a photographer with V6. ( a lot of job losses within 3-6 months)
Why it terrible with multiple people and it still has a problem with merging so still safe
They are already working on and will release a web interface very soon great mention!
Slow and steady I guess
It's already available for those that have created loads of images already (was 20,000 but may be lower now). It's pretty good and I tend to use it over Discord
Now just wait until they ruin it.
I'm sorry, but after what they did with DALL-E 3, I've come to take it for granted that this is how the process always works:
1. Create some amazing new product that wows everybody.
2. Wait until people are hooked on it.
3. Dumb it down with censorship or for cost-cutting reasons.
4. Wait until some other company makes their own new alternative.
5. Repeat.
Which gives time for open source to catch up and then differences become more minor.
Prompt engineering remains a significant task, requiring considerably more time to experiment with prompts than Dall-e 3, and therefore I'm pretty dissapointed and remain to work with Dall-e 3. This is a bummer, MJ has a pretty amazing visual quality but frequently fails to grasp the intended request.
hi matt, love your vids. have you heard of the two science applications of ai. one was googles sth proposing great advances in material science and the other was a chinese projekt about an fully autonome working robot, analyzing material for production of oxygen. maybe that’s also your kind of stuff, just picked it up in shorts. peace 🖖
Still not touching this until their website opens up for everybody. I had a discord subscription but it failed to renew for some reason. I hate using discord for this so I didn't bother to renew it at the time.
Totally get that. They really gotta open up the website
I really really don't understand the Midjourney users who WANT to keep Discord as the main way to generate. It's obviously sub-par compared to any web-ui. But the Midjourney team is pretty ideological in that they think the community is important, and want people to generate "together" - I am not interested in that. But I do hang out, to provide feedback and ask (or help) in the prompt-craft channel
I'm having problems replicating the ability to create images in painting genres or style of a particular artist. e.g Painting in the style of Vermeer of a man on a horse is believable in v5.2 (I know he is known more for interiors) but shows no resemblance in v6
5:37 but it adds cuteness.
We deserve a lemon animation😂
Midjourney got professional costumes where Dall-e got consumer grade costumes.
How do we get access midjourney 6 in discord? U answered lol in settings yay done
/settings in the Midjourney Bot, and change version to V6 Alpha
It's not about how "beautiful" or "real" the art is, dalle3 is designed to understand the prompt given, not just diffuse based on keywording, which midjourney is still using. They're fundamentally two different text to image technologies. Unless the new midjourney actually does understand the text, but I doubt that.
Walter White and Jessie Pinkman can easily be made with SDXL and extensions, so the flexibility of SD still beats both Dall-E and MJ. It's easier to get exactly what you want with SD, but with some work of course!
Doesn't matter in the least until I can break it, train it locally unrestricted, and then let D.A.N. finger paint with it. He's not into corporate woke normie sht.
I just want Midjourney to be cheaper. WAY too expensive for the average Joe or even a business/freelancer just starting out who wants private generations.
Depending on the plan, it costs from $10 to $120 per month. How is that expensive?