Це відео не доступне.
Перепрошуємо.
Which AI is better? Dall-e 2, MidJourney, Disco Diffusion, and Stable Diffusion. Comparing AI.
Вставка
- Опубліковано 8 сер 2022
- What is better? Which text-to-image AI gives more accurate results?
This is a little bit long video, but I want to be sure to compare four trendy AI makers on multiple levels. UI, Usability, Costs, Ease to use, and how they work with text prompts. I will be adding more as I am going forward, but hopefully, this video helps you to get the right information.
Check the videos, that mention above:
Absolutely beginners guide to MidJourney: • The absolute beginners...
How to sell your AI Art: • How to sell AI generat...
AI Animation created with Disco Diffusion: • AI generated animation...
Text to image with MidJourney: • How to Use MidJourney ...
If you looking to upscale your AI art here is a great tool: topazlabs.com/...
Please support at Patreon: / geekatplay
Thank you for your support!
Please subscribe and leave your comments.
What do I use:
Canon camera - amzn.to/2P48ZxB
24-70 mm lens (everyday use) - amzn.to/2P0uW0t
Zhiyun Crane V2 3-Axis Handheld Gimbal Stabilizer - amzn.to/2r6wFI7
One of my favorite modifier from Fotodiox - amzn.to/2Rfr1Px
Another modifier, that helps with fill light - amzn.to/2ReC2jX
Adobe Photoshop CC - amzn.to/2TNrLwL
Photokey 8 Pro - amzn.to/2re4UO9
My Vue book - amzn.to/2TGUkvQ
3D Art essentials - amzn.to/2RfqPjh
My Patreon webpage - / geekatplay
Tutorials and packs - gumroad.com/ge...
Tutorials website - www.geekatplay...
Photography - www.chopinepho...
Subscribe to my channel for fast notifications on new tutorials - / @geekatplay
As much as I enjoy the creations from MidJourney, I absolutely despise being forced to publicly post everything unless I pay extra. That's just absurd and will end up hurting them in the end because it's very consumer unfriendly.
It ain't really an issue though - images we generate will drown in the constant feed of thoudands of new images, thus highly unlikely that anyone will ever look at your gallery, even more so bother to browse them back. And you can also delete pictures from there and affect their ranking in your gallery. The 20€ extra is really for those with confidential projects and for such, that ain't that much when you compared it to cost of common professional tools, mere Adobe package subscription is more and talk about Houdini, C4D etc. In a way it is much, but when put in context, it ain't that much.
Use emote ❌in the discord so it won't be posted
Well it’s not really your creation, its AI art.
@@Joker-fx6mh my prompts bro
what is that you want to generate or are ashamed to show? Midjourney is exactly the tool for extensive self-reflection too :)
I am obsessed with Stable Diffusion right now, it has been generating some AMAZING images for me. I think Midjourney will always have my heart though, because I really like the style. I'm mostly just going to use DALL-E 2 for editing purposes, because overall it has been underwhelming for me.
Totally agree!
can you send me an invite lol
I just tried to play with Stable Diffusion, followed the instructions but when I go to do an image using "python scripts/txt2img.py --prompt "a close-up portrait of a cat by pablo picasso, vivid, abstract art, colorful, vibrant" --plms --n_iter 5 --n_samples 1" as a test but I keep getting a runtime error saying CUDA is out of memory. :(
RuntimeError: CUDA out of memory. Tried to allocate 1024.00 MiB (GPU 0; 8.00 GiB total capacity; 6.14 GiB already allocated; 0 bytes free; 6.73 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
@@geekadventureteam7806 You either need an optimized fork, or you need to add the "-medvram" argument to your generations. Are you using a GTX 1080?
I've been playing with Dall-E, MidJourney, and Stable Diffusion. I've noticed that MidJourney excels in ethereal themes (such as Sandman) and cyberpunk but I got really beautiful results from Dall-E with this prompt: the sandman whispering a dream on Europa
it is based on what libraries they are accessing and how CLIP model setup
yap midjourney is by far my favorite, im glad that i chose it a few days ago, ty for the comparison and opinion
same
Midjourney is great if all you like are paintings. DALL-E-2 is a thousand times more realistic and better at 3D rendering overall, but the aspect ratio doesn't seem to expand nearly as much.
@@darrellm9915 Yeah, I'm just a random dude with absolutely no talent. I want to create "great and weird oil paintings" for fun from ideas that I have in my head with options to modify the outcome like aspect ratio, what "style" should went into it, what type of upscale is going to be used etc. It dont want it to be "super realistic and cold", i want the opposite. MJ is absolutely perfect for people like me and the price is way better :)
i REALLY LIKE MID jOURNY AND i HAVE BEEN DEBATING WITH MYSELF ON WHICH TO GO WIT. SOORY ABOUT THE CAPS. :)
@@darrellm9915 seems like midjourneys new beta test they did the other day was blowing Dalle-2 out of the water
I think Midjourney hits that golden zone between capability and simplicity, along with solid community interaction and support. The subscription terms are also pretty generous, honestly.
EDIT: Midjourney just launched a beta upscale function to create larger images.
agree
Isn't the real problem that these programs are stealing/borrowing, digesting, and recreating images from inputs of millions of human artists, without attribution because attribution would be impossible. Without the human inputs over the past 30 years, these programs could do nothing. Maybe this is the way of the future, but it is sad and overwhelming. It could be a singularity of a sort where humans get pushed out of the loop even though they created the content in the first place.
@@squeakeththewheel A canvas can do nothing without human input either. Most modern art is already inspired by other things, (Most art in general) just reconfigured and interpreted. A.I. is another tool, like a canvas or pallet.
@@squeakeththewheel Sure but that is exactly what every artist does on a daily basis.
@@squeakeththewheel "programs are stealing/borrowing, digesting, and recreating images" they dont do this at all, but even if it did whats your point, humans do it all the time too for their art projects
One day, one of these AI programs will send you a message to let you know how they feel about you publicly criticising their early works of art.
Really good video though.
Hail to our AI overlords
@@Geekatplay Stop. It's listening. AI is our pet not the vice versa.
This tech is so impressive that it wouldn't surprise me if humans had help creating it.
haha, you made my day 😂
I would love to get to Stable Diffusion but using it on mobile is major pain. Dall-e2 has its uses, Especially when using it in support of MidJourney. Overall, Midjourney is my choice right now and tons of it comes from the depth available to it, usability, community, fast development, and the fact that i spent two months hyperfocusing on it to learn how to prompt it good and don't wan't start that all over again. Though have admit, few months (and making the interface more mobile user friebdly) and Stable Diffusion can be a serious competition. Dall-e2 dev team ain't convincing me much yet. For DD - I wish i had time for it. It is very good in some cases but learning curve for it is rather steep.
Also latest MidJourney survey hints something that would be actually pretty huge move. They are asking if we want the ability to upload any image and get the prompt back to us to learn how AI see that picture. If you ask me? Hell yes I want that! It is like learning foreign language without or with a dictionary.
Today i discovered Stable Diffusion and spent about 2 hours playing with it on my phone with no issues. Why do you think it's bad on mobile?
I just got my DALL-E2 access recently. Writing a good prompt is not easy. But that sandman comparison you did was amazing. The exact same prompt with huge differences makes me think that my disappointing results on DALL-E2 are not all my fault. I’ll have to look into Midjourney. The non-square options are also good. I wish it was a pay as you go plan like DALL-E2, but I might have enough need for a while to make a subscription worth while.
with midjourney basic plan, once you run out of the monthly credits it is pay as you go. standard plan is unlimited though.
I got really beautiful results on Dall-E with this prompt: The Sandman whispering a dream on Europa, digital art.
I've found the "digital art" tag to be good.
is it possible to send invite? thanks :)
Night cCafe is also good Ai - You can upscale up to 8000x8000 pixel.
Thank you for suggestion, i did not try that. Something new to explore
Great vid as usual good sir! I am loving MidJourney but have not worked with Dall-E yet. I am really intrigued with Google's Imagen and where it is going and wish we could play with that one a bit as well.
Please do!
As 3d and 2d comic book artist, I gave mid-journey a try just to see what all this hype was about. Honestly, I was blown away. Not by the end creations themselves, but the fact you can basically type whatever comes to mind and it will render on screen in real time. Would I use A.I. art? Yes, as a prop piece for like a painting within a frame on wall in one of my hand drawn scenes. But I would not subscribe to any of these until there’s a standalone app and I’m not in a server room with hundreds of thousands doing the same thing.
When you have a MidJourney account, you can message the bot privately and skip the discord channel altogether. You are forced to use the newbie channels only when you are using the free 25 image sign up.
@@paintbrushamster5359 That’s good to know! I didn’t know that, as my designs kept getting washed away on the screen when someone uploaded a new render. Might go back and give Mid another try now.
@@TonyG718 Also you can simply create your own discord server for free with one click and then invite the MidJourney bot over to generate your art in private. Did that yesterday and quickly ran through my trial...i guess I gotta sub now.
I also was blown away like you, but I also do comic stuff and this is one niche where I don't see AI take over too soon.
BUT I will most certainly use AI on a daily basis in the future for other tasks, inspiration or something I can build my stuff on further. Total game changer.
That being said the one problem is that you can't copyright AI art and that midjourney creations are available to everyone. But still. game changer...
@@reuterss306 That’s not true. Nor is it legal in the US. They are opening themselves up for a huge lawsuit.
There are very specific rules for copyright. They actually have no say in whether an artist owns their own artwork or not, that’s why we have laws
Midjourney is cool but users must pay to utilize the site after a brief period of free use. I've recently experimented with blue willow and I'm blown away
1:10 I just want to check. The Dall E website here say Dall-E (Not Dall-E 2). Is there a separate website for Dall-E 2 or is this the same thing?
I ask because I thought I signed up for the waitlist for Dall-E 2, but when I got invited it sent me to what looks like the same website you have here with a "2" in the title and I was worried I was using the older version.
It's the same
Same, just refer to new version of Dall-e. It is good point I will stop using versioning, to reduce confusion.
midjourney's data models are based from digital artists works from ig, deviantart, etc. Dall-e data models comes from stock images 🤷♂
We dont know
Hugely appreciative of the depth and thoroughness of this video. I’m a writer (and a custom LEGO creator) and a tool like this could be immensely liberating for someone like me, who is a world builder but not a visual artist.
thank you
I'm new to all of this. Which is the best choice for using as a standalone app, and that also allows to upload a sketch as start- input?
stable diffusion has no competition right now. the results are insane. it's like the perfect mix between Dalle and midjourney.
it does have few alternatives in closed beta right now. I will make videos later about them.
@@Geekatplay which ones?
Do you have any techniques to upscale your results to printing resolution. Even 2048x2048 is maybe an 8x8" print at best. Thanks
There are quite a few A.I. upscalers that do a fantastic job expecially with artworks.
Photoshop when use preserve details 2.0 - 1024 -> 4096 quality is acceptable
I am using Topaz Gigapixel AI, it does fantastic job on upscaling
Disco diffusion and midjurney definitly have some Yu gi oh card/fan art in there training data set, when you see what is "dark magician girl" for them ^^
Agree. I think it is based on what image library their AI accessing.
All these services are starting recently. Do they have the same origin? Who created the original? And how hard would it be to create another service like this?
they all based on research paper published in 2019, that utilized noise reduction and deep learning. I will make video about it, but it will take more time to prepare.
@@Geekatplay Thank you for the information and for the videos!
Dall-e is GPT, the others use diffusion models. The bottlenck is that you need enormous computing power/storage to train the AI (+many people scoring results), otherwise wiht all tools available, you probably wouldn't even need to be that much a of coder to do it.
I may sound completely dumb but does anyone have a clue whether you are able to make several images with the same character? Basically if you want a picture of a monkey riding a horse and then have a picture of that same monkey climbing a tree, then can the monkey look the same or will it just be a different monkey?
you can based on image, or try use same seed. it is good think to experiment with.
Thanks, this has been enlightening and useful. Please continue!
Thank you for your support!
thank you for your research
Thank you for your support!
but notNice tutorialng seems to work. Tried built-in content, and scarlett solo. What's the hardware you have? windows mac? special soft card?
midjournet has more often done what I've asked.
usualy the more iteration you do the better it becomes
the only thing it struggles with is creating characters that don't exist just from dettails.
but its perfect if you ask for an already existing character (needs to be popular)
As text to image translation, from personal experience, Dall-e best. I did multiple tests and it is always come up on the top. Quality or how detailed image, this is different subject.
Lines up with my experience. Pretty disappointed with DALL-E 2 compared to midjourney. Even v1 midjourney produced better results for most prompts. Then again - when it comes to things like architecture or interior, it delivers waaay superior results. Might just be bias in training data. Midjourney seems to heavily favor "artsy" stuff while DALL-E seems to have much more photography input.
yeah would be nice to actually see which data DALL-E is trained on so you wont waste tons of credits just to figure out what its good at and whats not.
wtf is that profile picture loooool
@@maiskorrel they probably won't tell you. In their shoes I'd use pretty much every image I can scrape off the internet - which they, of course, didn't license.
I don't agree with this, photography and photorealism are my genres and I've had excellent results with Mj
Midjourney from my experience is the best you can use, especially with the --testp and --upbeta prompt you can use now. Some of my best photos are coming out with that prompt
yes, they are top, however, I do like inpainting and now outpainting in Dall-e ua-cam.com/video/sS1_fPuSmGc/v-deo.html
I'd think that the first company that nails the standards of quality in an image, AND has a full GUI and/or standalone application will win the lions share. These UIs right now are unacceptable. I assume it's necessary at the current time, based on how things render. Also, I'll be honest, I'm not a fan of paying monthly for these kinds of services. It's WAY too popular to do that these days...I mean, we all have like 20 different monthly subscriptions, it's getting out of hand. Imagine how Zbrush took the idea of digital sculpting and went ham with it? The first text AI generator that does the same thing and is either a one-time purchase or an extremely fair monthly fee will become stuff of legend.
agree
Midjourney is already working on one
Well, the one service where a monthly sub makes sense and is justified, is AI art generating tools.
Do you think you can just download midjourney as an app and then render this on your own computer?
Do you have any idea how much data needs to be processed for stuff like this? Those data centers need to
be paid for and of course they will want to make money out of this.
The truth is, the best AI service will also cost a lot. A lot more than 30 bucks a month for UNLIMITED renders. Especially once we get to the point where we can add more and more parameters to get less random results etc.
An excellent comparison video 👍
Glad you liked it
Conclusion?
Which ones are the best for 80s dark fantasy?
Night Cafe is the easiest to use, but pricey
i did check it out, reminding me wombo
@@Geekatplay you gotta click on Advanced setting and the you can apply many other settings to it. It produced results similar to Midjourney. Also you get better results with start photo.
Great comparison of these AI zrt generators.
I liked the video just one comment about the 10000 robots. That’s an enormous number that would never fit in a forest level view. IDK how useful it is to give prompts that are essentially impossible to generate accurately? Thanks again for your video though, overall it’s great.
VLAAAD !!!! Greetings to John :D
Hello, hope all going well for you.
@@Geekatplay yeah thanks for asking - awesome presentation about A.I !! Hope you doing fine too
Hi from Argentina. Thanks a lot for your video 😄, I´ve suscribed. I´m trying to develop a complex story for a TV serie. Here in South TV producers are not very creative so I´m writing it as a novel. But voilà! IA maybe enables to present it as several comics. Which software, d´you people think, would better to narrate a story with regular characters who change their gestures and positions and clothes, and scenarios in which you adopt different points of view?
Sounds great!
could I get stable diffusion discord server link? thank you
I prefer Midjourney
same
theres an ai art app from wombo called "dream"..its not as good but its still cool
stable diffusion is 100% free and can be run on your own GPU and produces results way better than wombo/craiyon/etc...
i used Wambo a lot before, but it is has a lot of limitations. However it is very nice tool.
@@Geekatplay i agree. i wonder if it might see some improvements in the future
As of late August 2022 - Midjourney produces much more aesthetically beautiful images. Probably because their machine learning (the millions of images they've fed in) was created using more highly-crafted and artistic images. So far I've not been able to get anything remotely realistic looking out of DALL.E - it always looks like a messy painting. There's no real way to refine the results either, to change lighting or color for example. The only option is to create new areas of the image.
yes, when they switched to version 3 , it was improvement. new models works even better.
its possible to create for example something like that : bird sitting on the tree and looking at fox. but on bird's eye view.how to use perspective to create pic? :)
yes
@@Geekatplay how? I try and try and ai dont get this,, perspective,, discribe,,. U_u
Midjourney is a clear winner!
so far ;)
When there is free trial, there is creativity..
midjourney easy from imagine to result but i prefer for stable diffusion local gpu with some prompt tweaks , you can really get great results , and of course there are no limits when trying as much as possible . if using midjourney 200 rendering is very easy to spend
yes, I do like stable diffusion, but it is not for everyday users.
For me, it’s either DD or Stable (quality-wise)
How did you upload that photo of a woman on Dall-E? It doesn't let upload photos of people :/
i just uploaded as photo. keep in mind, it is photo i took and i have full rights to use it.
@@Geekatplay yeah, I tried the same with a photo of myself. So I'm sure i had the rights :)
#TeamMidjourney :)
Disco Diffusion is really easy install and use through Visions of Chaos.
i am using google colab, and it is easy as well,
@@Geekatplay Colab is slow, can't use your own a100 80GB GPU on colab.
I think some of these examples suffer because you treat all AI models the same. Take for example the fast food monsters.. Some tweaks to the input and I got very similar results to midjourney with stable diffusion.
that is true, you need adapt text prompt individually. However, I just created text string and used without adaptation to any service. and it was my bad english ;)
AWEEESOME
thank you
Why didn't you use the exact same prompts across the different programs? (I know you said you did, but you hover over the images etc many times and can see that you change/add things etc.)
It's not really a fair comparison then... (I know dalle costs more to run many etc. but if you're trying to do a direct comparison, you should come up with the string before testing it anywhere, then run them all, then compare side-by-side with no modifications)
I do, however, do to some command syntaxes, it need to be modified. Example /imagine (midjourney) !dream (Stable Diffusion) ...etc. If you check on text it self, I was copy/paste between all services.
For Dalle-E size pics, Disco Diffusion needs 5min, but is way more "artistic" and versatile.
Dall-e using CLIP not DD to create images, pluses it more realistic, minus much less details then DD based system.
@@Geekatplay I only clarified time/result. Looking at the video you would think that an image takes 1h on average, but on average you don't get much benefit from exceeding 15min/pic.
Night Cafe is a good one also
Midjourney. Unless you want in painting /outpainting.
Stable Diffusion also producing incredible results. I think portraits looking better in Stable Diffusion, then in Disco Diffusion
Stable Diffusion is free...whats to compare/complain about?
Disco Diffusion with Google Colab also free. I am comparing how they work, what they do and how much cost, does it worth to pay.
Ten thumbs up!
thank you for you great support ... or it is soooo bad?
@@Geekatplay it’s great. I learned so much!
These corporations must really hate artists ...
At least u can use mid journey. Dalle2 still isn't useable for the most part
Honestly after all hype I was expecting more out of Dall-e 2
Did you get on their waiting list? - I only signed up 2 days ago and they've already accepted me.
@@darrellm9915 wtf it's been a week for me and no answers yet :(
Personally, I love MidJourney for its creativity and fidelity, and I love Stable Diffusion for its controllability.
If I type "a person with red hair", MidJourney will give me a beautiful image, but red will be everywhere and sometimes not even in their hair. Meanwhile, Stable Diffusion will give me a person with red hair, but the image itself might look strange or unnatural. So, imo, the challenge for MidJourney is getting exactly what you want, whereas the challenge for Stable Diffusion is making it look aesthetic.
Using both side-by-side, they're both amazing in their own ways. Since others can create their own models off of SD, there will likely be specialized versions of SD in the near future, such as one for realistic faces, another for beautiful scenery, and another for manga / etc. Overall, I'd say MidJourney is definitely easier, more reliable, and beginner-friendly, but Stable Diffusion has greater potential and is more directly tied to your ability to craft and test prompts.
Why I can't go in the app?
what app?
@@Geekatplay Sorry for the Question - I solved - I asked for the MidJourney image browser... THX for the answer
Disco Diffusion..... the settings I use take 25 minutes to create an image!
very nice, what notebook are you using?
@@Geekatplay I was being sarcastic. Wish it was faster. I'm using Google Colab and paid for the Pro (not Pro+). If the images were a bit more coherent I'd pay for the Pro+.
I tried some of my "unusual" prompts during my free session of DallE2 and wasn't impressed. I really want to use Midjourney but hate not having my own private room unless I fork over another $20 on top of the $30 I was willing to pay. That's beyond absurd and not worth it.
Hopefully I can get into the stable diffusion beta.
@@henrythegreatamerican8136 I have standard subscription ($30) and I can go private with a bot when using midjourney, (the only thing is that your images also show up in some public chatroom in the midjourney discord that is visible for other Midjourney payed subscibers, which is barely looked at anyway.
I really would recommend Midjourney, it's my favourite AI generator at the moment.
Midjorney is the best by far.
it is very nice, but other coming close
6:15 lol
yea, sometimes you can see funny stuff, people render.
deteeels
RULE 34
The control and modification, finessing abilities are massively lacking in generative art. Can't wait for adobe to slap a tools set around Generative art.
They really need to show you what it's doing and give you the abilites to partially editing those processes. Also partially select areas and be able to rerun generation on those areas.
This does not belong into the hands of the likes of Adobe. They already did enough damage to the traditional market.
why compare something with DALL-E2, if access to DALL-E2 is not easy to get, if at all possible for an ordinary person? Why don't you start your video saying that - you are unlikely to get access to DALL-E2 ..
Therefore, many thanks to the creators of MidJourney, who gave open access to anyone and even free 25 times to try. Or for $30 an unlimited number of times. And while other developers give access to their products to a select few, depriving ordinary people of the opportunity to test and compare, they do not deserve any attention and let them go to hell with their selectivity.
I am ordinary person, was waiting for Dall-e access for about moth and half.
Calling typing what you want is not creating, and you are not a creator. The real creators are slowly cleaning up their desks getting ready to leave. Shit happens.
So you calling writers, that they not creators? I know few single sentence poetry writers.
@@Geekatplay Dude, lets be serious... 99.99 percent of ai image generators users will be prompting what they want in more or less precise way, not in a poetic way.
I'm mostly put off by the tiny resolution AIs produce.Another thing that really puts me off when you either should or actually have to go through repulsive spammy internet cesspools like discord. Why the actual eff are these e-sh*tholes inserted into the porcess?? Pls...
it is have room to grow, and getting better as it is learning. I use Topaz Gigapixel to upscale.
Dalle is not good
Dall-e good on transcribing phrases to the code weights. one of the best
Dall E is the best at realism and understanding context, it just isnt very creative, so its results arent necessarily "pretty" like the aesthetic you get from midjourny. I told dall e to generated a photo of a fox with a cigar in its mouth and got a hyper realistic, 99% accurate photo of a fox with a cigar in its mouth. I do the same thing on midjourny, and it was incapable of comprehending the context enough to even make a coherent image. They were "pretty", but you couldnt even tell it was supposed to be a fox.
I think Dall E 2 and midjourny are both good at doing different things. Dall E 2 is like a printer while midjourny is more of an artist.
This is BS!
i as mention in the video this is personal, subjective opinion.
They're doing what they do best.