RIP Midjourney? Open Source & FREE AI Art is About to Take Over! DIRECT COMPARISON 

An open source, trainable AI I can run on my own system is the holy grail for me.

Timestamps courtesy of Deciphr AI 👨‍💻 0:00:39 - Stability AI's SDXL 0.9 as a serious competitor of Midjourney 0:01:28 - SDXL’s features and real-world applications 0:02:23 - Comparison of SDXL 0.9 and Midjourney through image examples 0:07:17 - Comparison using ClipDrop 0:09:03 - Comparison with different prompts 0:12:21 - Comparison with a creative prompt 0:13:47 - Comparison with a character prompt 0:16:26 - System requirements and availability of SDXL 0.9

These AI tools are getting so crazy, that if you don't use them to your own advantage you might lose your job. To be fair it's good for content creators (especially if used with Famester, which is great for increasing presence on social media). Create content, make it go popular straight away, I like it 😍😁

The ridiculous censorship on MJ, along with the monthly cost, led me to finally learn Stable Diffusion and I'm so glad I did. I've learned so much about how text to image works, how to craft prompts to match up with models,etc.

12:12 If you look closely at the reflection in the eye on the right, the model decided to place the character in a cave. Truly amazing attention to detail.

I feel like midjourney 5.1 and 5.2 are still way easier to get a wide variety of excellent images, but for getting to a specific image you have in mind, between various models and extensions, SD already takes the cake. MJ is more fun casually or if you have a vague idea you need a starting point, SD is better for drilling down on a concept you want but don't have the photo editing chops to execute.

Midjourney isn't a raw model, it's a whole system, for example it could edit you inputs to produce what it thinks is "better" outputs and you wouldn't know it, it also has certain settings that you can't change that makes it more consistent but more inherently limited, it could even be picking and choosing which model to use to produce the most compatible results, it could even be using 3d overlay! stable diffusion is a raw model on the other hand that can be used to make other open source systems, that's why the specific stable diffusion models just the ones that are actually being used, currently are specialized models that can surpass midjourney in many cases and it is widely used.

You can actually use negative prompts with MidJourney, you just type "--no " at the end of your prompt. */imagine blue bear --no brown*

This is huge because now even novices can produce high quality SD images. Most people who try out web-based SD apps give up very quickly because the image quality is really bad compared with the images shared by expert users. Now anyone can produce high quality images with a simple prompt!

Let's hope. I can't afford midjourney

It's all about censorship or no censorship. I'd rather less quality with no censorship than amazing quality with full censorship

nah since loras introduction SD come ahead of MJ, easier to change model/lora for specific use, with often better results

This video is the little push that you needed to reach the 200 k, congratulations in advance ✌️ Edit: here's my honest opinion about this open source image generator. I think there is still a huge bias involved in the comparison tbh, midjourney looks a lot, and I mean A LOT better still tbh

I actually like the SDXL lemon better: it's actually the shape of a lemon. The MidJourney one doesn't even look like a lemon: it looks more like a stumpy yellow hotdog.

Stable Diffusion + "mods" has been blowing Midjourney out of the water for a at least a few months now

Looks like a great option compared to the old standard models for new users but I think I'll be sticking with my current setup and 1.5-based models for now. I can already get photorealism with my models and extensions and I have models that can produce images that feel like they were actually taken by a human whereas this model can produce photorealism but it tends towards looking like studio photography with lots of photoshop and compositing. It looks real but also often very contrived which isn't always a bad thing but I tend to prefer a more candid look. I can't deny the quality achieved by Midjourney but I can actually get what I ask for with my current models and with a similar level of quality some of the time vs Midjourney which tends to decide what it wants to give you with little regard for what you're asking for in terms of stylistic direction. The only thing my current workflow lacks is consistency. 1 in 10 or 20 images will be beautiful and then the others will be puddles of writhing flesh and limbs longing for the sweet release of death so maybe models training on SDXL can improve in that regard. I wasn't surprised by SD's idea of a 35 year old woman, if you ask it for anyone over 30, expect to get an elderly person and you get someone that looks 18-20 if you ask for younger but some of the models improve on that.

I feel like the MidJourney devs knew this kinda tech wouldn't take long to make its way to home use and actually be comparably good. It's why they charged the 36 dollars while they could.

A very good hardware, software and open source analysis. Unfortunately I can’t say the same for the artwork comparison. And it doesn’t just come down personal preference. Based on the examples shown in this video, Midjourney is still notably better at subject /background composition, including focal point and control of eye movement around the canvas. It has superior sophistication of asymmetrical balance, better proportions, perspective, light and shadow, colour, textures and of course detail. Perhaps the biggest difference is the ability to distinguish physical property behaviour such as how hair flows and fabrics drape. The shirt on the women in grey is exhibit A. The Stable illustration was a brutal attempt. Midjourney is still weak on hands. SDXL may be better in that regard but I didn’t see enough examples to make that determination.

Really looking forward to this release, thanks for the reminder. I may be one of the exceptions to the rule, and it's all subjective, but I've never rated MJ as highly as many others do. Sure you can get great images from it but you can get them from SD too. I don't care if MJ is tecnically better, the pros and cons for each don't make MJ worth it for me over SD, and that's before SDXL lands. Each to their own though.

I'm always interested to see what hive moderation says. On a photographic style I generated it reported 99% AI generated, but not as generated by SD. Instead it came back as 98% likely to be midjourney, which I think gives an indication that SDXL is indeed in MJ territory. When hive update their models I imagine they might then be able to discriminate, but XL does look good and excited for 1.0 and general availability.

Open Source Devs are modern Heroes.

Stable Diffusion was already worlds ahead of Midjourney. Midjourney is only for people who don't want to do their homework and actually learn to use Stable Diffusion and its many game-changing extensions. Midjourney was never comparable to what you could do in SD 1.5 with just Controlnet, not to mention all the other community extensions, models, etc.. And with SD XL launching with support for Controlnet, Kohya, and other community extensions, it's gonna remain unbeatable because it's open source.

SDXL is 1024 native :)

I'm glad you finally came to your senses👏👏👏👍👍👍

He's competing with Midjourney now, and he's free ❤️🎉

I was a big fan of MJ since the beginning. In the past I would see and try Stable Diffusion and would get really bad results. It turns out the original SD models are actually not that good, or may be they are just too broad and basic. After I discovered custom models and loras, all of that changed. SD with custom models, loras, controlnet and other extras is just better in every way. It gives you extreme control over everything you create. Also the upscaling and resolution of generated images is way way better. MJ still have very limited resolution. Something else with MJ is that it is very much focused on close up portraits - it uses blurred backgrounds, shallow focus of small area (the face or part of it) to imitate quality. It does well with lighting too. But when we talk about details, MJ is just still very much struggling. Try getting an animal in forest with flowers, mushrooms and grass for instance. You will always get perfect animal and everything else will be just smudges and deformities or just super blurred background. This is problem when upscaling because the details are just not there. What about inpainting and outpainting? MJ is very much super late with these tools. They released new version with zoom out feature that is basically useless, nobody asked for it and is good only for you to play for fun for and hour or two. The most asked thing by the MJ users is the resolution, option for real upscale and inpainting. To this day I can glimpse an image and in 2 seconds I already can tell if it came from MJ. SD on other hand have so much models and styles it can do virtually anything, and new models are released every day. The reason so many people stick with MJ in my opinion is that it is just easier to use. But that is what makes it so limited. Yes, launching SD involves a lot of steps, but it is totally worth it. I use it on a cloud GPU for cents per hour and can generate images in 10 seconds. I can generate 8 images simultaneously. Everything's working super fast. The user interface is a little intimidating for some people but in reality it is very easy and useful. I have 300GB custom models right now and I can run a single prompt through all of my models and get comparison of the results and pick the one I like most.

I tested this with my standard prompt test set using the browser interface. The quality was good as advertised, but occasionally the AI failed to understand altogether what I was asking it to do, even in cases where its competitors succeeded well. The human hand may still have six fingers. Photographic images use disturbingly often a large camera aperture.

I guess it really depends on what you're looking for. When I use an AI generative art program, I don't WANT a hyper-realistic, photo like result. I WANT the more art focused result. To say that one is better because it give a hyper-realistic result instead of a more art-like result is ridiculous.

I'm relatively new to this, like less than four months. I never tried Midjourney, but most of the people I came across in the BlueWillow community may use that AI as their favorite, but they actively try every AI that they could use. Many of the users are coming from third world countries, me included. Part of the fun is make something amazing with the tools you have. Many AI bots are thriving in Discord, lol, probably there is a niche on the market for everyone.

Dude, there's no such thing as "Midjourney domination", only people who have no idea about the field might think so. I follow actual practical AI art very closely and make my own, and I have never seen anything even remotely interesting made using Midjourney, not even once. Everyone using Stable Diffusion, because it is just better

Interesting, can't wait until a few months after release when all the Loras and Checkpoint models get updated with this.

SDXL is nice. But imagine the custom models based on SDXL... Stable Diffusion will murder the competition.

All I want for Christmas is a locally run AI text to image generator that is either free or a one time payment with absolutely no censorship

These stable diffusion images have some great texturing. Every image had good looking materials

title is pretty accurate, was not expecting that, good show!

I can't wait for the full release

The real problem I find with Midjourney, it produces good quality images, but it seems to have a mind of it's own in the images it produces, in a sense, you have some control over the image you want but I feel Midjourney overrules a lot of what you are asking of it, with the aim of producing a pleasing image, and that's fine if you are after a random like image to the given prompt, but it's not giving you much control whereas Stable Diffusion does. If this new open source one delivers, and you have all the benefits that Stable Diffusion offers, whiles also being able to run it on your own PC, no restrictions and no cost, that would be a far better option then what Midjourney is offering for most of us, and I do feel that Midjourney monthly subscription is going to be hard to stomach for a lot of us when there are alternatives that are free, but if I was an artist, it's the control that would grab me, something the open source ones give whiles Midjourney doesn't, after all, Midjourney is good for creating random like images to the given prompt but it seems quite weak on delivering what you actually want, whiles also taking a lot of liberties on the prompt you give it, all in the aim of creating a pleasing image, but I suspect, more direct control to get the image we have in our minds to the screen is what we really want. Anyway, looking at the images shown in the video, both look impression and close enough to not matter, especially with a few slight alterations to the prompt, you can get closer to the result you want with either of them, that does offer a big advantage to this open source one over Midjourney, and once you throw in the kind of control that Stable Diffusion offers, I really have to wonder why anyone serious about creating good images, would use Midjourney when there is this alternative that offers you far more control and is free.

This is actually fantastic. I only tried midjourney a few times and for some reason it couldn't do my hyper realistic photo (for a novel cover) but this sdxx did it in the first try. WHOA>

What I want more than anything is a text-to-image generator that can use plain English instead of unintuitive prompts.

Thank you for talking about open source with this passion. I am so happy that I can participate in this AI revolution without ever needing to touch any closed-source software other than the f-ing nvidia drivers on Linux

Asking for Apple glasses is a bit hard as here probably isn't many images of them to reference to

Why do you all feel so strongly that MidJourney is better than Stable Diffusion, Adobe Firefly, etc? I mainly run SD locally & trained on my own images, or a library of client images when relevant. I’ve played around with Midjourney but haven’t found any reason to pay for it 🤷🏼‍♀️

Thank you for the video. I think the platforms' target audience should be touched upon as well. MJ's is not the same as SD's. Those differences are a very important part of why these platforms perform how they do, what they allow, how much control the synthographer has, ease of use, the learning curve, the pricing, etc. Good video, but I doubt either of them is going to RIP. They'll finesse the USPs and strengths they have over the others. MJ's going to have to keep a keen eye out though, that's for sure.

You can dream, but midjourney has no competition

Matt, you're always better than Santa in person because only you know how to make me really happy. Thanks, Matt. Your videos are real gifts for me. Do you want some cookies and milk, maybe?... 💀☠💀☠💀

Im very certain Midjourney can run on any standard PC with a good graphics card. Midjourney itself runs on Stable Diffusion. Midjourney is mainly for casual people whom cant afford a high end pc, and dont want to nerd out and tinker with the extreme complexity of installing and running Stable diffusion. Stable Diffusion has a huge learning curve, with powerful options, while midjourney is plugandplay with limited options. Im running my SD on a old 1080ti and its perfect; still able to make hyperrealistic art. Ive been using SD everyday since Jan 2023 and I still not even anywhere close to halfway efficient at using it. Casuals that dont use Stable Diffusion keep thinking Midjourney is its own unique AI art software, but all Midjourney is doing is taking the Stable Diffusion name and pasting Midjourney over the top of it. The only thing unique they do is train their own custom models, for the casuals that dont understand what that means, its basically like making a mod for a game and Stable Diffusion is the base game, and Midjourney is the mod maker. People that use Midjourneys product are basically paying the Midjourney team for their custom models, most of which are free opensourced on CivitAI. SDXL isnt opensource yet, there is literally no mods for SDXL on the most popular Ai art Mod site Civitai. Its also highly censored which is why. If anything midjourney team would profit the most from SDXL, not the opensource community, because the team already uses stable diffusion to run all their operations and midjourney itself is censored so they'd have no problems with SDXL coming prepackaged with censors. Until StabilityAI starts charging commercial license use for their Stable Diffusion product people like Midjourney and hundreds of other AI art generator websites will piggy back off their hard work completely free of charge.

I cannot wait for this. SD has always been my go to for most AI stuff but this model is really good.

Can you do a deepfloyd vs SDXL comparison video?

This is definitely better than some other alternatives. I keep trying to create a realistic image in Mid-Journey of a person wearing a white t-shirt, but no matter what, it always puts a collared, button up short on the person. This one actually worked and put a t-shirt on the person, and made a very realistic person as well. I did noticed it still has trouble with hands, where Mid-Journey has definitely improved in that area.

I mean some of the community trained stable diffusion models and negative textual inversion prompts already allow for stunning images with coherent hands and honestly sometimes completely indifferentiable output. Since you can't train textual inversions on Midjourney, I think a fairer comparison would be Midjourney vs fully modded stable diffusion rather than out of the box stable diffusion.

about time, we need to democratise free AI MOREEEE

I consider both essential. They have different feature sets that compliment each other very well. Mj has the prompt generator and a new blend feature, SD can evolve reference images. MJ is king when it comes to fantasy style art, subject placement, perspective, balance, etc. I like to do stuff that looks authentically hand painted, and that's where.mj falls behind. SDXL 1.0 in dreamstudio rules when it comes to that and that's what i like to do. But i can generate amazing images in mj and then evolve them toward that authentic hand painted look in SDLX.

Stable Diffusion and other open sourced solutions will still have their place because they offer customization for those that need it. For those that don't want to mess with "all that", there is midjourney.

I'm actually blown away by how impressive this is. It's as good as MJ, maybe even better.

The only reasons why I still use Midjourney are: - Custom aspect ratios - Define the tone of an image kind of consistently - COMMERCIAL usage (SDXL is currently under non-commercial, research-only license) If those are possible with any other software, I'll quit ;-)

8:00 btw, the resolution of SDXL is not 1024x1024 but 2048x2048

I tried it on ClipDrop after watching this video. It's pretty good.

No, what a bummer. I am running windows 11 with an AMD Radeon RX 6700XT (with 12GB of VRAM). So I can't run this model😔. Or does someone know a way to still run it with my specs?

Super! Very nice! I am looking forward to it...Open Source rules!!

Loved your video! You might find Stylar AI interesting for logo and interior design. It's still in development but worth a look.

One annoying thing about Midjourney is those "Midjourney Fingers" lol. BUT!!!!!!!!!!!!!!!!!!!!!!!! Midjourney has prompt weight.....Almost like the Niji Waifu Function....A lot of people dont know that withing any prompt(s) you can add weigh/importance by adding ::# After your prompt, hit space the 2 Colon and then, NO SPACE and number....The number ='s importance.....You can do this a MULTIPLE of times throughout a prompting sentence. Also, Midjourney DOES NOT recognize Commas....They are just used to neatly divide all prompts and section out your sentence.... Try it for yourself....Run the same prompt set twice+........Once with your regular prompts and, then again with a space and ::# (2 Colons) and then a number (No space)...Remember...The number ='s importance/emphasis in your prompts

How can I make it generate my face into one of the ai generated pictures.

Yo guys how to put my private nsfw LLM on external gpu with webgui/app for conversations?

This is extremely promising. I hope this will also run as awesome as SD does on my M1 Max.

It's a serious contender it has some great features built in like background removal and I'm getting closer to what I want quicker

Wow, the rate of advancement is really amazing!

Thank you for sharing about it. Waiting for it to be released…

A better prompt for SDXL for the glasses would probably be: "wearing augmented reality glasses" or something. Apple vision pro is too specific, and dependent on what text prompts are recognized .

I was just thinking of updating my image software. What wonderful timing!

is not bad at all is improving and will eventually compete with Midjoruney

The timing of this is so good

If mid journey can just get body parts correctly and really do what you input when it comes to multiple things

... except the lemon for the MJ image doesn't look like a lemon - looks a little like a short banana. The one on the left - the Stability image - actually LOOKS like a lemon!

After watching your video yesterday, I decided to try it out and I have to say, I'm hooked. I like it so much that I paid for the pro plan...thanks for this

Bruh with all due respect, if the world world was ending, I'd agree. Midjourney has some shit coming that will revolutionize the game

I like the UI buttons of stability and the extra tools ie uprez etc...and just bought a subscription for Pro for work. But I have to I've been getting better images with Midjourney ( also subscribed ) overall with less oddness on second regens. The resolution of Stability though is better in some instances and I like that it's just a site and not discord, very fast with the tools. I can see Stability progressing fast and catching up quick to Midjourney.

The thing is that sdxl that you tested is just a raw model without any add-ons like vae/lora/etc. And it is still not v.1

5:00 Correction: Yes we do really get negative prompts in MJ. Have you even used it? 🙄

*getting back on my chair after falling down laughing at your absolutely "random" comparison prompts*

The "v.s. MidJourney" marketing ploy is tiring. As a comparison you see so many RU-vid videos of people asking "were the 'Rolling Stones' - 'Queen' - "Led Zeppelin" - "band x" better than the Beatles? The common ingredient... The Beatles. MidJourney is its own animal and so are the others. Claiming to "dethrone" MidJourney is just re-enforcing how popular and good it really is. I'm all for trying other A.I. engines (and I do) and there are plenty of good free options out there but last time I looked you can legally use more than one. The current MidJourney model doesn't even use Stability (to my knowledge). Comparisons are weak if not irrelevant. Yeah Midjourney charges and can be considered expensive but when I look at the results and the costs for anything with results that are near comparable it's well worth it. I'm not saying you can't get good results with other models, I'm just saying it's silly bordering on manipulative to keep pushing this "better than MidJourney" trope over and over and over. Build a reputation on results over a period of time and the followers will come.

but when i loaded sdxl 0.9 into my stable diffusion it says it cant be used for commercial use. thats not really free

those stable diffusionXL images didn't have a negative prompt. I bet you can get even better images if you could input one.

I always dreamed of becoming a concept artist. So glad I didn't bother pursuing that career 😅 nobody's paying you to type in a prompt.

Testing for Clipdrop is limited to 400 images a day, or ~100 prompts. The price however is very reasonable. I subscribed for Relight alone.

Excellent content kind sir. Gladly subbed and looking forward to more content

I appreciate you showing me the licensing agreement even though you didn’t mention it 🤓

It's not unlimited, there is a limit, and it sometimes can't follow my prompt. Especially if I ask it to make art of a specific person or place.

midjourneys selling point is jack of all trades, being multimodal (even the multiple versions can be remixed into eachother) also midjourney encourages learning how the AI understands you instead of getting it to understand you. To put it in perspective, midjourney favours eg; if you prompt red and you get blue you now learned that red gives you blue and you roll with that remember that.

I’d argue the silver-haired woman by MJ looked exactly 35, a healthy one. But the SD one looked not a day under 50

so.... things like control net won't work on SDXL out of the box, right? Seems it has to catch up to a looooot of features that make SD1.5 so awesome to use right now

What i am missing in MJ is real capabilities to create replica of portrait (with training) and adapt to prompt as i have seen some demos as controlnet and dreambooth but better as MJ should be able to generate with minimum effort. Their upload image and blend doesn't work so well. Would pay for it

Nightcafe has that new model now.

Yes, I have tried it in Clip Drop at length, and it is a major improvement over the previous beta SDXL which was used, but still no AR or img to img. But why OH why more freaking aliens! Aren't there any good looking humans left?

Midjourney works better tokenized. Otherwise it's a beast to master. You will have to "BEND" your mind a bit and tweak. Tweak. Tweak. But being able to use similarly styled Midjourney Prompts with SDXL definitely is definitely going to be EPIC!

midjourney has alot of style-options, dont know why you would say that they havent

Will copy wright issues damage Midjourney? It reminds me of the start of the rap hip-hop industry.

Death to Midjourney! Long live Stable Diffusion!!!

It's funny to see: With cameras many people say "I need higher resolution". With AI it's "an upscaled 512*512 image is good enough". That's 0.26MP. Of cause, there isn't much that is better in the AI world, yet. Still, it's funny to see the difference in perception.

It cannot be installed on a Mac right? I tried to install the libraries but one of them was only for Linux...

ok so how come I'm seeing the exact same images for this model that they're showing for the MJ update??? The alien, the wolf, and the hand holding coffee are all being shown as MJ generated images in other videos on YT..