This is a MAJOR Win! Open Source & Uncensored: SDXL 1.0 is OUT!

[@Terry-Adams]

Watching this with my 80 year old dad and he is finally trying out himself on playground

[@arcader9184]

Ha ha that’s beautiful. Imagine see the world back then to now, that’s awesome ❤.

[@ThymeHere]

Thats amazing!

[@eb-dv8ts]

So cool

[@raybrandt]

man, your content always makes my day.

[@MattVidPro]

Thank you!

[@MysteryGuitarMan]

Thank you, Matt! Glad you're a fan :D

[@PlanetCat]

No replies? Lemme fox taht.,

[@coloryvr]

I installed it locally yesterday in 1111, tried it out the whole day today... and....... I am totally blown away! ...very, very WOW! I hope this will run with Deforum soon! Happy colored greetinx!

[@PunxTV123]

can you do tutorial

[@gameswithoutfrontears416]

It's a really great model and will only get better with the community models that come from it.

[@gameswithoutfrontears416]

I haven't tried XL 0.9 so I can't compare

[@realcinnamongold]

@@gameswithoutfrontears416it's a WHOLE LOT better

[@ozerune]

Give this maybe a month and hopefully we'll have as much control over XL and we do with 1.5 and it will be the greatest thing ever.

[@ChandravijayAgrawal]

another nice and useful video no clickbait

[@DiceDecides]

it's mindblowing how it can generate text so well now, most signs are perfect now, I remember when it used to be complete gibberish

[@flareonspotify]

The down side is if you only have 8gb vram switching back and forth from the refiner and base model is time consuming because they have to redownload every time you switch to the other one because there’s not enough room for both

[@mazmahjoobi]

Eyes and limbs still need a lot of work. You still often get funky eyes and extra limbs, or funky fingers. Midjourney is still way ahead in regard to these. But, I do agree, you can get some really nice images that are different from Midjourney.

[@nuvotion-live]

What about with negative prompting? Any better?

[@mazmahjoobi]

@@nuvotion-live Yes, it gets a little better with negative prompting. But still not great.

[@abandonedmuse]

The paid version of clip-drop the year is out for $49 through iPhone. I totally got it. I love their tools. I also have Midjourney. I have both because i want different settings for different things.

[@DrewMedina]

SDXL is working great in “Visions of Chaos”, thanks for the vid!

[@vredrick_]

🎯 Key Takeaways for quick navigation: 00:00 🎉 Stability AI just released the open source Stable Diffusion XL (SDXL) 1.0 image generation model. 00:58 🖼️ SDXL can generate high quality 1024x1024 images with realistic details like accurate bokeh blur. 02:20 ✏️ SDXL seems to recreate text better than other models like Midjourney. 03:56 💻 Being open source, SDXL can be run for free on your own machine if you have a capable GPU. 06:42 🐈 Test images show SDXL generates photorealistic results like a cat dressed as a chef cooking. 10:34 📸 The AI community has produced impressive celebrity mugshots and photos using SDXL. 12:39 📝 SDXL seems very good at generating coherent text on signs and images. 14:58 🆓 Open source models like SDXL will push commercial AI companies to improve their tech.

[@Vpg001]

Ooh, would love to see some comparisons

[@1EAS1World]

6:43 wow who would have thought a cat would be ready to make him a delicious meal

[@ChristianFuchsBlues]

Thanks first of all for all your work, Matt! I find, SDXL ist a little hyped by the KZbinrs. All video titles once more went overboard like it's the second coming of the Messiah. So I went straight over from Midjourney to SDXL to find is still has far too many hands and limbs wrong (and eyes, expressions etc) . And if I ask for something that is not as common as a tree ( I tried piano) it is massively lost. "A piano player at the piano" on MJ gets 4 out of 5 generations usable, SD gets maybe 1 out of 6 just OK. The good examples shown by KZbins in favour of SDXL, so I presume, must be cherry picked. Yes, Midjourney doesn't do what you want, but what comes out is more correct, a level more correct, I must say.

[@MonzonMedia]

Regarding Playground SDXL is 1.0 👍 SDXL is very promising!

[@nahiddotai]

Very impressed with the text generated by SDXL, I thought it would still be too early to see this in an AI image generator! SDXL being open source is such a huge win for all the AI art generators using SD, they're going to get a MASSIVE upgrade

[@TurbineFlyer]

It took me 15 minutes to generate 1 image on a1111. And I have a 3070. ComfyUI was WAY faster for me.

[@thFaust]

you should run with --xformers and --medvram.. generation speed is okay then, but still slow. For me, sdxl is not worth it yet, with 1.5 and img2img upscale i get more flexible and better results so far.

[@lucamatteobarbieri2493]

Did you try InvokeAI?

[@rolarocka]

Similar here with the 3070, it just takes really long for one picture

[@asher9888]

Takes 24 seconds per 1024x1024 image on my 3060 with A1111. I don't even have "--medvram" or "--lowvram" enabled.

[@umcarafilipino]

SDXL faces look strange, but it can do chromatic, metallic textures for objects, which I like.

[@drowsy4400]

Download realistic models😉

[@emoneydatruth1]

Open Source will win this AI race.

[@lucamatteobarbieri2493]

I prefer open-source. The big problem is when an AI becomes so powerful that it can destroy humanity. For example no sane person would want open-source nuclear weapons nor open-source protocols to create deadly viruses. Be careful for what you wish for.

[@mycelia_ow]

@@lucamatteobarbieri2493 I think that's a reach

[@Rynix_]

this new clipdrop is amazing!

[@tvkhaled7067]

Finally... I have been waiting for this video for so long.

[@plazzmusic]

so long is 2 days now ;D

[@stephenrodwell]

Love your videos! Thanks Matt! 🙏🏼

[@southcoastinventors6583]

Matt you need to generate Minecraft art of lemon guy relaxing on the beach by a pile of diamonds to celebrate the occasion. Great video as always

[@zainkhalid3670]

I would definitely want to see a further video.

[@hjups]

Just to point out, closed models are not only to enrich the stakeholders... That may be how the bigger companies see their models (i.e. Adobe and Microsoft), but for smaller companies like MJ it's likely to fuel growth. The cost to train one of these base models is enormous, especially as they keep getting better. So if a company invests those resources in training, they should be able to see a return on the investment (or at least break even to justify further training and improvements). Making the model fully open source would hinder that ability, so the only companies that do so are either hemorrhaging money (like Stability) or can afford to write off the training as effectively "petty cash" compared to other revenue sources (like Meta). Obviously open source models are preferable, but I think we also need to understand the economics behind the decisions made by players in the field.

[@jopansmark]

OK, Midjourney fan

[@hjups]

​@@jopansmark Not a Midjourney fan, I don't even have an account with them. But I'm a ML researcher that understands the mind boggling cost to train a model from scratch. The cost to train some of these models is higher than the GDP of some countries, and that's only the raw hardware cost, not including the datasets and human capital costs.

[@rh906]

Thankfully for midjourney, no one using their service has principles. Especially as much as they badmouth (surprised they don't ban) anyone that mentions that their June icon is still up months after June.

[@davidpurple3698]

Thanks a lot for a great video. How can you download to a Mac M1 ? Seems a little too troublesome now, and as you mention, Im still waiting for it to be easier - like 'click' and install auto.

[@michealhsujejdud]

Hi dude your videos all very useful for me , i have one doubt about SDXL 1.0 AI images i can use commercial purpose

[@njrivetelite]

Oooh I'm gonna DL this.. It would make nice use of my 4090

[@Hasblock]

I fuck with MattVidPro heavy 🔥🔥🔥 one of the best AI channels out there today. I especially fuck with his emphasis on open sourced technology. keep at it my guy

[@chariots8x230]

I like the fantastical stuff. I hope the AI becomes even better at creating images for comics and graphic novels. I prefer these styles much more than I do realism.

[@KolTregaskes]

4:05 Midjourney uses its own model, not SD. :-) I think v6 is specifically targetted for text interpretation and that is due any week now. So it will be interesting to compare SDXL 1 to MJ 6.

[@spideyguy3454]

Magespace added SDXL today!

[@rh906]

Yeeeeeeeeh, when it can work on the same hardware as 1.5 I will get excited.

[@jozsefbenko7852]

Wow, thanks!

[@justinwhite2725]

We just need control net and a few other quality of life things we've gotten accustomed to in 1.x

[@ZhafranAzzamir]

Please make a video how to install to our local machine and a minimum specification for ram, vga etc...

[@truejim]

It’s interesting how speaking with “radio voice” has taken over KZbin. I remember when people would speak with a radio announcer intonation just to be funny.

[@Mike-ne6nt]

Is it possible to use SDXL 1.0 without any limits through the official API? (If anyone knows how to do this in php/js, please contact me as I need someone's help :))

[@jooptablet1727]

Just ordered a 4090... Impossible to use with my 8gb VRAM

[@noobgod3216]

The 8gb version or the 16 one?

[@jooptablet1727]

@@noobgod3216 24gb

[@seniorp9444]

Next level

[@executivelifehacks6747]

Billions and billions of parameters!

[@executivelifehacks6747]

I've been making similar comments... is he now leaning into it? 🤔

[@questor55]

JARBALLED

[@bazebuster8330]

I did. also See, that it was released 🤩

[@AnthonyGoodley]

So when we download this does it come with everything we need from the get go such as the training data? So once it's properly installed and setup it will just work, right?

[@noobgod3216]

How are you verified with 1k subs

[@AnthonyGoodley]

@@noobgod3216 I don't see an answer anywhere in your response to my question. Your responding to a question with a question.

[@noobgod3216]

@@AnthonyGoodley ok

[@abraham9301]

@@noobgod3216 It's google+ verification it's very old

[@12uniflew]

Bro i cannot believe you didn't try the infamous lemon on a beach prompt!! Thanks kids are crying now

[@jcorey333]

You should cover the Civitai challenge to build Loras and base models based off of sdxl

[@seanrodrigues8184]

It's pretty good with photorealistic stuff. Anime... Not so much.

[@jonth3978]

like the mugshots... was going to do one of Trump but... why AI when you get the real deal

[@drendelous]

beatiful minds, please share where to look how install stability ai

[@drendelous]

oh Matt is talking sbout it in the video srr

[@GS195]

Me with models and LoRAs that run on 1.5: 😞

[@Raulikien]

1.5 moodels are still way better until they fine tune the new XL and release stuff for that

[@diadetediotedio6918]

@@Raulikien They are not "way better". They are generally better because of the more control, but they are objectively worse in the sense that they do not have the same generation context size.

[@eg3730]

lets Go?!

[@andywest5773]

The image at @3:30 looks absolutely amazing and yet... the man has six fingers. So close, but still not 100%.

[@ryglitheegg]

Nice

[@WrldOfVon]

That ain't Drake that's Dramie Foxx 😭😭

[@1salacious]

It's great, but man can it tax your gpu. Some 2x upscale's have hung my 3090.

[@IchiOne-b4u]

The Predator!? What?? LOL What a dork.

[@kenrock2]

I realize SDXL uses more RAMs than usual, it utilize almost all my 16GB RAMs when changing any checkpoint

[@hypersonicmonkeybrains3418]

Hell Yeah.

[@malarkey102]

What websites allow NSFW using SDXL

[@frazy4487]

Just how many times does this guy say "Open source" in his videos 😂

[@LouisGedo]

👋

[@RockyBergen]

Midjourney is WAAY overrated for my interests, at least.

[@MilesBellas]

Midjourney's celebrity likenesses are deliberately inaccurate?

[@MilesBellas]

"Midjourney founder David Holz tells Forbes Mostaque gave a “very small” financial donation but that is all. " Emad help finance Midjourney's inception.

[@keisaboru1155]

WHAT ! TF !

[@jackfrosterton4135]

"Ensemble" is how you spell ensemble.

[@gerkim62]

First comment?

[@Anti-HyperLink]

I'm so sick of github. I don't understand how to get shit off of it unless it's regular files.

[@Jumpyfoot]

It's always a chore to get files off of GitHub. I just have to dork around aimlessly and hope to get to the right files.

[@AutumnRed]

I'm finding it worse than the user made models, no Lora training allowed, much heavier on the GPU - running out of memory pretty quick - and worse results, not sure what people are finding so good on this SDXL update tbh

[@mirek190]

LOL only you

[@munchkintaker4263]

What a BS. there are many LORAs already. VRAM consumption isn't that bad, especially on comfyui. And the composition of the results are generally better than even the best 1.5 models.

[@ilyass-alami]

It is slow to generate images

[@FlamespeedyAMV]

1.0 kinda sucks compared to 0.9

[@ThymeHere]

1 day 1kliks 206k subs

[@michirex8766]

Kkkkkkkkkkkk

[@realthing2158]

All images of females on Clipdrop come out with overly masculine features for me. Seems skewed.

[@scjjml]

Hate clipdrop, the resolution is terrible.

[@williammixson2541]

"Own machine" 7900XTX. 24GB VRAM, useless junk in Windows or WSL2.

[@anewman1976]

1.0 is rubbish 0.9 was better I noticed that the resolution is a lot worse and detail even before you download HD is rubbish.

[@diadetediotedio6918]

They made the model more easily trainable at the cost of some quality, so for finetunnings the general quality would be better. Also, it is not worse than 0.9 in a broader sense, it can generate text fairly well for example.

[@reflectivish385]

This is the worst combo ever. AN ADVANCED AI IMAGE GENERATOR UNCENSORED??

[@IPutFishInAWashingMachine]

👋

[@thegreatdelusion]

It's not uncensored. There's no NSFW. And celebs have been filtered out.

[@jopansmark]

I generated Xi Jinping with SDXL and it's literally Xi Jinping. So, it generates celebs easily