SDXL 1.0 IS HERE! Were to get it, how to use it, and what to expect.

The latest model from Stability.ai is out, SDXL 1.0! This is a huge upgrade to models of the past and has a lot of amazing features. In this video we cover where you can download it, what tips and tricks we use internally when testing it, and some other tips that will help you get the most from this amazing AI art tool. #stablediffusion #sdxl #comfyUI
Grab the SDXL model from here (OFFICIAL): (bonus LoRA also here)
huggingface.co/stabilityai/st...
The refiner is also available here (OFFICIAL):
huggingface.co/stabilityai/st...
Additional VAE (only needed if you plan to not use the built-in version)
huggingface.co/stabilityai/sd...

[@johnmcaleer6917]

Having a play on ComfyUI and it's very impressive, I generally do 'photo realistic' images and have been very happy with the community created models using SD1.5, very much looking forward to what comes from this model..

[@sedetweiler]

I think you will love what this model can do!

[@LeonvanBokhorst]

Great! Thanks 🙏 you're the reason I got into Comfy :)

[@tomschuelke7955]

i also would suggest to try out invokeAi 3.0 which just came out.. it combines a far easier interface like A1111 with a nodeworkflow that now still is beta, but sure with 3.1 will get much more solid.. and also allready was able to use SDXL 0.9 as i heared. and.. yes the interface is realy a charme.

[@sedetweiler]

I am excited for the nodes in Invoke as well! However, it needs a lot of work yet to be useful. I still think Invoke is a beautiful piece of software and probably the best for inpainting over the others.

[@tomschuelke7955]

@@sedetweileri will learn all of them..

[@jaredbeiswenger3766]

Lots of great tips. Appreciate it. Looking forward to trying it out.

[@sedetweiler]

Glad it was helpful!

[@MonzonMedia]

Saw you on the call Scott! Then quickly recorded a video! hahaha! Runs so well on ComfyUi even on my lowly 3060Ti. Tried it on Invoke AI but it looks like I need to optimize things in the backend, was running super slow...couldn't wait. I'd love to see different SDXL set ups on ComfyUi.

[@sedetweiler]

I plan to start giving those away soon! Glad you caught the call! Cheers!

[@MonzonMedia]

@@sedetweiler looking forward to it! 👍🏼

[@vilainm99]

Love ComfyUI & Stable Diffusion!!. Got the SDXL model. If I load this, the VAE output still uses the 'old' VAE? Do I need to download the SDXL VAE for best results?

[@sedetweiler]

You can use either.

[@marjolein_pas]

Thanks for the update. And yes, I've noticed that negative prompts are mainly BS

[@marjolein_pas]

@Don_Coyote thanks for the tip

[@bAEtman.]

@@marjolein_pas can you please help me regarding how to write prompts.. The AI does no seem to generate images according to my words, a lot of things are always missing, the faces and eyes are uncanny, plz help 🥲

[@VioFax]

Do you think 1.5 will still be useful after XL is more optimized and widely used? what about custom models? or LORA?

[@sedetweiler]

It might be, but really it is going to start to look pretty dated soon. Custom models will come quickly, as I said this will be easy to train.

[@VioFax]

@@sedetweiler My mind automatically goes directly to "make them all work together somehow"

[@nullKass]

ComfyUI has become my favorite already, the only feature I'm still looking for is batch solution for ControlNet, that's why I still have to use A1111 at this moment.

[@sedetweiler]

Yup, but it is coming. It's just all moving so fast.

[@nic-ori]

Thanks.

[@sedetweiler]

You're welcome

[@ChrisAllenMusic]

Congrats 🛳

[@sedetweiler]

Thank you my friend! Hope you are doing well!

[@appolonius4108]

nice video :D

[@sedetweiler]

Thanks! 😁

[@Arlechinu]

Any chance of an ‘official’ workflow for ComfyUI for using SDXL? Cheers! Excellent videos!

[@sedetweiler]

Working on it!

[@Arlechinu]

@@sedetweiler that is fantastic news, thanks! Looking forward to the release 👌

[@sedetweiler]

Just posted it! :-)

[@Arlechinu]

@@sedetweiler already watching it, and liked it 👌 thanks again, really enjoy your videos, keep it up!

[@tomschuelke7955]

Great. now for ome reason i wasnt able to create the best results with 0.9. but i am courious how it works out now. Furthermore i just today got some increadible greate architectural images from a colleque from midjurney, and just did throw the prompt to SDXL 0.9.. first time trying out SDXL 0.9 a bit more.. moods and architecural quality was nearly as good as midjurney allready with the beta.. only two issues.. straight lines where often a bitt wiggling and not straigt. And it was nt able to creat a Square structured patterns in the ceiling that where stunning great with MJ.. never the les also with the beta a big step.. and now lets see what the final XL will be like..

[@sedetweiler]

Let me know! I think it works pretty well for architecture.

[@tomschuelke7955]

@@sedetweiler well i just tested the 1.0.. but it stays this way... overal a great improvement in overal light and compositing., but just tried to creat a skyscraper and all lines are totaly ... unstraight... searching for words.. maybe my jason file for comfi ui is wrong? some values wrong? is there somewhere a great start jason available?

[@umarudoma1811]

so you mean i just download the SDXL model and paste it on to my previous SD1.5 folder and just select it as checkpoint to use the SDXL ?? no need to reinstall anything from square one?

[@sedetweiler]

Correct.

[@REDINKmysteries]

Does Comfy have img2img,outpainting and drawing features?

[@tsutsen1412]

yes it does

[@sedetweiler]

Yes, it does all of the standard operations.

[@beatemero6718]

Can you link the specific vae? I dont understand the name you are saying of the guy who made it. Sounds like Calyuga, or something. Would be nice to explain who that is, or show it in screen.

[@sedetweiler]

I will get those links posted today.

[@beatemero6718]

@@sedetweiler Thanks!

[@sadelcri]

I have a few questions: does it generate 1024x1024 images as fast as with sd 1.5?? can it generate images smaller than 1024x1024? and bigger images?

[@IlRincreTeam]

It's supposed to generate 768*768 and above, smaller ones ain't good with this model

[@sedetweiler]

I would imagine they are a bit slower because the models are much larger (thus the XL in the name), but native 1024 images is the benefit. Making smaller images might yield poor results.

[@lakislambrianides7619]

Can I use the same workflow of SDXL 0.9 with the new model?

[@sedetweiler]

Yes.

[@professorpenne9962]

I have a lot of fun generating war machines and mecha in SD will this model do good for that application?

[@sedetweiler]

I sure think so!

[@UcheOgbiti]

Please I'm confused. Can anyone explain the difference between SDXL 1.0 and Stable diffusion 2.1?

[@sedetweiler]

New model with a LOT more parameters (thus the XL part of the name). Much easier to train and a lot more flexible.

[@ZenBenzineX]

If anyone know how to get a perfect tileable pattern, please, share me the trick... I tried but it seems SDXL doesnt created it.

[@tomschuelke7955]

i struggeled with something like that on a ceilig.. a square pattern ... MJ exeled here.. SDXL well that was a faile, while everything else of the beta was a huge improvement ..

[@ZenBenzineX]

@@tomschuelke7955 I´ve been trying and they simply dont match when putting the tile next to the other. Is really not easy. So if anyone can find the trick... PLEASE LET US KNOW!!!

[@Definesleepalt]

do we still train it using dreambooth ?

[@sedetweiler]

Yup! That will work.

[@hobologna]

Automatic1111 won't load the 1.0 model or the refiner for me.. any ideas?

[@sedetweiler]

No idea. Like I said in the video, give it a day.

[@Cutieplus]

I have no problem using the latest A1111 v1.5.0. It clearly state that it upport SDXL.

[@None_ya_B]

Disable extensions and restart.

[@hobologna]

@@None_ya_B what worked for me was using --medvram. I still get some weird OOMs at time, or tensor errors. I am cloning Comfy again and using it instead for now, and keeping 1111 for 1.5

[@Elwaves2925]

I couldn't get the model to load on A111, it would process for a few seconds then revert to the previous model. This is with an RTX 3060 12Gb and 32Gb RAM and I believe A111 is on version 1.5.1 (not sure how to tell exactly). If anyone has a solution, I'd appreciate a heads up to it. Tried it on ComfyUI and it worked straight away. Huzzah! Not sure about the quality of people yet but other generations seem to work great.

[@rndll.]

It doesn't work with Hires fix enabled.

[@Elwaves2925]

@@rndll. Thanks for the suggestion but I never use that, so it isn't enabled.

[@sedetweiler]

you might try putting --no-half-vae in the COMMAND LINE args in the bat file.

[@Elwaves2925]

@@sedetweiler Thanks but it turned out my A1111 install wasn't updating, so I did a fresh install and now it works. Cheers.

[@Moh_ha]

Can use it for commercial purposes?

[@sedetweiler]

Yup!

[@monbritt]

hi, where i can download this workflow?

[@sedetweiler]

I will post it soon for everyone

[@monbritt]

@@sedetweiler thanks, the first is great to, thanks. Hugs from brazil!

[@sedetweiler]

Thank you!

[@techniquebeats_dennis]

how about controlNet ?

[@stefanb7015]

Stability is working on it.

[@sedetweiler]

Those are coming soon. They need to be trained and we might have some nifty ideas of our own to toss into the mix.

[@BikingWIthPanda]

Stability promised controlnet would be ready at launch a month ago

[@LouisGedo]

👋

[@sedetweiler]

🍺

[@case0h]

workflow?

[@sedetweiler]

Coming soon!

[@BikingWIthPanda]

a finetuned sd 2.0-v or 2.1-base have far more detail than SDXL 0.9 or 1.0.

[@BikingWIthPanda]

in fact the v-prediction of 2.x allowed us to fine tune it on zero-terminal SNR and there was no need for offset noise. when you say negative prompts are a thing of the past, you didnt mention that they no longer work? they do not work properly anymore. 'watermark' doesnt give you watermarks in positive prompt, so, negative prompt cannot remove them. sdxl was trained on so much watermarked data. you guys ought to be ashamed.

[@alejandrobarbas6396]

AMD compatible???

[@sedetweiler]

I believe there are options for that in comfy, but I don't know directly.

[@andrer4508]

why do i get missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'} when trying to use SDXL 1.0 it takes super long time when i Queue. then it says, DDIM sampler, 0-100% downloadin, then it restarts again with Setting up MemoryEfficientCrossAttention. and does the missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'} again. dont know what s wrong..

[@sedetweiler]

I would avoid that sampler.

[@frankschannel2642]

Eyes and hands are still poor. Body proportions appear elongated. So far, I'm finding that 0.9 gives better results.

[@DomagojStivic]

jup strange tho idk

[@libbychang413]

6:55 that even sounds lame...