New Model! SDXL Turbo - 1 Step Real Time Stable Diffusion in ComfyUI

Stability has released a research model that is a real time model, that can deliver images in only 1 step! This is not the same as a latent consistency model (LCM) as it is using Adversarial Diffusion Distillation (ADD). In this comfy tutorial, I will show you how to use the custom sampler to implement a more focused method rather than using the normal ksampler (which still works), but this is much more true to the ideal workflow for this model.
#comfy #stablediffusion #aiart #sdxlturbo
You can get the SDXL Turbo model here: huggingface.co/stabilityai/sd...
Interested in the finished graph and in supporting the channel as a sponsor? I will post this workflow (along with all of the previous graphs) over in the community area of KZbin.
/ @sedetweiler

[@havemoney]

Turn on an audio book, let the images from the book flow seamlessly from one to the other.

[@havemoney]

audio >> text > pic

[@sedetweiler]

Totally!

[@ArielTavori]

A Pinokio script would make this seamless, especially with a small, fast llm, fine tuned to summarize the most recent paragraph into a visual description...

[@ArielTavori]

Or alternatively text>audio/image...

[@sirtimatbob]

Also adding in background noise, such as a busy metro area if that's where the people are talking, etc.@@ArielTavori

[@UltraStyle-AI]

I think this is one of the pieces to making this more accessible and useful, along with using less VRAM resources. Seems like we're headed in the right direction. Thanks for showing us!

[@sedetweiler]

You are quite welcome. Cheers!

[@ahtoshkaa]

I've just finished testing LCM and here is a new model that takes it even further. Woowee!

[@sedetweiler]

Yup, huge steps forward every day!

[@elysilk4538]

Thanks to SDXL Turbo, I produce in 9 seconds what used to take me twenty minutes! I now generate 24 images of bioluminescent organisms, as an example, in almost the (slow) blink of an eye. Thanks to you and others for highlighting the new SDXL!

[@sedetweiler]

Rock on!

[@petec737]

No you aren't. The quality difference between the two is incomparable.

[@daniwanicki]

Amazing, it's like painting with words. Thanks for showing how to set it up!

[@sedetweiler]

Glad it was helpful!

[@superlucky4499]

Thank you for your tutorials. I am learning so much. Please keep doing what you are doing!

[@sedetweiler]

Glad you like them!

[@RaphaelAlba]

This is insane! So great that the technology is moving so fast!

[@sedetweiler]

It is a run-away train some days!

[@jaredbeiswenger3766]

That's very cool. Appreciate this little insight into something I never would have found on my own.

[@sedetweiler]

You bet!

[@thomasgoodwin2648]

That is like 'CRAZY' with a capitol W! ❤

[@sedetweiler]

Indeed!

[@Q8CRAZY8Q]

Very nice to practice prompts efficiently

[@sedetweiler]

I agree! I think that is one of the more interesting things to come out of this.

[@ImAlecPonce]

thanks!!!! this is going to be perfect to actually see which prompts actually DO something!!!

[@sedetweiler]

I think we will all learn a lot from it!

[@Renzsu]

Fun times! Kids are going to love this.

[@sedetweiler]

I think my wife might enjoy this as well! Cheers!

[@edsonjr-dev]

Scott Detweiler Best teacher in the world ever, meu like e watchtime 100% 💜💜💜

[@sedetweiler]

Aww! Thank you!

[@Firespark81]

Thats really neat! Thanks for the info!

[@sedetweiler]

Sure! Thanks for leaving the comment as well! Cheers!

[@murphylanga]

Thank you very much, I always discover great new things here🤩

[@sedetweiler]

Happy to hear that!

[@blender_wiki]

For who is wondering, on our early tests look like is working also with the animediff workflow withsome twiks, speed gain are not really clear yet we need to do some more controlled comparisons.

[@MannyGonzalez]

Damn, this is super cool and fun to just type. Good to see what the model has hidden inside. LoRAs work dynamically too and you can just add them via pop-up prompting as you type. Thank you!

[@zoemorn]

Helpful video but for newbies: to perform the copy of the nodes and links together, hold CTRL while you highlight the nodes, use CTRL+C to copy, use SHIFT+CTRL+V to paste which pastes the nodes and connected links.

[@yourashero]

Thanks for the tip. Do you know how to update the comfyui so the correct sigma shows up? I just installed the latest release and the fp16turbosdxl but I don't have the turboxl sigma

[@zoemorn]

@@yourashero hey sorry i'm still quite the newbie, but ran into the tip i posted just due to my trying to follow the video. i'm not sure what you mean by 'sigma' and googling isnt help me understand the question. i do see the sdxl-turbo in the Load Checkpoints node of the workflow.

[@Weroleytor]

Thank you :)

[@zakblue]

updating images on the fly like that is a game changer! wow

[@sedetweiler]

I agree!

[@magenta6]

Just BRILLIANT!!!

[@sedetweiler]

Thank ya!

[@SaadKhanAhmed]

Your videos are amazing!!!

[@sedetweiler]

Glad you like them!

[@Champignon1000]

this is super useful for prompt exploring. i guess you can always use a better slower version after.

[@sedetweiler]

Exactly!

[@pn4960]

Super impressive !

[@sedetweiler]

Glad you think so!

[@JimPenceArt]

Awesome! It even works on my slow system. Average about 7-15 sec per image.

[@sedetweiler]

Great to hear!

[@johnmcaleer6917]

Amazing...tonights fun all sorted then...😄

[@sedetweiler]

Hope so!

[@agnesslovehealz]

Super great tutor

[@fredlef100]

Great video - my wife showed me a video of someone demonstrating this a week or so ago and here I'm showing her - sort of - my M1 Mac isn't nearly as fast as yours - can't wait to try it on my soon to arrive M3 Pro Max. I'm really new to ComfyUI and Stable Diffusion in general so your videos are all very helpful. One thing on this workflow is that I don't get images in the sampler like you do. Not sure what I'm missing to cause that.

[@sedetweiler]

In the manager, set the preview to slow taesd. That will give you previews. Cheers!

[@fredlef100]

Great thanks!@@sedetweiler

[@yashaven]

This works great on my 3gbvram gtx1050 , thankyou

[@sedetweiler]

Good to hear

[@aimademerich]

Phenomenal

[@sedetweiler]

I love it!

[@snatvb]

wow, amazing

[@sedetweiler]

Thank you! Cheers!

[@-invisio]

Very nice workflow, thanks. BTW, you should tell us what hardware you use to have such fast image generations: I guess an RTX 3090 or 4090?

[@sedetweiler]

I am using a 3090, but 1 step will be super fast on pretty much anything that can run stable diffusion.

[@ThePromptWizard2023]

Holy c*** that is amazing, I cannot wait to test various prompts. What a shame this is not possible with all models.

[@sedetweiler]

You can use that LoRA and it should work on most models. :-) Cheers!

[@obscuremusictabs5927]

You should do a video on what to do when comfyui nodes fail to load and how to deal with node conflicts. Thanks for your work.

[@AkoZoom]

hey cool ! .. under the wood i've gone a part on Fooocus which integrates comfy engine .. then the double UI (to make nodes and config all the "tripal things" then to use an easy interface would be a cool thing. Even comfy makes it comfortable but a bit tricky. * About the SDTurbo : is that it is only a safetensors model that could be used as it without any other trick ? i tested .. some images but not fast - the turboscheduler is not yet implemented - is this "just" adding this "sigmas" thing to make it ok ? -- oops tested in sdXl and 1024²

[@Pauluz_The_Web_Gnome]

Hi Scott, Thanks for another great video! Love it...Question: How do you get the colored nodes by default? Mine are all gray... Thanks!

[@PatrickIsbendjian]

It's part of the ComfyUI_tinyterraNodes Just install the custom nodes and you'll find the option as "Default BG Color (ttN) when you right click on a node

[@Pauluz_The_Web_Gnome]

Thanks man! Gold!

[@lzrdfrt]

Your insights to the settings are incredible and have helped me greatly! Question i haven't found elsewhere. How important are the OS NVidia driver versions for rendering? If at all?

[@sedetweiler]

I would always stay current. I don't know if there is an advantage in game or studio here, but I tend to go with studio.

[@ReddXYZ]

Excited to see what my thumbnails will look like with this new tool!

[@sedetweiler]

I have been having a ton of fun with this. Way more than I initially thought I would.

[@vincentvoillot6365]

I'm still not used to LCM XL in 4 pass, still recovering from stable video and now they roll-out a turbo one pass version . oO.. the progress in 2 years is incredible.

[@Nasser-bp6qf]

more like 1 year

[@angryox3102]

I could see this being useful for trying to hone in a prompt, then using that same prompt for high res generations.

[@sedetweiler]

Yup! I think that is probably the golden nugget here no one expected.

[@bwheldale]

Absolutely, I just played around with word weights (ctrl+up) discovered in realtime the effects and which words do nothing

[@alpaykasal2902]

I can't wait to try! Do Loras work? Thanks!

[@sedetweiler]

Not sure. I have yet to try them.

[@nic-ori]

Thanks.

[@sedetweiler]

Sure!

[@bwheldale]

On increment instead of fixed was spectacular until I realised 'save images' was applied then spent the next half an hour sifting through so many crazy generations!

[@sedetweiler]

Oh wow! That had to be a chore. :-)

[@redregar2522]

I think the speed up is impressive. For me the question is can I use it also for Higher Quality and image sizes without the need of masses of gb ram

[@soulxslayerchan]

Damn, that's fast!

[@sedetweiler]

indeed!

[@TPCDAZ]

It's 1 step of course it's fast. any model is fast with 1 step

[@soulxslayerchan]

@@TPCDAZ "any model is fast with 1 step" but it can't produce a good outcome. new to SD? ahahaha. the point was, that it produces good results fast, which other (except LCM) models don't.

[@TPCDAZ]

@@soulxslayerchan You never said quality. You said fast. New to reading? Learn to read before assuming people want your opinion.

[@soulxslayerchan]

​@@TPCDAZ So I assume by your logic, one can make a 128x128 picture with 1 step and SD1.5fp16 model and be happy with whatever a mess you get. Because it's fast, and acceptability/quality of generation is not logical. You, sir, are a true genius. 👏

[@SylvainSangla]

Very interesting, thanks ! What is the difference between what you did (copying the nodes 4 times) and setting 4 to the "batch_size" on the Empty Latent Image node ?

[@sedetweiler]

Layout, but I can also change other settings for each version. So I would prefer this method.

[@zackio2683]

Thank you for this awesome workflow! One question, does this workflow work with lora? I tried to add a "load lora" node and connect it after the "load checkpoint" but the output image was messed.

[@sedetweiler]

All of these should work with a LoRA.

[@EuropesHeart]

What can I do if the TurboScheduler does not appear in the Sigma list? The model is loaded and ComfiUI has also been completely restarted.

[@drformalyst5222]

Tried it, works amazing - thanks for sharing! I am wondering, would it be realistic to try to use this model with AnimateDiff or at least img2img - I tried but obviously results are not quite good given that we can't control the denoise strenght (or start steps) given we only got one step :) Any ideas of a workaround?

[@sedetweiler]

I don't see why not!

[@drformalyst5222]

@@sedetweiler I'd really LOVE to see that workflow then :).. My assumption that it's not possible is that the sampler only does on step, so it's not possible to fine-tune how much of the original image (say in img2img workflow) we want to maintain as we do by adjusting start step in Advanced Sampler (or denoise weight, which I assume does the same as start step). I tried img2img, controlnet and AnimateDiff workflows, and I can confirm that it does not work. CNet (without img2img input) flows do result in a visual that sort-of follows the original one, but heavily smudged (almost like a normal SD doing one sample, more/less). Anyway, thanks so much for your channel!

[@wisdombox4]

Hi Scott Can do a video about how to add prompt styler any workflow, I look on KZbin no one made a proper tutorial, Thank you

That was very exciting again! Another push to have a reason to leave AUTOMATIC. I just still really need the img2img/ Controlnet/Tile/Ultimate SD Upscale workflow to be able to take images of large size and resolution. Can I hope that there will be such a solution?

[@AlecuGrigore]

My friend, all those are in comfy for some while now…

[@sedetweiler]

Yea, those are in the community graph area, and as a member you can go download all of those. Also, check the live stream replays, as I do that quite often as well.

@@AlecuGrigore I have less practice in Comfy. Can you provide a link to a workflow?

[@sedetweiler]

There are a ton in the member area here on the channel (hint hint) 😉

[@1111Scorpitarius]

Something's awry with the Queue Prompt. If I set it to Auto Queue it just runs on a constant loop, not the best for my slow- assed PC. If I DON'T set it to auto, however, it cues 2 at a time!!!! Thoughts?

[@babbagebrassworks4278]

Running SD XL on a Raspberry Pi5 with the Onnxstream mod. Does landscape images better than people, but only 512x512.

[@tomschuelke7955]

Well i updated comfy ui, but i still cant get at sigmas a sd turbo scheduler.. i have reroute, basic, karrras, exponential and polyexponential.. in the manager i clicked on update, and it told me it was up to date... any idea?

[@DJBFilmz]

Me too. Any word on this?

[@patagonia4kvideodrone91]

Something updates and disappeared the scheduler, but with the Basic Scheduler it works.

[@RamonGuthrie]

Is there a ComfyUI node that allows me to render an Output video of the Denoising stage, you know the Preview you see on the KSamplers at the Denoising stage if you have Preview method set to Auto or TAESD is there a way to output that as a video?

[@jibcot8541]

It's pretty cool, faster and more flexible with Loras than Nvida TensorRT. It a shame there are limitations on img2img upscaling (denoising strength x steps must be 1 or greater).

[@sedetweiler]

Give it time. It has been less than 24 hours. 😂

[@dhrumil5977]

Yooo that's coooll but is there any notebook for this? Now I want to start learning comfy ui

[@sedetweiler]

You can run comfy in a notebook, but I do not have experience using it in that manner. Check the comfy git, as there is a notebook ready to go on there.

[@dhrumil5977]

@@sedetweiler thanks!!

[@AM-yk5yd]

Their Huggnigface provide the code on how to use it.

[@jonmichaelgalindo]

Wow. Yeah, it really can't get forms right, but the speed is amazing.

[@sedetweiler]

It is just another step forward.

[@jonmichaelgalindo]

@@sedetweiler Actually, at 4 steps cfg=1.2 it can do hands / eyes etc. almost as good as SD 1.5. Why did you say 2 steps max in the video? Even reading SAI's paper they specifically designed it to work best at 4 steps.

[@Voidermusic]

It doesn't generate anything in real time for me, did all steps exactly like in the video. Is it because I am on a GTX 1070?

[@Holubiik]

I did everything but my Image preview stayed empty, nothing was generated idk where i went wrong

[@wrillywonka1320]

Can turbo create multiple angles of one object?

[@keepitshort4208]

My python crashed while using comfy ui during prompting, what can be the initial cause of it to happen ?

[@patriotfett]

is there a way to create multiple images at one time using comfyui

[@tezromania]

will it let you create a loop in the graph? could you feed the result image back to the beginning as img-img input?

[@sedetweiler]

Naa, just add to the end. You want to avoid a loop.

[@tezromania]

@@sedetweiler well, I was sort of hoping for loop behavior though where it would run continuously...

[@ronin_1375]

what is the difference between 16fp and the other one that is 13gb is that fp means fps frame per seconds or something else

[@edsonjr-dev]

Hello Scott, what is your computer configuration please? I want to know which computer to buy, thank you.

[@sedetweiler]

I actually have a hardware review coming up soon!

[@edsonjr-dev]

​ @sedetweiler Thanks for the answer, I'll wait anxiously.

[@erics7004]

This is amazing! I'm making images in seconds with only 4gb of vram. It's insane! I can't use LCM because my GPU only accepts full precision. I can't use fp16. (1600 nvidia series doesn't allow that).

[@sedetweiler]

Great to hear!

[@sokolum]

What is next recommendation

[@littleowl2912]

Any difference between sd_xl_turbo_1.0.safetensors and sd_xl_turbo_1.0_fp16.safetensors ?

[@McSozialphobie]

I have built it exactly the same, but it takes about 15 seconds to create a martian landscape for me. (checkpoint etc already loaded) Im running it on a 2070rtx 8gb. Is my card too slow or what could be the issue?

[@sedetweiler]

hmmm, that seems slow. Perhaps create a new install in another directory and see if you have something hanging around that might be causing the issue.

[@McSozialphobie]

thank you for the quick answer :D@@sedetweiler

[@webdvex]

What pc setup do you need for this speed

[@leaphardotnet]

amazing but i have a question : each time i change my prompt it have to reload the model so i don't get the instant result like you, what do i do wrong ? got prompt Requested to load SDXLClipModel Loading 1 new model Requested to load SDXL Loading 1 new model then it render the steps in like a franction of a second

[@sedetweiler]

You might ask that on the developers git. That isn't the typical way this should work. You might actually try reinstalling, since it's pretty painless if you have the models locally

[@leaphardotnet]

i reinstalled a clean comfyui and actually i think it's just normal that the model is reloaded after each prompt change, so on a rtx 4060 it take 0.3 sec if i don't change the prompt and about 3to5 sec if i change it, not bad@@sedetweiler

[@TheMahdiayman]

it created a folder after cloning it ((generative-models)) but cant find any turbo related node or model in the comfyui

[@sedetweiler]

Double click and search for it if it isn't in your menu.

[@JavierGarcia-td8ut]

3:09 I updated ComfyUI but I can't see SD Turbo Scheduler

[@krisDag17]

I had the same issue, but if you search for it, it should come up. There is update.bat file you can try to run as well.

[@user-pc7ef5sb6x]

My images are blurry. cant figure out what I'm doing wrong

[@zengrath]

uh, i did a fresh install of comfy ui this weekend, all up to date, and this turbo sampler isn't there. not in the manager too. I searched all the available sampler plugins i can't find it. if i just downloaded and installed comfy ui using git clone of the repo and all how is my comfyui out of date? how exactly am I behind the times? What am i missing?

[@zengrath]

ok in video shows you dragging from sigma and it's there. for me i have to search for it, but I found it.

[@uncertainultradian]

I think it's worth noting that the huggingface page shows a preference for 4 steps rather than 1, for best performance and quality; and "preferably, the model generates images of size 512x512 but higher image sizes work as well." So it's not 512*512-only, they just don't really specify further as far as I can see. ...And there are two custom checkpoints based on SDXL Turbo on CivitAI already. XD

[@sedetweiler]

I have found that 4 is a bit much. Give it a try and tell me what you think.

[@erics7004]

I've tried with 2 steps and it worked well. 9 seconds with 4gb vram! Crazy!

[@dafimperator]

@@erics7004 What GPU did you used?

[@zintoki8211]

Yeah 4 looks like sharpness filter. Best results are 1-3. Also images with higher resolution than 512 are tiled so no it's not made for higher resolutions.@@sedetweiler

[@TheBogstaverne]

How'd you get auto queue to only go when you changed the prompt?

[@sedetweiler]

Once you have auto-queue clicked and start the job, it will run whenever something changes, like the prompt, seed, etc.

[@zintoki8211]

Is it possible to use it as base for SDXL refiner?

[@sedetweiler]

Sure! Not sure it will be great, but you can try this with any SDXL mechanism.

[@WhiskerClub]

Is there a way to create 16:9 images? Or would i need to generate 512x512 and then convert them to 16:9 or something?

[@sedetweiler]

You can try, but the model will work best with the 512 square.

[@tom.shanghai]

is img2img possible with this? there is no denoising strength option in the custom smapler..

[@sedetweiler]

I would probably use this as the first pass, and then encore the result for img2img with another ksampler

[@tom.shanghai]

@@sedetweiler I've tough about that, but.. then the speed advantage is gone :l

[@kkryptokayden4653]

Have you tried this with animations yet? Cant imagine the speeds.

[@sedetweiler]

I have not, but I bet you are right!

[@aminh101]

I went through the comfyui manager to update all but I still don't see the SDTurboScheduler :(

[@sedetweiler]

Search for "turbo" your list might be ordered differently.

[@tofu1687]

@@sedetweilerthx: searching for turbo finds the node - it is just not in the default list of samples...

[@o1ecypher]

i have OCD can we get the windows to snap in place with a minimal space in between

[@sedetweiler]

I have OCD as well, but this was just a quick demo. I often leave the spaces so people can clearly see the connections.

[@MartinDenStore]

You can snap the nodes to the grid in your own workflow when using Comfy :)

[@SSPARROWW]

I got this working in A1111 and after watching this, I've been on a binge trying to find something that works like in the video. Where typing a new word in the prompt auto generates another image. Anyone know anything about this possibility?

[@Elwaves2925]

I don't and haven't looked yet but I am curious if it's possible. My brain is telling me it's not possible because of how A1111 is designed but what does my brain know. 🙂

[@SSPARROWW]

@Elwaves2925 so far the only thing I've found is a bandaide type fix. Right clicking generate you can make it infinitely generate, then typing in the prompt will update the next generation. Leaving the seed the same will keep the image pretty stable and leaving it as a random seed kinda works just a little chaos. But it doesn't stop generating images is the only downside and when it's constantly generating its easy to "lose" the real good ones that fly by lol

[@Elwaves2925]

@@SSPARROWW Thanks but I think I'll pass on that way. Infinite generations until we stop it is not what I'm after. When I get some spare time I'll do a search myself but I suspect if it was possible, it would have been mentioned in a video.

[@tomschuelke7955]

Now... i would like to see this in invoke ai, in conjuction with the unified canvas, where i would like to paint in and on the other side this will get interpreted in realtime..

[@sedetweiler]

That would indeed be amazing.

[@Aristocle]

can I use ip-adapter with this model?

[@sedetweiler]

Yup!

[@leonardhinkelmann5629]

Does someone know if this model is/can be compatible with controlnet?

[@sedetweiler]

Try it. It's brand new, so someone will need to test things on it. With only 1 step, it might be hard to control.

[@Steve.Jobless]

hey scott, can you make workflows for img2img with sdxl turbo, is it possible?

[@sedetweiler]

Sure! We will probably do that on the live stream tomorrow.

[@Steve.Jobless]

@@sedetweiler nicee i can't wait

[@matiosjed]

I updated Comfyui by using a manager but I still have only 2 sigmas avaiable : flipsigmas, splitsigmas. Do you know why?

[@sedetweiler]

Did you restart afterwards?

[@seanc7676]

Same with me , I've updated and restarted and It's still not showing up@@sedetweiler

[@sedetweiler]

Search for "turbo" and see if it shows up.

[@aminh101]

@@sedetweiler I faced same issue but searching for it showed up!

[@Toskyval]

How much VRAM minimum? Would a RTX 3060 12GB be enough?

[@sedetweiler]

Oh heck yes. Plenty.

[@the47er]

when i write a prompt nothing happens, it does not generate an image. What i am doing wrong? i have a geforce rtx 4080

[@sedetweiler]

Not sure.

[@bogdahn689]

Hello, I hope you're doing well

[@LouisGedo]

👋

[@sedetweiler]

🍺

[@ProzacgodAI]

Can we do img2img yet?

[@sedetweiler]

Sure! You have been able to do that since day one, and it didn't change with this model.

[@MasoodOfficial]

What is your pc spec?

[@sedetweiler]

3090 nvidia and 64gb of system ram

[@CarbonBeast]

how to use this with automatic1111

[@sedetweiler]

Give it a few weeks and I am sure someone will come up something.

[@patagonia4kvideodrone91]

I was doing videos in a very short time yesterday with other methods, and well when this will see if it works in my workflow with this, but if they took 1min with this it will be instantaneous? XD is already at any time we put a chip in our heads with Elon Musk and send us to dream! XD was looking and I think I don't have the Sigma, I didn't understand if you have to update Comfyui for it? I do not see that I show them to me in the manager, but I have a stack of how 1200 images in production, so when they finish a vow if they update, if you get to walk and only if you go, you may have videos in real time also xD?

[@BG4-R]

动画也应该可以完全自己控制了。。。哪怕60帧的。。。

[@sedetweiler]

It's not animation here, but I bet we could use it for that.

[@pedroserapio8075]

wow, got 0.32 seconds on my RTX3060 12GB

[@sedetweiler]

Rock on!

[@FusionDeveloper]

⚠ **BLACK IMAGES** unless you change this:⚠ I have to add "LOAD VAE" and choose the SDXL VAE safetensors file or else i just get a solid black output image. Connect Load VAE to VAE Decode (instead of from the Load Checkpoint) and be sure it's an SDXL vae.

[@sedetweiler]

Hmm, that is odd. I tried with both and I have the same images from either one.

[@FusionDeveloper]

@@sedetweiler yeah, that's why it's helpful to have comments like mine and yours, because things don't work the same for everyone. Maybe one of us overlooked a detail or have a conflicting file or different hardware or drivers, it could be anything. People can try what I said if they get the issue I had. Thanks for the video.