SDXL Lightning is SUPER FAST! Only 2 Steps to Render!

SDXL Lightning is extremely fast and works with any SDXL model. You can render in only 1, 2, 4 or 8 steps. The amazing thing is: You can also render that fast for upscaling.
#### Links from the Video ####
More AI News on Twitter: / oliviosarikas
SDXL Lightning: huggingface.co/ByteDance/SDXL...
My Workflow on Patreon: / 98927565
#### Join and Support me ####
Buy me a Coffee: www.buymeacoffee.com/oliviotu...
Join my Facebook Group: / theairevolution
Joint my Discord Group: / discord
AI Newsletter: oliviotutorials.podia.com/new...
Support me on Patreon: / sarikas

[@OlivioSarikas]

#### Links from the Video #### More AI News on Twitter: twitter.com/OlivioSarikas SDXL Lightning: huggingface.co/ByteDance/SDXL-Lightning My Workflow on Patreon: www.patreon.com/posts/98927565

[@TR-707]

olivio, thank you for your tutorials. You are really great at them. Thank for you concisely and eloquently going from 0 each time there is a new workflow or model and not assuming that everyone already knows it because they watched a tutorial 5 years ago. Its quick for veterans yet detailed enough for newbies

[@WD_MyImage]

Thanks, Olivio. I'm going to try this out soon.

[@pauloak121]

Just want to say I love your page, your videos and the detail and explanation you put into them. Please keep up the great work!

[@SpiralingOut23]

the fast upscaling is impressive

[@TheMartiansLoveKuato]

So handy for upscaling

[@sb6934]

Thanks!

[@jbrocktheworld]

I've try and love it, No more LCM turbo for me now

[@frankhovis]

The images you have of the bald bloke with a beard in front of the microphone are utterly amazing.

[@ArmanVesona]

Wow you so fast

[@povang]

Whats the difference between lightning lora vs LCM lora?. Im using LCM lora 2 cfg and 8 steps.

[@OlivioSarikas]

i feel like the quality is a lot better with Lightning

[@SionMusica]

How is this any different from SDXL Turbo models/Lora's on Civitai?

[@NotThatOlivia]

exactly

[@Elwaves2925]

I guess it's supposed to be quicker than turbo models. Personally, I've found the turbo models lacking in quality (although others haven't) so I don't use them. The speed increase isn't worth the qualty loss for me and it's not like they're slow on my system. I'll stick to normal models.

[@procomp113]

turbo is trained on 512x512 I believe (at least the base model). This one is trained 1024x1024

[@Elwaves2925]

@@procomp113 That's a good point. I didn't think about training resolution.

[@gotenks157]

@@procomp113 turbo is trained over sdxl, so 1024.

[@gordonbrinkmann]

Since you mention not having Juggernaut version 9 yet, I just downloaded and made a series of 10 portraits with a starting seed and set to increment so I can use the same to compare it to version 6, which was the one I had before: in my opinion, the results of 9 are high quality, but much too similar over the 10 portraits. With version 6 they had much more variation, so I guess for some things I'll stick to the older version.

[@eurgaptis]

is it available on A1111 ?

[@raven1439]

is there an on/off plugin for comfyUI that, for example, you can turn on the upscaler or any other nodes?

[@thechadbuddha]

you can bypass(right click note bypass) or mute the node (ctrl+m)

[@Emonk2010]

You can mute or bypass nodes by selecting them and pressing ctrl-m or ctrl-b - if you have nodes grouped you can right click on the group and choose the menu option to bypass the whole group

[@raven1439]

@@Emonk2010 I didn't know that you can group nodes, it will make a lot easier, thanks for the answer.

[@groucho828]

You could also check out some of the RGThree nodes for example if you want to add a node that can quickly mute/bypass/enable, not sure if it's that much faster than keyboard shortcuts but could be helpful if you're doing something repetitive

[@raven1439]

@@groucho828 I didn't know these nodes, thanks for the hint

[@hc5843]

hi where to get sgm_uniform scheduler?

[@rolarocka]

I bet if u add deep cache node right before upscaling it's even faster 😁

[@OlivioSarikas]

interesting, i will try that

[@StrengthCoachFelix]

I only use LCM workflows anymore. Usually a 2 step into ultimate sd upscale. Excited to try this one out now

[@97BuckeyeGuy]

Am I a bad person if I say, "meh," to this? I mean, I don't get it. It's not THAT much faster than SDXL Turbo (4 steps vs 8) and the result isn't something better. Also, I *really* wish they would develop more of the SD Cascade path. Cascade does SO much better with prompt following than even SDXL. Prompt Following is the real advancement we need, now.

[@NotThatOlivia]

SDXL TURBO has same amount of steps, but - no commercial use allowed, in this case - something similar with commercial rights (model on Huggingface is for research purposes only) so ... what is the point?

[@Elwaves2925]

Ignore my other post, I misread your comment. Obviously it's for commercial use if it has those rights.

[@OlivioSarikas]

Turbo with 2 steps?

[@HistoryIsAbsurd]

No offense man but you keep saying 2 steps this 2 steps that your video had 4...@@OlivioSarikas The quality is pretty poor on the 2step model NGL.

[@J3R3MI6]

@@HistoryIsAbsurdshut up nerd

[@NotThatOlivia]

agree@@HistoryIsAbsurd

[@billybouki]

Anyone tried img2img with this ? If the resample is above 0.66 competely ignores the input image , if below is identical with it.

[@NotThatOlivia]

Superscale HAS VISIBLE tile difference - and no matter how you set it up - no 100% clean final image can be achieved (depends on the initial upscaled image very much)

[@tradehut2782]

I feel this video should have been longer , showing how this lora does with different art styles.

[@OlivioSarikas]

you can use the lora on any SDXL model. instead of me spending hours testing different models and prompts you can simply download the lora and test your favorite models and prompts

[@FusionDeveloper]

My favorite SDXL Turbo model is "UltraSpice".

[@tradehut2782]

Same

[@valomar00]

Mine too! It's my go-to jack-of-all-trades. There are other models I use if I want a specific look/style. After copious amounts of testing, UltraSpice is the most consistent.

[@lucianodaluz5414]

So a basic question: It only works for Comfy Ui?

[@Kryptonic83]

looks like the lora files are now safetensors so i'd image will work fine with a1111 but haven't tried it myself.

[@lucianodaluz5414]

@@Kryptonic83I tried here, it works, but only with 8 it worths.

[@totempow]

This is pretty cool. Pretty, pretty cool. Thanks!

[@Mike-xe8ql]

The loras already have safetensors version.

[@WackyConundrum]

How do you compare Lightning and Turbo?

[@OlivioSarikas]

Turbo needs more steps and thus is slower. But they both have their ups and downs

[@MarekNowakowski]

20seconds for 4steps on 3080ti? that spounds very slow... faster than my old gtx1080, but barely 3times?

[@Steamrick]

I think he meant including upscale

[@OlivioSarikas]

yes, obviously including upscale. i think i did mention 2048 x 2048

[@MarekNowakowski]

@@OlivioSarikas I understand, but I do 3440x1440 20sec/iteration. I really expected more from 3080ti, I'll tamper my expectations for a GPU upgrade

[@nic-ori]

Thanks! Useful information. 👍👍👍

[@OmerAbdalla]

Looks like they already replaced the .pth files with .safetensors files which are quite large.

[@gordonbrinkmann]

No, the safetensors files have the same size. Rewatch the video. There are large "unet" files (which are full checkpoints) and small "lora" files - which are the ones Olivio shows in the video. They have only added checkpoints without "unet" or "lora" in the name and they are a bit larger than the "unet" checkpoints.

[@486DX]

I think negative prompt not working with this.

[@hk78901]

Euler is pronounced "oil-er" by the way!

[@Burnrate]

Yoooler

[@blacksage81]

This would be cool if I cared about speed over quality. However, my generations need to be commercial ready, and all these experimental models are fun to play with, but I am not lighting up my gpus daily to play, I'm in this for real.

[@OlivioSarikas]

well, there is a reason there are rockets and trucks. one does not replace the other

[@valomar00]

I think you could use your standard work flow as usual, but then use this for the upscaling, detailing, etc. Correct me if I am wrong.

[@motherofallemails]

Aren't these highly biased though? I tried a 3 step one and it's just the same face same haircut even.

[@guns1inger]

Fast, but the quality is meh, and the accuracy quickly drifts away as the prompt gets more complicated.

[@RaumBances]

Can you please include Forge integration information in to your future videos for SDXL Models? So far, it is the only interface that actually let's me run SDXL effectively. Thank you for covering Lightning.

[@piratediffusion]

I'm Loving It (tm)

[@glitter_fart]

:)

[@khaledmoshref1385]

How to get ultimate SD upscale

[@OlivioSarikas]

install the node pack for ultimate upscale

[@geterpriffin]

This is very fast! Granted not the best quality, but you can really take advantage of quick prototyping in 75% less time. You can always Image to Image with more steps (& other settings) for best quality.

[@alcocat]

Talking quality it's garbage same as cascade, for now...

[@OlivioSarikas]

you are missing the point and also missing that this is a new release. If we expect every tech to be perfect before it is released, there won't be any updates for the next 10 years. So lower your expectation and be glad we can play with this for free

[@alcocat]

@@OlivioSarikas "for now" . Sdxl was also bad on start, now it rules. Just keeping that in mind.

[@Cylume.]

No

[@odderphase]

This is just another useless gimmick made for flexing and showing "what is possible". Upscaling is NOT the way to get proper details and it will never be. Check out the upscaled image again...differnet eyes, bad teeth... and for what purpose? Just to show something can be done in 8 steps or whatever? AI img gen is going in the wrong direction with turbos and similar stuff, bc it's promoting bad quality and "time saving"? I guess amazing, detailed and correct images were the main goal, not some rushed results made for the sake of demonstrating wrong "tech achievements"?! Not a single serious creator is using turbo models for their works. People just don't understand CGI fundamentals, so they are easily accepting almost anything as "amazing"...

[@jl25735]

turbo models let you explore a bunch of results before working on refining the creation. This lets you explore prompts, seeds, etc. The way I see it, you don't use LCM to get a final result, but it's great to begin your work.

[@OlivioSarikas]

Just because you don't know how to use it, doesn't make it useless. There are many use cases like live render. Fast preview, use in cases where quality matters less, live canvas design and more

[@odderphase]

@@OlivioSarikas Oh man, really? 🤣 You are too old to react as a 16yo kid, but anyway... Every single comment which is not supporting EVERYTHING regarding AI is automatically responded as "you don't know how to use it", haha, sure! Exactly the same situation with exactly the same sentences as on CivitAI. I don't get why people get offended when someone has a different opinion. BTW all of your reasons are useless to me...fast preview (of what), quality matters less (when exactly)...? I understand that every topic in AI is pure gold for you as a YT creator, but dude... 99% of the things in this world are NOT "amazing", so I really don't see how every single video about AI is "SO AMAZING" for creators... I guess most of the people forgot about clickbaits... Now I will step on your toes even more... ComfyUI is a waste of time if you know anything about img creation, editing and working in certain Adobe products 😊. Before you start screaming at me, let me explain... one month invested in learning Photoshop is 100x more useful in anything related to img creation, than that same one month wasted on learning how to connect hundreds of nodes and create dozens of "workflows" in ComfyUI... Now I have my popcorn prepared, bc I guess you will try to "teach me how to live" immediately, or just delete my comments 😊 Cheers!

[@149315Nico]

The result itself may not be amazing, the inference time to get there tho is. Not everything is about the perfect award winning image, and this reduces time for testing different prompts/Loras. Sure you will use a normal sampling method to get your final image but this will save a lot of time until you reached that point.

[@149315Nico]

@@odderphaseolivio has photoshop skills. So do most people. Difference is, many users in ai are able to leverage their PS and ai skills. You seem not to be willing to learn ai skills, so this will bite you in the future as adobe can also leverage ai, using the same fundamental models for CN as a1111. Almost everything in firefly is a direct copy or heavily inspired by lllyasviels controlnet and the original inpaint/outpainting code for SD. I agree with you that PS skills don’t hurt, but what also doesn’t hurt is knowing how to automatically detect hands and reredender them in higher res so you don’t have to do that manually each time in PS. The stupidest of individuals believe that they can either use ai alone or ps alone to create fantastic art, the smart ones will let machines do the boring stuff so they can focus solely on the part of image creation that is their main drive, be it composition in which case you will need tools like Controlnet, or adding small Storytelling details into your image, in which case ai will help you with the basic image so you got more time for that. You’re like a guy saying mathematicans don’t need calculators. Yes you can do complex equations by hand, but your boss will hate you for taking twice as long to accomplish anything than the other employees.

[@Renzsu]

Made by Bytedance… no thank you.

[@harshitpruthi4022]

What is problem with that

[@Renzsu]

@@harshitpruthi4022they make tiktok, a data harvesting app masked as fun social media. I refuse to support or use anything that comes from them.

[@Drokkstar_]

Thanks, didn't know that! Very interesting wiki-hole on them.@@Renzsu

[@lefourbe5596]

I Guess xformers and llama by Facebook is fine then.

[@MilesBellas]

​@@harshitpruthi4022 It's PRC : TT etc..

[@GamingInfested]

no, tried forge last its trash basically a1111 still best choice for stable, stable diffusion

[@EriIaz]

What?

[@ClashMafia]

Forge is better if you have less than 8gb vram

[@lefourbe5596]

What*2 ?

[@lefourbe5596]

A1111 can't run SDXL as fast as Forge. On a 3090 it takes me 3x less time to highresfix from 1k to 2k. It's comfyUi level of effeciency. I Can go much higher res since the VAE decoding is not tanking my VRAM anymore.

[@benjamininkorea7016]

@@ClashMafia I have a lot of problems making things work, especially some extensions.

[@unknownuser3000]

I just don't get the love for node based editors. Comfy seems terrible of a ui.

[@thechadbuddha]

skill issue

[@benjamininkorea7016]

Everything in ComfyUI is reduced to its inputs and outputs, which means there are no limitations in how you can use everything. A1111 has options like "Send image to inpaint" or whatever. A1111 you can send a single image to a Pose preprocessor, a depth map, an instruction to save the image, an inpaint image input. This allows for very specific creative workflows that aren't possible in A1111. It does take a bit of a computer-brain to understand what to do with all this freedom though. I can see why many might not consider it worth the effort to figure everything out.

[@HistoryIsAbsurd]

Its about speed and modularity. Not everyone likes A1111 and how poorly optimized it is. If you have even the slightest bit of coding experience comfyui is a game changer. You can program nodes yourself. Its not just about AI images or Video. If you cant see the benefit to that than i dono lol.

[@sn0wbr33z3]

Might be a good time to get into Invoke. Invoke 3.7 supports shareable linear workflows. Which means users can share their workflows, hiding their complexities and expose only important nodes as panels. Invoke community content is lacking so far compared to other UIs, but it can be an ultimate tool if people start investing time in it.

[@OlivioSarikas]

it's a professional thing. If all you need is standard click-to-render wait for the later version of a A1111 extension. This is comfyui, because comfyui is almost always first with new tech

[@lennoyl]

i prefer to wait for safe.tensor versions of the files.

[@gordonbrinkmann]

They are there now.

[@LordDweedle]

more comfyui. can we have for forgeui its better

[@Stablediffusion-ai]

Thanks!

[@OlivioSarikas]

Thank you for your support 🙂

[@Stablediffusion-ai]

You are welcome @@OlivioSarikas we love your work you deserve 10x more