THIS is Crazy!!! Juggernaut XL Lightning in only 4 Steps - Automatic 1111

The new Juggernaut XL V9 + RunDiffusion Photo 2 Model is out of this world! It can render high quality images in only 4 Steps. It does NOT need the Lightning Lora anymore can natively used in Automatic 1111 (A1111). It is the best SDXL Lightning Model
Follow me for more AI news: / oliviosarikas
Support me for awesome Rewards: / oliviotutorials
#### Links from my Video ####
civitai.com/models/133005/jug...
/ stevewarnerfl
Upscaler: huggingface.co/gemasai/4x_NMK...
Best settings:
Res: 832*1216
Sampler: DPM++ SDE
Steps: 4-6
CFG: 1-2
#### 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

[@ColorblindAdam]

Thank you for liking my Dragon Prompt! :) I created that while testing Jug9 with Kandoo. I appreciate the shout out for the image.

[@HansEgonMattek]

Kandoo's models are always ahead of everything else. It is probably due to the massive amount of Images and steps used while training.

[@freekhitman9916]

Great Video @Olivio Sarikas ;-) I also love the new sdxl lightning this will be the new standard till SD3.... this is why made my own Lightning model called Mjölnir don't need any lora too but higher cfg 2,5-3 at 4 steps

[@nietzchan]

This is crazy! Thank you for sharing this as always

[@71Jay17]

pretty massive improvement! amazing stuff

[@elysilk4538]

Beautiful! I like to test realistic models with coral reefs, small fish, and a crab or two. I have never seen better! Thank you and the developers.

[@FuriousDevi]

There is also a Dreamshaper XL Turbo version that creates beautiful fantasy art with 8 steps in 1024x1024 in about 8 seconds on just a 3060ti (using Forge)!

[@ksiobrga]

Thank you for sharing! Keep it up ♥♥

[@user-qg4gx7be2j]

Thank you! I checked it out and it works!

[@ColorblindAdam]

Thanks for trying it out! :)

[@renierdelacruz4652]

Great Video, thanks very much

[@electronicmusicartcollective]

yesterday say when the juggernaut xl is comming and one day later voila you tell us this great news. TOP. thx

[@krawlak]

The pictures you created look awesome! I wish there was more variety though when creating several variations for the same prompt. Do you have any tips on how to set parameters to get more different results to choose from?

[@undergroundxp]

Absolute game changer! Now imagine this speed with SD3, truly next-gen stuff!

[@MilesBellas]

Bytedance ?!

[@albert2006xp]

Speed is never really important unless you're generating massive Animatediff videos. Quality is more important for photos. These are great but they're pretty niche use and not game changing.

[@kpcreative2542]

Strongly disagree on this point. If you think from the standpoint of design theory, than exploring different options and iterating on an idea/concept is essential to creating the highest quality image/artwork. The faster you can do this, the faster you can reach the best possible design. @@albert2006xp

[@zaq_hack4987]

@@albert2006xpIf you play AI Roguelite, this is 100% game-changing.

[@Tofu3435]

@@albert2006xp unless if you run in CPU mode

[@NAKOOT]

Thanks Olivio 🤘

[@NineSeptims]

Is this magic? if it gets any faster we are gonna opening some doors into some really cool possibilities! Thanks for keeping us so up to date Olivio!!

[@edmoartist]

Try using the 2 step Lightning lora at 0.5, works great at 4-6 steps. Realities Edge Lightning is great too

[@runebinder]

Thanks, I'll be checking this out. I take it this as easy to setup in ComfyUI and can be used in Fooocus with not special setups as it looks the same as loading a standard SDXL model in A1111?

[@user-je7qy5ey3y]

It's so beautiful and so fast and the most is everything low

[@pmtrek]

Nice. Do you plan to do a similar vid for ComfyUI workflow ?

[@Gabriecielo]

Juggernaut is my favorite, now even more!

[@ColorblindAdam]

Thank you so much. :)

[@BeetjeVreemd]

Thanks for this one, i'll try it :) By the way, you say the model is 6.6GB, but i only see a 7.11 version of the model on huggingface? Is that the right one?

[@Tempestan]

Wow, just tried it, simple short prompts, no negative prompts, sdxl styles used photographic, suggested paramaters, have a 2070super. Took a couple of seconds and I could not tell if it was AI generated or real. Kind of freaky because I used simple prompts to describe my first girlfriend and it looks almost 100% like her, no joke. Reality is going to get a lot more complicated from now on.

[@igor_timofeev]

Cool, wonna try it in Fooocus

[@cyborgmetropolis7652]

I played around with the model the past couple of days and it's very cool. Can you explain how you can get the same scene but from different angles, like from that bedroom scene? Is that just using a different seed or something else?

[@PADATWO]

Hello, do you know how to use emote portrait alive? A tutorial on this would be possible? I think we should be able to use it with ComfyUI, right?

[@FusionDeveloper]

I saw it and will give it a try and see how it compares to UltraSpiceSDXLTurbo 1.4

[@tdslaine]

are the outputs with hotshots/animatediff viable to use on low steps models?

[@paulsheriff]

What is minimum spec ect vram as only have 6vram so can only get a small image any bigger just crashes

[@AIKnowledge2Go]

Thanks for the video and keeping us informed, Juggernaut XL Lightning is great. In my opinion Dreamshaper XL produces slightly better images. Takes 3 - 6 Steps though. Can't wait for SD3 to come out to improve this further.

[@Elwaves2925]

As someone who felt the turbo models sacrificed quality for speed, I've been using two lightning models (this and RealVis) all afternoon and they're great. Not sure why it's different but for me it is.

[@Powerlevelover9000]

Same for me, Lightning is simply faster and of much better quality than turbo imo. Tho for realistic to semi realistic checkpoints. In anime related checkpoints, I see some reduction in quality, like when combined 4step lora with animagine checkpoint, it does work but with somewhat compromised quaility. But for realistic checkpoints, I see minimal compromise. But even in animagine case I use that for sd ultimate upscaling along with anime tile control net lite, it's pretty decent in that.

[@OlivioSarikas]

Lightning for president

[@lefourbe5596]

​@@Powerlevelover9000i love every mention of Animagine XL :). But you mentionned Tile in the upscalling process ? Is Tile XL a thing or do you use SD15 for it ?

[@siddharthmishra8283]

Awaiting for your SUPIR installation for A1111😊 12gb version

[@wykydytron]

I still think best is pony and autismmixes, due to their ability to do pretty much any concept out of the box. I find it hilarious that models nade for furryporn are extremely competent in everything and understand most stuff even with super long prompts and mixing natural language with tags. If only we could have ponys understanding and juggernaut realistic/3d abilities.

[@AgustinCaniglia1992]

Dreamshaper lighting also works with no additional lora it seema

[@GS195]

That's great. I just wish hands were easier to generate (they still aren't).

[@andrewUA1234]

Hi! Can you make comparison Turbo-Lightning-LCM? Thanks!

[@pvgmartin]

Olivia what was the prompt on the crystal Dragon? It's just beautiful

[@Powerlevelover9000]

It's on the civitai model page.

[@rickarroyo]

All prompts are in the CITAI link (in the description of this video). Just click on ( i ) on each image from the creator that has all the parameters used.

[@ColorblindAdam]

we put it on the civitAI page for everyone. I am glad you love that one! I created it while testing Juggernaut for Kandoo :)

[@Showdonttell-hq1dk]

The Lightning models are really great fun. Normally the IP adapter embeds need very high CFG values and you have to wait a long time for the images to be generated, but with the Lightning Lora alone it's incredibly fast and in high resolution. Together with the Controlnet, this is a fantastic combination. I've just downloaded the Juggernaut XL Lightning model and am about to start testing it. Thanks for the tip! :-) Update: Hm, doesn't seem to work so well with the suggested resolution of 832:1216 in ComfyUI. With the recommended DPM++ SDE Karras sampler and 1024:1024, however, excellent and better than with the old Juggernaut model and the Lightning Lora. The generated images are created in exactly the realistic style I prefer and even from painted source images. Cool!

[@ViscLive]

with VAE or without?

[@davidrosbo4519]

it looks like it will work in forge which is awesome. no matter what I do in automatic 1111, comfy it takes about 5 mins to render an image. But Forge works great for me taking about 5-7 seconds.

[@OlivioSarikas]

Yes, works in forge the exact same way

[@user-qg4gx7be2j]

+ Thank you! I checked it out and it works!

[@Shackrah]

Wow! Can you provide the prompt for the dragon? I want to make more of those and to experiment around that!

[@OlivioSarikas]

All prompts are on the civitai page of the model

[@trkx444]

Can this be installed on a Mac OS? If yes, is there a link?

[@marcihuppi]

it gets faster and faster so fast :D

[@OlivioSarikas]

Crazy, right.. Realtime is possible with lcm also already

[@MarcSpctr]

Apparently Bytedance has a cluster of over 12k A100 GPUs. How the f did they get access to that ? According to EMAD they can train LLM models like GPT3 within 2 days time. and here StabilityAI is struggling to get enough compute to train their Video Model.

[@lefourbe5596]

A100 is two+ years old and bytedance is under the hood of TENCENT the biggest chinese mega Corpo. They hog the demand for GPU for 4 years or so leading to the US restriction to NVIDIA. Basically NV should not be selling the best card in china anymore. Of course they have 12k unit of theses GPU.

[@Digital_Toolbox]

Will it work well In comfyui?

[@_gr1nchh]

Will this work in ComfyUI? It's definitely working, I can see the progress looks fine in the preview. However when it decodes into the saved image, it's very distorted and comes out blue and not at all like how it looked in the preview window as it was denoising. Any ideas what could be causing this or this just doesn't work in Comfy?

[@OlivioSarikas]

Yes, 8 already have a video about that on my Channel for comfy. Just in this case you don't need the lora

[@_gr1nchh]

@@OlivioSarikas Interesting. I'll go give it a watch! Thanks for all you do for the community, Olivio!

[@mayorc]

So using the same model, you need the lora in Comfy but not in automatic1111? @@OlivioSarikas

[@RS_0_7______]

does anyone know if we can use stable diffusion for freelancing and professional work? like are the artworks we create allowed for commercial use or not?

[@jdesanti76]

Depends of the model, no all have free commercial license

[@dendenisification]

and some upscalers can't do commercials too(according to it's license)

[@lefourbe5596]

You have to produce actual meaningful content. A pretty pic is easy to generate but a coherent composition and consistency is not. I, in my case Can train a concept for your need and your own use, i provide the computer and the Time Taken for the training data and then i deliver it. I Can do character sheet, i Can do some complex scenery with extra times

[@yolcuadem]

Is there a code link for Kaggle?

[@kaiz0099]

Does it work on forge?

[@Cyberdjay]

Works great in Forge 1111 but in my standard automatic the images take over 1 min with the same settings. I am unsure why? (6.6 secs on a 3060 ti on forge vs 1 min 30 sec on standard 1111)

[@Powerlevelover9000]

Forge is simply much more optimised than auto 111. It is so optimised. In fact, if you use "never oom extension," you can run sdxl in meagre 2 to 3gb Vram, or even less.

[@OlivioSarikas]

Forge is more optimised for speed

[@Cyberdjay]

@@OlivioSarikas Thanks Olivio. Thanks for all the great videos to keep up with the speed of progression

[@Skettalee]

I dont understand, I downloaded this model from the civitai link+ extension and i put all the same settings you showed and this thing is taking forever to generate my first picture of "a baby monkey rock start, highly detailed, 8k, cinematic lighting" and it took 6 minutes and 33.7 seconds to render the image. isn't it suppose to be fast? Because any and all of the other methods I've used A1111 other models, comfyui, forge, fooocus; they all work way faster than this with my RTx 2080 SUPER. What would i be doing wrong?

[@NotThatOlivia]

1 - do not use automatic 1111 (like never ever) 2 - Lightning SDXL is ONLY Great for closeup generations 3 - learn ComfyUI for real magic

[@aynrandom3004]

@@NotThatOlivia so it does not work with a1111? i'm confused. Title says a1111

[@lefourbe5596]

​​@@aynrandom3004on A1111 it works for GPU with 12Gb of VRAM for SDXL model (like Juggernaut here). It is not your old usual VRAM friendly stable diffusion 1.5 anymore (like dreamshaper 8). Your weak 2080S will suffer MASSIVE SLOWDOWN due to lack of video memory. That is why @OliviaFromTrivia told You to dump A1111 forever (or until Automatic fix his code). ForgeUI and comfyUi are MUCH more optimized for your GPU. I expect you to divide by 15x the Time to generate a picture by making the Switch. ForgeUI is basically your A1111 UI but fixed with more extension built in.

[@LouisGedo]

👋

[@Ben-hu2jp]

But it's slow. One of the slowest samplers is DPM++ SDE.

[@gicknardner]

Low CFG and short prompt means you're not really get a lot of input into the results, right? Fewer steps but definitely at a trade off.

[@Powerlevelover9000]

In lightning base model cfg 1 = sdxl base cfg 6 as it converges. Well either way you can use dynamic cfg extension to boost cfg value without effecting image quality.

[@ViscLive]

my eyes and mouth look bad. I have adopted the settings exactly. Why? Any idea?

[@lefourbe5596]

You should try the extension adetailer. It does an automated inpainting on the face at a higher resolution.

[@ViscLive]

Thank U 🙂@@lefourbe5596

[@angrypolygon8410]

Can this work in Forge UI ? I switched to Forge because of you now...should I regret that???

[@Powerlevelover9000]

Ofc it can work on forge. Forge >>> a111.

[@spartanx169x]

Yes, but is it any better with hands, feet and limbs?

[@OlivioSarikas]

You can fix hands with controlnet check my channel for hand fix video

[@DoctorMandible]

Juggernaut has restricted license

[@squirrelhallowino29]

Now all we need is a prompt generator, let it run overnight, pick the best images, make it into amazing stunning pictures w post fx, profit?

[@OlivioSarikas]

You can do that with a llm or wildcards or prompt batch

[@RetzyWilliams]

The lightning Lora is better because it can speed up any model, including all the Juggernauts.

[@bluecloud111]

Is it possible to use this in Comfyui?

[@chillsoft]

ofc

[@alcar32sharif]

Image generation with lightning is now 4x faster then upscaling and nearly 5x faster than VAE decoding. I need a faster upscaler and a faster VAE decoder. 😂

[@OlivioSarikas]

You can use the upscaler with the same. Settings. Makes ultimate upscaler super fast

[@YouMostTubeWanted]

@@OlivioSarikasbut does ultimate upscaler work with forge? that would be huge!

[@marcos13vinicius11]

How can I use it in comfyUi?

[@OlivioSarikas]

Just use it like a sdxl model with the settings I showed here

[@markh5268]

How is an open sourced model offering commercial license when it's trained on content they don't have the rights to? Is this a grey area that is being exploited right now?

[@CHARIOTangler]

GREAT question🧐👍

[@MilesBellas]

Bytedance = don't question, imitate the dance/behavior pattern/paradigm ?

[@martin231181]

Definetly a grey area. But stable diffusion itself can be used commercially so if this is not a breech of the copyrights on the material it’s been trained on, then neither is this. But if it is a breech then so is this. So in that sense it’s no difference than Stable diffusion in general.

[@markh5268]

It's like bippers going in stores in SF and taking stuff under $950 and then selling it to the resellers pawning it out on the sidewalks.

[@thejhonnie]

Yup. 😂

[@RenoRivsan]

I dont like this model for img2img.. wahts the best prompt for beautiful faces

[@serioserkanalname499]

Forge on low vram cards + lightning models/lora already made sdxl more than twice as fast as last year... and we're only 2 months into 2024 :O But... I cant help but feel like lightning does worsen the image a little compared to normal.Also, can we at this point say lightning is just better than lcm?

[@Powerlevelover9000]

From what I have seen lighting is much better than turbo and lcm. If you need more quality jsut use the 8 step lightning lora.

[@goobydntfear]

bruder, lightning ist donner, lightING ist beleuchtung

[@ernstaugust6428]

Wrong. Lightning is Blitz

[@snatvb]

RTX 3070 - stucks on start, just loads full memory and steps hang out maybe smt wrong locally

[@Powerlevelover9000]

Even on a111, 8gb vram should be enough for this. Either way, try forge(similar ui to a111, buy much better optimised) or comfy ui.

[@snatvb]

@@Powerlevelover9000 yes, I tried both, maybe because my monitor is 4k, or something broken inside. I tried dreamshaper lite and it works, but too slow, around 5 min for 4 steps

[@Lykon]

DreamShaper XL Lightning came out days earlier

[@mlsterlous]

I spent a lot of time comparing this model to some other lightning models and noticed 1. When you create closeup portraits, there are some strange lines/artifacts, usualy on sides of forehead, and under nose. 2. This model seems a lot LESS realistic than for example Wildcard-XL or RealVisXL V4.0, portraits just look very artificial(not sure, maybe needs specific promt) 3. Don't know why but CFG 1 renders a lot faster on my 1050ti than CFG 1.5 or 2, maybe someone can explain why?

[@Powerlevelover9000]

Higher CFG leads to higher compute times, it's true for pretty much most models. But some require cfg to be 6 ~ 8 range tonfunction properly. Higher cfg means more adherence to prompt. Sdxl turbo and lighting model, cfg 1 = cfg 6 in sdxl (it converges faster, but still better in performance) Let us look at an anology. You are a sculptor who is a given a blue print, to make a sculpture. In this case cfg is how close you want to be the blue print. The more closer you want to be to the blue print, the more the effort it takes from your side to make it as per the specification of the blue print (high cfg) Bit if you don't need to strictly follow the blue print, you can take on a more easier route, and sculpt is usually close to the blueprint instead of following it to the tee, which would involve relatively less effort (less cfg) Ao less cfg = Higher iterations/s. Hope this analogy helps.

[@mlsterlous]

@@Powerlevelover9000 thanks :) I really noticed this first time. The difference between cfg 1 and 2 looks to be ~8sec/it vs ~16sec/it (1216x832).

[@Powerlevelover9000]

​​@@mlsterlous​if you want even more speed, then you can try "euler sgm uniform" sampler(only for lightning models) its like 70% faster than dpm+++ sde, but you would loose some skin textures, but for non human generation, the quality difference is barely noticeable. Also if you are using A111, try forge, it's much faster and more vram eff, especially for older gpu's like yours. It's ui is basically like a111.

[@Tomasin19]

Can AMD users have a go at this?

[@OlivioSarikas]

A1111 runs with amd

[@VisualWebsolutions]

i use for my 7900xtx comfyui with directml and i take with that model about 5 sek.

[@Tomasin19]

@@OlivioSarikas sorry - I meant if they can use it with vanilla A1111 cos the civitai page says it runs well on Forge but my 580 can't run that :(

[@mick7727]

You'd have an incredibly poor time with that card unfortunately, if it could run anything. AMD is kinda terrible with local diffusion.@@Tomasin19

[@withcarename]

No local deployment?

[@OlivioSarikas]

It's in A1111. That's a local ui.

[@Powerlevelover9000]

​@@OlivioSarikas I'm just asking. Is there any reason to use A111 over forge ? From what I have seen forge is simply faster an and much more V ram efficient than a111, while giving the same benefits and ui feel. By enabling the oom extension(which I dont need to), its only using 2~3 GB of vram out of 16 GB vram, for sdxl, with barely any reduction in speed.

[@walidflux]

It's not Pony

[@antiplouc]

It's bad. I don't see any reason for using it. The real juggernau model is much better. He even says so in his own post.

[@nokiaairtel5311]

Open source but cannot use commercially? Sorry, I'll pass

[@mick7727]

It doesn't do everything though, and tbh realistic images get kinda boring. Who cares if it's fast, when art is about YOUR quality?

[@mlnj144]

biased images, i dont like it, all similar

[@cliffdickinson7091]

I use to follow you for photo manipulation advise. Affinity photo, etc. You are lost in this ai fantasy world that's not human creation.

[@foulwin9719]

Grow up