Updated the Prompting Guide
For business inquires, commercial licensing, custom models, and consultation contact me under [email protected]
Join Juggernaut now on X/Twitter
Juggernaut Ragnarok on RunDiffusion
Juggernaut XI & XII on RunDiffusion
Prompting Guide for Juggernaut Ragnarok by Adam
Prompting Guide by Adam for XI & XII
A big thanks goes to RunDiffusion and Adam, who diligently helped me make it work :) (Leave some love for them ;) )
Hey everyone,
It’s been 8 months since the last version was released here on CivitAI.
Of course, I haven’t been idle during that time . I completed several projects to ensure I’d have the financial means to keep exploring new architectures and possibly do full finetunes on them in the future.
Juggernaut Flux (and its many sub-variants) was a ton of work, but ultimately, I’ve wrapped that chapter up. The training process gave me way too many headaches. To keep my sanity, I spent my spare time working on Juggernaut SDXL with the hope of maybe releasing one final version for you all.
And that day has finally come. :)
Juggernaut Ragnarok has improved in many areas: photorealism, digital painting, poses, hands, feet, and much more.
That said, it’s still an SDXL model, and I don’t recommend comparing it to models like Flux, Reve, or Sora. For example, it still has limitations when it comes to text rendering or faces at a distance.
I recommend using it as part of a pipeline for your projects. Example setup:
FluxDev / Pixelwave / Jug Flux Pro → Juggernaut Ragnarok
A quick personal note about Juggernaut:
Honestly, I don’t know what comes next.
After the release of Sora and similar tools, the open-source image generation space feels a bit dull in comparison.
Nothing has really excited me enough to dive back into training (yes, I’m talking about HiDream too).
So I’m seeing Juggernaut Ragnarok as a kind of farewell, especially since it’s unclear where things are headed with CivitAI in general.
(You can download all Juggernaut versions from HuggingFace, by the way.)
Last but not least:
Have fun with the model, share your creations, and good luck with your projects!
And in case you’re wondering: Yes, you can do anything you want with Juggernaut : merge it, train it, sell the image outputs, etc.
Just a simple shoutout is all I ask. :)
And now, here are the recommended settings:
Recommended Settings(VAE is baked in):
Res: 832*1216 (For Portrait, but any SDXL Res will work fine)
Sampler: DPM++ 2M SDE
Steps: 30-40
CFG: 3-6 (less is a bit more realistic)
Negative: Start with no negative, and add afterwards the Stuff you don´t wanna see in that image.
VAE is already Baked In
HiRes: 4xNMKD-Siax_200k with 15 Steps and 0.3 Denoise + 1.5 Upscale
And now, have fun trying it out. As always, I'm eagerly waiting for your pictures in the Gallery :)
If you liked the model, please leave a Like. In the end, that's what helps me the most as a creator on CivitAI. :)
Last but not least, I'd like to thank a few people without whom Juggernaut XL probably wouldn't have come to fruition:
Dreamlook.AI (Trained 3 Side Sets)
Description
Recommended Settings:
Sampling Method: DPM++ SDE
Sampling Steps: Recommend 6 but 4-6 possible.
CFG: Recommend 2 but 1-2 possible.
Negative Prompt: None
Prompt Size: Do not exceed 75 Tokens
Tips: Use more tags with Hyper than the standard model.
FAQ
Comments (203)
AttributeError: 'Options' object has no attribute 'k_sched_type'
Time taken: 0.7 sec.
Damn!! thank you very much for this 😁😁😁😁💗
I tested is not working with lcm Especially with my Laura bubble watercolours LCM He come too much Sharp But still it's amazing
I saw your image in the Gallery. You used a too high Stepcount. For the Hyper Model 4-6 is way enough
Hi I'm using your fantastic lightning model, as Sampling method I imposed DPM++ SDE, like Schedule type what I have to set since the webui requires a choice between: Automatic, Uniform, Karras, Exponential, Polyexponential and SGM Uniform?
Many thanks in advance.
You can leave it as automatic, with dpm, karras or exponential should be fine.
The I believe the official scheduler for Lightning is SGM Uniform but others can work well too.
@BilboTaggins thanks, actually, I tried all alternatives and I found SGM Uniform the better too, but wanted to know others experiences. Thanks again
I think Karas works better for human skins. Especially with freckles. Freckles look like painted spots with SGM Uniform.
Anyone get Hyper to work with Forge? Getting errors. :(
Yes I do all my testing and prompting in forge. I use Automatic 1111 (Forge) on RunDiffusion and it works great for me.
@Colorblind_Adam Cool, I'm sure I'll sort it. Thanks for the reply, cheers.
Its work great on Forge
How do I get similar or even decent results in ComfyUI? I can't seem to reproduce these examples in ComfyUI at all. Always comes out horrible.
Thank you.
Jug 9 is my go-to favorite model. That's why I've spent so much time reviewing the recommended guides trying to get good results out of Jug X. Like many others, I have only been able to do it occasionally after a lot of effort.
I think the idea of making a model that is super responsive to extreme levels of specificity is interesting! But I've actually found that anytime I've gotten a prompt to be thorough enough to work on Jug X, if I switch to jug 9 the result is similar and sometimes better. So it's unclear what Jug X is adding vs the fact that I just spent 20 minutes refining one prompt. I'm confident there are improvements, they're just hard to manifest.
I've observed a trend of separating the level of prompting-difficulty from the assessment of model quality, but intuitive prompt-steerability IS part of the model. And if the model is difficult to steer, then that's a flaw to iterate and improve on. This is 10X as true if it's a model intended for professionals, which should involve some level of user-testing with said professionals who haven't used the model before. Otherwise, the risk is that saying the model is "for professionals" functionally just lowers the quality bar during training on making the model intuitive to use.
To end on a positive note, because the Juggernaut line really is my favorite model line, it could be helpful to create a GPT or something to assist with prompt-design. (could be the creator or someone who's managed to get consistently good results and understands the prompting!)
I agree completely. I find myself using Pony often because it's incredibly steerable. Juggernaut will get a better overall picture quality, but it can be frustrating if it doesn't quite understand what I'm trying to do within the first 5 minutes.
Sometimes I just use both with ControlNet crossing them over, though SDXL ControlNet seems to be missing some features/quality compared to 1.5. Perhaps there's updates I haven't found? Idk. I just get off coloring with depth and of course OpenPose still isn't an option afaik. That's kind of a side rant though; the point is I have to work with another model and work around the problem instead of doing it all-in-one with Jugg from the start.
Hi Ivorsky thanks for your nice comments. It is possible to create GPTs to create prompts for models. I have done so for multiple models including Jug X. Have you had a chance to look at the new guide I just released? It should make designing your prompts a lot faster and easier so you don't waste time and can be deliberate. Jug 9 is a beautiful Model and I too love using it. Jug X however has a bit more to offer in terms of control. Remember this is the new base that future iterations will work off of and it'll improve :)
@Phraxas Hey could you check out the new guide I gave on how to write prompts for JugX? I have found it incredibly responsive if you talk to it the way I wrote int he guide. I think that will help you a lot. Now that I have the "magic formula" for me it works great. Hope that helps guys.
@Colorblind_Adam yes, I had read it this AM before writing my comment. Thank you for putting it together. To your credit, your guide helped me get from basically not knowing what I was doing over the past week to being able to get some good images. Though it also helped me see just how much effort it was going to take each image. I think I'll try what Phraxas mentioned: using it with controlnet and Pony and see how that helps things.
I think 𝕏 + hyper lora looks cleaner than the merge
its a good prob to have, when I have to challenge myself to not merge Juggernaut lol Well done friend
is this the support you requested? just adding a comment on the civitai confirming it's a good checkpoint?
Still one of the cleanest models I’ve found. Although I’m still trying to find the best loRA to match it for realism
nice
I have older version of this Model maybe from Oct/Nov 2023. Should I delete the older one after I download this later one or what?
I mean will this new Model create same/similar faces like the one in the previous Model version?
will it create same faces? maybe but mostly likely it will be different since newer model has a different face neural network that interprets noise differently
@aiaicaptain so what is your personal choice or opinion or preference? I mean, lets says if you downloaded a Juggernaut Model back in January and now in April a new version of it was released and you downloaded the newer one, do you keep the previous version as well or you delete it?
@Loneranger87 I would delete older models. I can always get them online if needed be. Saves space and confusion on what to use. I keep just 1 turbo, one base, one lcm and one inpainting model. that is all you need.
@aiaicaptain Right. Got it.. Thanks for your suggestions and comments and your time. Appreciate that. 😊
@aiaicaptain @aiaicaptain btw, can you explain a little more about these Turbo, LCM (which I think is LORA) and Impainting model. I mean I know that there is LORA for improving art style furthermore during SD Art Generation, but is Turbo? Can you suggest a couple of them for Juggernaut lets say?
In a lot of ways I like the images I'm getting with the X/Hyper model, EXCEPT - it seems completely unable to render hands. Over dozens of generations with steps from 4-6 and CFG 1, 1.5, and 2, with and without LoRAs, not a single one had normal-ish hands. They're either completely malformed or have 6-10 fingers each. Worse than the worst SD 1.5 model in that regard.
nice
In testing out the X/Hyper models I am very disappointed with the number of images I have to throw away because the number of fingers simply are not 5 in total. 6 and 7 seems to be the norm. Why is this happening. It seems like I'm not the only one.
Although I don't have a problem with the new prompting system to get the composition, lighting and poses I want through prompts alone. I find that Jug X more often than not has issues with disfigured faces on small or distant characters, extra limbs or mangled bodies. It's not awful and inpainting with a different model or running the image through a different model (Jug8, 9 or Zavvy) usually fixes things right up, but it's still an issue not present in previous versions of Juggernaut.
I don't think it's my prompting because I understand the new prompting system well enough to get the desired results, I just find it to be more problematic with multiple subjects and smaller/distant subjects than previous versions.
Nothing that can't be fixed but it's enough of an annoyance for me to avoid using X if I can get the desired composition with other models and I often can. However, Juggernaut X is fantastic when it comes to camera control.
Hey, first of all, thank you for your feedback :) Currently, Version X's dataset consists of 2.7k images. Juggernaut 9 had a total of 15k images in its set. So with Version X, we have only reached about 20% so far. Perhaps it would have been better to wait another 1-2 months and train with 50% of the original dataset, but that would also mean that there would have been 3-4 months between V9 and Version X. However, I am more a fan of regular updates to keep the community involved on this journey.
Exactly the things you are addressing will improve in future versions with more content in the dataset. Version XI will be much larger than Version X and should show improvements in all areas while still being quite accurate in responding to prompts :) So, the journey from Juggernaut on SDXL definitely continues, even after the release of SD3 there is still much to do
@KandooAI wow thanks for the insight, this went much better than my reddit post about it a week or two back :) yeah this makes sense, I'm sure when you guys get it all trained up it will be just fine. Because as far as prompting and it understanding prompts goes, that's great already.
best
I love this model! However, I'm facing an issue with the hyper version where my backgrounds are often dull, requiring extra effort to enhance them. Despite following the prompting guide, suggestions like "in a forest" haven't been effective. Does anyone have any tips?
hyper is a model for faster generation at the cost of slight quality. if contrast/dullness/minor details etc matter to you, I suggest you switch to a regular or full model.
as for prompt "in a forest", keep what is most important first or increase steps. large prompts needs more steps for model to generate all the details.
22
@aiaicaptain144
Tried to use in Fooocus. Very bad result. Maybe I did something wrong
Make sure to use sampler dpmpp_sde_gpu and scheduler karras as well as default_overwrite_step 4 steps. You're welcome to join the discussion at https://github.com/lllyasviel/Fooocus/discussions/2813! :)
@mashb1t Thank you!
@mashb1t This is what ive been looking for. Finally some like-minded people who use fooocus and give some information about the settings 🙏
Clip SKIP?
if they dont specify is mostly -1 but the only other usal option is -2
Layer of CLIP are irrelevant in SDXL
my most used model. however, automatic1111 keeps running out of cuda memory, even with --medvram, and shared GPU memory never gets used like with other models. any ideas why that is?
that is because A1111 code is stupid and need to be revamp, however extension compatibility rely on it sometime (lora CTL and regionnal prompter for example)
for a beter experience you have 3 option :
Forge UI : exact same as A1111 but doesn't fill your VRAM when outpouting, so much more efficient (also faster)
SD NEXT : unlike A1111/ForgeUI get constant improvement and bug fix. also very fast and memory efficient
ComfyUI : node based UI also fast and versatile. hard to use however
@Le_Fourbe gee. thanks for dropping by. I'll check it out
isn't there any inpainting version ?
Hi, I love to use the lightning one, any news about the X version ?
Kando, your model enocurages me to train a model of my own, and for that I have a few questions if you have the time:
1- I have 2000 HQ studio photos of the subject, would them be enough?
2- Do I need to merge the result checkpoint with another model to assure variety?
3- How much money do I need to put into this, roughly?
4- How to even do it? I searched a lot but to no avail.
Thanks.
there are several guides on the home page for this:) see "articles"
A combination of the Align Your Steps scheduler node (at the default 10 steps) + the Perturbed Attention Guidance node (at default 3.0 "scale") + DPM++ SDE GPU as the sampler + a CFG scale of around 4.0 in ComfyUI with any regular SDXL checkpoint (like no Hyper or Lightning anything whatsoever anywhere) is IMO much higher quality 100% of the time and also not slower than any Hyper or Lightning config is.
TLDR the config describe above renders Hyper and Lightning literally obsolete and useless at least in my case, I'd highly recommend people try it if they're currently using Hyper or Lightning based stuff for performance reasons.
not defending all those fast models, even though they have their use, but -
did a lot of XY plotting with all possible combinations of AYS, PAG and ACFG - although increase in speed can be significant (especially with Automatic CFG), drop in quality or change in style just not worth any of it
The Prompt Accuracy Is Great!
That's what I'm saying!
Everytime I try to use the new Hyper model in Forge I get "TypeError: 'NoneType' object is not iterable" every other model works fine.
really one of the most realistic generators in here
How accurate is Inpainting?
It's great!
Question on prompting: How should it be worded if we already have an image we want to use? E.g. take this image of the woman, put her standing behind chair...
I think you need to describe what you want to see. e.g. "Woman standing behind a chair" and put the image you want.
I have yet to try but the PDF guide with the neural map image is epic. Thank you so much for this creation of yours.
nice
INSANE, INSANE, INSANEEEEEEEEE.
Hats off to the developer for this fantastic model and now I have realized how to get good images from a few tweaks of settings here, I'm getting fantastic results, wonderful job :)
Great job!
Very good, very feel
With hyper I get the best quality results from DPM++ 2M SDE Huen + SGM Uniform. Wonderful.
Is this new Hyper model extremely slow? 5 steps, 2 CFG, DPM++ SDE Karras, Automatic1111. 2070 Super. The other Jug lightning was much faster than this. Is this to be expected?
thanks for your work
should we expect future updates on this model or its a done deal?
nice!
I haven't really used SD in awhile. Are images generated with this model supposed to be a bit blurry looking? I can't seem to reproduce anything as sharp and detailed like other users have.
same. idk what the secret is.
4 steps / cfg 1.5 / DPM++ SDE Karras - Use xyz plot to fine-tune as more complex prompts affect how the sampler behaves. Good Luck.
Even if you have EXACT same numbers in sampler settings - the samplers itself can be different and generate poor results with that numbers.
Also recommended settings here - most of the time not correct.
Copy pasted numbers never properly work for me, even recommended sampler type is not correct, you must find numbers yourself for your setup.
If you need crystal clear - start with Euler/, low CFG + high steps number.
@vicksInhaler Prompts have nothing to do with image quality. Type "Box" in prompt and it will be blurry win wrong config. If you will add "box super duper 4k sharp, high quality, masterpiece" - it will remain blurry until you put proper sampler.
Okay, sorry but I'm a little confused about the names of the different versions. Several different questions here.
- What does the "RunDiffusion" in the name indicate?
- As for the Lightning version, the only version I can see is labeled V9RDPhoto2. Is that the only one? I assumed the versions with "V" in them were previous versions. Does the "Photo" in the name indicate a different version?
Thanks and sorry all the questions but it's been driving me crazy. :)
RunDiffusion refers to the company who created the photo model used for training.
Naming Convention:
V9 = the base version of the Juggernaut Model (model, clip, vae)
RDPhoto2 = RunDiffusion Photo Model 2
Lightning_4S = Stable Diffusion Lighting XL - 4 Step model
Hope that helps.
Thank you sir.
@nathanieljgarcia469 Sorry if i'm being dense i'm new to this it's fascinating. Is the Lighting model something which is included in juggernaut or does this need applying as a seperate module so to speak?
@Grimdihlum My answer could be clarified further. I'll use an analogy to explain this more clearly. SDXL 1.0 Lightning 4-step is a checkpoint. Think of it like a suit or tuxedo you could buy off a rack in a store. You can wear it as it comes or it can be tailored to fit a specific person's body. That checkpoint (or suit) was used as the foundation for training/creating Juggernaut XL V9. Think of Juggernaut XL V9 as a tailored suit that now fits just the way it's designer wants it to and has some new styling. The tailor added functional buttons to the cuff of the suit, a new colorful lining to the inside of the suit, and a few other hidden touches. RDPhoto2 was what the tailor used for the lining, functional buttons, and sizing for the suit.
In practical terms whether you buy a suit off the rack or get it tailored you only wear one suit (checkpoint). Loading SDXL 1.0 Lightning - 4Step is the off the rack suit and Juggernaut V9 RDPhoto2 - Lightning_4S is the tailor styled suit. So if you're just doing a standard workflow in A1111 or ComfyUI or some other tool you load the single checkpoint. You don't put on two suits.
Now the analogy falls apart in more advanced workflows where you might use more checkpoints to do more complicated things but in general this should hopefully clear things up.
@nathanieljgarcia469 incredibly helpful thank you!
@nathanieljgarcia469 If only more people on the internet were this cool! Thanks for that analogy!
Is there a specific reason why the Lightning model delivers by far superior images compared to the Hyper model? I tried various samplers, various numbers of steps, different CFG scale settings, but the Hyper model results always look quite poor and unfinished, whereas the Lightning results are just plain awesome.
If you're referring to V9's lightning version I'd guess the difference is how you are directed to prompt for version X as it was re-captioned with GPT-4 and requires really precise natural language prompting. I've had good results but my prompts follow the guide they linked much more closely that what I'd write for V9.
@nathanieljgarcia469 I didn’t mean the prompt accuracy. I’m talking about the overall image quality. V9’s lightning version delivers way better quality images than version X Hyper. Also very noticeable in the images posted here, if you compare them.
which one of your sdxl models have the most wrights from rundiffusion ?
I can't get Loras to work on the Hyper version, it worked on the Lightning version, am i doing something wrong? I also have massively moreunderwhelming results on Hyper than on Lightning. I have followed the prompting guides, but for now, I can't get anything as good as on the previous version (non-X). Is anyone in a similar situation?
where do i get the hires models from? The default ones dont got the recommended model
Question - Is hyper less refined than lightning? I ask because you say "for those with lower PC hardware".
My hardware is good. Am i better with Lightning ?
No, Lightning is more or less the same as Hyper. The difference in quality stems from the fact that the Lightning model is V9 and Hyper is V10 (or X), which was trained from scratch. Use the regular checkpoint of either version (without Hyper or Lightning in the name).
@Weltleere Thanks for the answer, but I cant understand it. They are the same, and "difference of quality" ? So what ones higher quality?
Also are you suggesting use the regular version for best quality?
@RabbitRed Yes, regular checkpoints have the highest quality.
Top tier
Very good work
for some reason v9 is still best and far better result than 10 and 11 hyper
Can you elaborate on the reason why 9 is better, thank you very much!
@dj9wqurf has better prompt adherence and softer more realistic look, but with sharp, vivid, deep colors, will bring it to a hyper look, but the prompt adherence makes big win ie red jumper, yellow jacket will have more images of that, hyper will often pick one color and getting the requested is rare. im fussy love prompt adherence and control.
Love it. Giving me the most satisfying results so far.
Absolutely amazing results
Amazing results, any tips or default training parameters on how to fine-tune it on a specific dataset ?
i was trying to run Juggernaut 7 and i cant find it just 9 and X.. do you know if it will be added back for cvitai generator support?
did a large sample test of 7, 8 & 9 with old, new and crafted prompts and found 7 and 9 have very similar qualities and outputs, 9 is often more refined for realistic photorealism. 7 is more flexible for other styles.
8 will every now and then throw a curve ball and do something different, it wants to wander to nsfw more often than 7 or 9, if they weren't so big I would keep 8 just for those times you just want to try a different variant.
9 is more strict on SDXL sizes and will deliver expected result if you size accordingly. 9 also needs to have the lower cfg which I found combined with controlnet needed to go even lower to get good results. Still it had really good prompt adherence, all 3 did and would often succeed with color placement when others failed.
Again 7 is more flex and will just work on more sizes. I would say for fast experiments 7 and below might be best.
are you gonna make Juggernaut 3 now
The short answer is: We are working on it.
The long answer: It will take some time. Simply using the Juggernaut Dataset will not work satisfactorily. The SD3 Base has different issues compared to the SDXL Base, so I need to change a few things in the dataset. This will happen over the next few weeks (after all, I need to extensively test SD3 myself so I know what I need to work on). In the meantime, RD is taking care of the licensing stuff; there are still some uncertainties here. Overall, we hope that a first version can be released in 4-6 weeks.
@KandooAI Glad to hear that. My own test on SD3 comes out really bad. Don't know if it's about the settings or model itself...
@KandooAI That's great to hear! We're all waiting on the finetunes to fix the human anatomy issues as a result of over-censorship.
hey i heard that if you gonna make juggernaut 3 through fine tuning the base model you still can't use it for commercial purpose is that true?
This is a good Video regarding the License:
https://www.youtube.com/watch?v=r8oA42saTNI
Just one detail about the Jugg_X_RunDiffusion_Hyper version. In ComfyUi, for realism, it works better using euler_ancestral + exponential. Perfect model for realism! Thank you for your hard work on this model!
Is Hyper faster with lower quality, or just faster?
Can you create an XL model with a similar visual effect / VAE to the one in "Juggernaut Final Baked Vae" (SD1.5) ?
You say version X was trained on GPT-4 captioned images but just curious how exactly you did that is there a guide you could point me to because I would like to use GPT-4o to caption my own images for training
Captioned it with this Tool:
https://github.com/jiayev/GPT4V-Image-Captioner (From the HelloWorld Creator)
brand spanking new to this whole AI thing, so please for the love of god be patient with my dumb ass. is it supposed to take 3 hours to make 25% progress on a 512x512 image using the exact recommended settings on the page? I have a 3060ti & a Ryzen 5 5600x
I have the same card as you. It takes 17sec for a gen on 1024 1024
absolutely not, what ui are you using and how many seconds each iteration takes?
you should install (or dual boot) linux, 60% of all your dependency problems will disappear. Then there's the pain of learning linux. :-)
What's the error message?
@philion easydiffusion, and I haven't been able to time it because it's never finished anything. I gave up on it after it stopped at 38% after 5 hours
@RodentOfUnusualSize have you tried a different ui? fooocus also uses juggernaut xl and is pretty capable and easy to use
Fantastic model
Quality of this model is amazing.It would do anything you can throw at him.You can do porn,anime,cartoon,fantasy,SF,Juggernaut would never fail you.Thank you for amazing work!
Is there a way to use this model on SD1.5?
Err no... this model is already SDXL... you just run it on A1111 or ComfyUI
I would like to explain something ...
only TOP Creators have the possibility to use the civitai-generation service for checkpoints ... all others do not ... BUT that does not mean that these models are the best "by far not"
What does this mean absent the names of the supposed rightful heirs to the throne?
Hello, i maybe missing something cause im getting bad image quality if i use same prompt that works with run v7-v8. Where is the Hyper point? anyone can explain me please? thanks
Like all the samples need to keep steps at 6, cfg at 2 or it burns the image. But that makes it quicker.
I'd love to see another version of this, v9 and the lighting give better results than v10 and hyper, each is good still.
but the v9 lightning amazing images look professional love it still.
Version 10 is full of "holes" that lead to extremely strange results, captioning with only detailed captions straight up doesn't work, you need to do detailed captions immediately followed by tags provided by a strong tagger like WD-VIT-V3 or what have you.
Great model mate!
I also wanted to add a shout out to @garand343635 for
Nicola Samori - ArtStyle Lora.
Used it for most of my images.
Works wonders with your Checkpoint.
Amazing!
I am new to AI image generation and appreciate your efforts in creating the Juggernaut checkpoint. I would like to ask if I have permission to sell the images generated from this model.
Sure no problem :)
@KandooAI Thank u so much
@KandooAI 你就是我的上帝。
Model works incredible! Although I am trying the recommended upscaler settings in ComfyUI with Ultimate SD Upscaler and results have a lot of noise/artifacts. Any other recommendations?
I usually work with 4xNMKD-Siax_200k as Upscaler. So i would recommend trying this Upscaler :)
@KandooAI I tried the recommended 4xNMKD-Siax_200k as upscaler model in ComfyUI with the Ultimate SD Upscaler node. But it has a lot more options than in A1111. How are your results in A1111. Are you using any special extension?
Amazing!
V9+ RunDiffusionPhoto2 it's in my top 5
Did anyone else experience elongated bodies or body parts when generating images? I trained a LoRA for this checkpoint and while the images are fairly ok from the composition PoV, the bodies are dis-proportionally elongated (mainly torso but the face and neck, and arms/legs too).
I also saw some long-ish faces on a LoRa I trained, but not only as long as actual long faces, nothing grotesque 🤷♂️.
Yes, it seems to like disproportionately long or short torsos and legs
Awesome model, just works. Though I do warn you, don't use the LMS sampler. At 25 steps on LMS it breaks at step 23 and then saves a overcooked glitchy mess of an image.... though you might be after glitch art soooo uhh...
LMS sampler NO BUENO if you want actually usable realism, tho it is kinda cool for forked up glitch art.
Keep up the work Kandoo, JuggXL is like my go to base model for the majority of my art.
text2image，only works well when Sampling method is ‘restart’
I try to give it a simple prompt, and yet it gives strikingly similar images, ignores negative prompting, and the subject has the same face every time. I'm not sure if it's a me issue, or something with the model.
This is what I'm using for the positive: "realistic, 8k, cinematic, long dark brown hair, young woman, large breasts, long white transparent gown, red lace panties, red lace bra, high-heel brown sandals, large kitchen from the medieval era"
Doesn't seem to have a grasp of what "medieval era" is, so I guess it just doesn't have that.
This has happened to me before and i nearly slapped myself when i realized I was using a seed. Set seed to -1 for random. (for A1111 or forge)
I am not sure why, but it keeps generating nude models. I have NUDITY, NUDE, TOPLESS, BOTTOMLESS in my negative prompt, and no luck! :(
Same, even increasing weight doesn't seem to help.
NSFW in negative, SFW in positive
also try naked in negatives, that helped me a lot,
so neg should have "nude, nudity, naked, unclothed, NSFW, nsfw"
ans dpos should be "Clothed, SFW, sfw" <- both need to be there for some reason, dont know why but it works... -.-
Ok. Thanks for the tips. ^_^
Sometimes adding a delay to negatives helps. For auto/forge, just do something like {(negatives):0.3] which will make them ignored for the first 30%. I'm not totally certain on this, but my understanding is that negatives add a concept at very low weight, rather than actually subtracting it, and early steps are more chaotic than later steps. Could just be a placebo, though.
Looks like version XI is going to be permanently paywalled.
Although the results are more often than not, quite realistic, sadly I find all the faces to be consistently sharing similar traits and failing at accepting specific face description to palliate this issue.
All women faces have same attributes, oval/rounded face, soft cheekbones and jaw, making them quite lookalike. Attempting to include "sharp cheekbones", "High Cheekbones", "Sharp jaw", "thin face", "thin lips" etc... constantly fails.
Mate You need some type of lora or IP adapter
i like the quality when you face swap, it shows really more realistic then the standard base model juggernautXL in fooocus, but i see a lot of crazy hands and fingers and eyes that are not so good
yes i want to know how we can fix the deformed hands, body parts, is there a fix for this ? or use a combination with other loras, or something its kinda dissapointing when i see all those deformed parts? anyone can give some tips or support thanks
did you merge v10 with dpo? images seemed way too saturated and loras dont' work too well with it. works well on v9 and below
Update:
Hey everyone,
It's been a while since the last update. Some of you may have already noticed that Juggernaut XI has been available on Nightcafe & RunDiffusion for a few weeks now, and recently on Octo as well. With the help of a lot of feedback, work on Version XII has already begun.
We're now in the final stretch, and the release is approaching.
This brings me to the main point of this message: With the release of Version XII as an API, Version XI will be available on Civit (including for download). XI will be available on CivitAI sometime within the next 30 days. :)
The guide will be updated again at the time of release and should inform you of all the important details. However, it should be noted in advance that it will once again become significantly easier for you to get images of the same quality as Version 9 and the Prompt Adherence from Version X :).
do you have any plans to create fine tunes of any other models such as Flux or SD3?
@pigeon_ During the development of XI and XII, we have also considered other models. These include SD3, Lumina, PixArt, and Flux. The latter only appeared a few days ago, so there isn't much to say about it yet, except that fine-tuning Flux will likely be much more expensive than the other mentioned variants.
With SAI SD3, I am still holding back a bit. SAI has already announced that they want to release a kind of SD3.1. Personally, it gives me those SDXL 0.9-SDXL1.0 vibes, just with a larger time span in between. Therefore, I am currently using SD3 more for testing, similar to what I did with SDXL 0.9. It is more likely that we will release something after SD3.1 appears.
We are currently once again in the recaptioning process for training. All of the above-mentioned models work significantly better in "natural spoken language," and we want to adjust our dataset accordingly. We have already started our first attempts with all models except Flux (which will happen in the next few days) to see how they react to training.
We will probably do the first major run with the adjusted dataset later this month. Currently, I think we will do this with the Flux model, but I can't guarantee that. Exciting times ahead :)
Take your time, you release great models, Juggernaut 6 to me is still one of your best models it's still perfect followed by Juggernaut 9
Hey, any update? 🙏
@binauralhealing100139 Won´t take much longer, its coming this Month. Maybe in the next 7 Days :)
@KandooAI Really nice, looking forward to it 🙏
Hi :)
Great checkpoint for photorealism, but can I make a request for future updates?
Can you please add some framing info in the data? "Full body" usually gives me a portrait, and "cowboy shot" ALWAYS dresses my subject as a cowboy (negative prompts for cowboy hat and cowboy outfit have no effect)
Thanks in advance :)
which one should i download for comfyui sdxl
Any of the versions works, recomended the last version, you can choose the normal last version if you have a good pc, if not choose the fast one
@derpmagician I guess i will download the latest version, And shall i place the file into "Checkpoints"?
when i use V9+ RunDiffusionPhoto2
it gives me the best face quality
when i use the nsfw version X its bad quality faces resolution but good quality body parts... i wish there would be a nsfw version of V9 rundiffusion photo2
what does that even mean_???
@Ayaz89 its already fixed juggernaut X uses different ways of prompting so i was getting different results
@KandooAI
The version notation is complicated and I don't understand it. so could you please explain it to me?
[Question 1]
Does the "+RunDiffusion" in the version notation mean that RunDiffusion has been merged into Juggernaut?
[Question 2]
The latest version [Juggernaut_X_RunDiffusion] does not have the "+" notation, so is it not a merge?
[Question 3]
Are RunDiffusion and RunDiffusionPhoto 2 different models?
[Question 4]
Also, why is the latest version's X RunDiffusion and not RunDiffusionPhoto 2?
Version 6-8 have RD Photo 1 merged in it, Version 9 has RD Photo 2 merged in it. Version X (and the upcoming XI) have no RD Photo 1/2 in it.
Version XI will run under "Juggernaut XI by RunDiffusion"
@KandooAI
Thank you for the concise and clear answer. I understand it now. I think it would be ideal if this concise explanation was at the top of the model page.
Why is RunDiffusion in the name even though it has not been merged into Version X? I think most people would mistakenly think it was merged into X as well.
@fantaseed I will fix this with Version XI and make it more clear :)
RunDiffusion´s Name is still in X Filename cause they are my "Sponsor". They didnt ask for it, but i wanted to do it anyway as a "Thank You" to them
@KandooAI
I understand, thanks for letting me know.
Hey,
great model:)
In the beginner's guide there is one german sentence which you might want to change to English:
"Ein paar Keywörter die ich noch empfehlen kann:"
"A few additional key expressions which I can recommend:"
Cheers
Yeah, i never fixed it :D With XI there will come an updated Version of the Guide, so it will be fixed by then :)
cool : -)
Great model, if only it could generate proper male genitalia (without using external LORAs or having to mention them) it would be perfect for realistic and classic art generation...
Will they release a version of FLUX juggernaut?
I would love to see a JuggerFlux as well!
I'm only interested in Juggernaut Flux. All other models seem to exert extreme bias in one direction or another at the cost of Model depth.
Hi!
Could you add the recommended params for the lightning version please? Or is it the same as the hyper version (so please mention it in the description).
Great job BTW! Thx for all these models ! :)
It's in the "About this version" under the model details and files on the right. That box adjusts depending on what model you have selected on the top.
@uncledude lol wow I never even knew that I would always look at the generated images to see what was used in terms of scheduler and steps and all that. Thanks!
Any news on new version?👀
This Week :)
Details
Files
juggernautXL_juggXRundiffusion.safetensors
Mirrors
Juggernaut_X_RunDiffusion_Hyper.safetensors
juggernautXL_juggXRundiffusion.safetensors
juggernautXL_juggXRundiffusion.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
houssam hyper-XL.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
juggernautXL_juggXRundiffusion.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
JuggernautXRundiffusion_Hyper.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
jxlmod.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
jg_hy_LI.safetensors
juggernautXL_juggXRundiffusion.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
juggernautXL_juggXRundiffusion.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
REAL_XL_Juggernaut_Rundiffusion_Hyper.safetensors
juggernautXL_juggXRundiffusion.safetensors
juggernautXL_juggXRundiffusion.safetensors
juggernautXL_juggXRundiffusionHyper.safetensors
Juggernaut_X_RunDiffusion_Hyper.safetensors
JuggernautXRundiffusion_Hyper.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.