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:
Res: 832*1216
Sampler: DPM++ 2M Karras
Steps: 30-40
CFG: 3-7 (less is a bit more realistic)
Negative: Start with no negative, and add afterwards the Stuff you don´t wanna see in that image. I don´t recommend using my Negative Prompt, i simply use it because i am lazy :D
VAE is already Baked In
HiRes: 4xNMKD-Siax_200k with 15 Steps and 0.3 Denoise + 1.5 Upscale
FAQ
Comments (141)
X Hype train, let's go
i suspect that JuggernautX might become the king of Civit for the next few months.
plz god be nsfw. the worst thing with juggernautxl was the censorship.. thank youuuu
It's not a Pony variant, it can't do multi-person NSFW properly at all if that's what you mean. It could always draw nude neople like most photoreal models though
No version of Juggernaut has ever been censored
@RunDiffusion Wow thank you for your reply!! And yes you are right! I do understand that you can do nudity in juggernaut, but I was thinking more about the intense stuff. So yeah, juggernaut is uncensored in the sense that it can do nudity and a lot more, and you have always been able to do it. But I want quality "home made porn". At this point I barely even watch pornhub, I watch my own home-generated content instead. But you are right, to a certain degree, juggernaut is not censored.
It's an interesting problem, what is censorship? What is NSFW? The female shape? A silhuette of female body? Analsex? Swastikas? Obviously there is a spectrum of extremity here. I would love to read a blog post on how juggernaut trainers think about this topic!
Obviously I am a big fan of your work and juggernaut has always been ahead of the curve. Thanks and keep them coming!!
I enjoy using this model!
Thank you for trying it! Kandoo put in a lot of work training and I spent many hours testing. So when we get these positive comments it feels great. :)
I can already tell version X is going to be legendary, let's gooo
outstanding as usual
Agreed! Kandoo always makes the magic happen.
Also, Can't wait for the lightning model of version X, I'm working on 8 gigs of vram and it's a 3050, so basically a recipe for disaster xd
your lightening 9th model is a life saver really, for real excited for the turbo version of this one.
Lightning Version for Version X will come next weekend :)
@KandooAI Keep up the good work, I will be eagerly waiting!
3060 with 6GB here. I feel your pain, haha
For some reason when using version 10, I am not getting as good of a results as version 9. I am using Fooocus. Switching back to version 9 :(
Are you following the prompting guide. Create a complete prompt you have much greater control with X. So prompt your background, foreground and little details. It'll shine! :)
We've tested this model extensively and while a few baseline generations might be better with 9, X outperforms the other models. Plus it is a new foundation that is MUCH better as we continue to build on top of it.
@RunDiffusion what settings do you recommend for fooocus? I usually dont get good results with the recommended settings that are given
@Colorblind_Adam I do use detailed prompts
@RunDiffusionMaybe because I use fooocus and it doesn't give as much control as other applications
@deluxelex134 Can I please see an example prompt of what you are doing? I am one of the primary testers for this model so I have probably a 100 hrs testing this and I tested Jug 9 as well before launch. This gives far greater control to prompt engineers. But if you enjoy Jug 9 more nobody is saying stop using it. I just am convinced once you get the hang of this one you'll love it.
@Colorblind_Adam Sure, here's a meta data example. Results in version 9 are better than 10, both detail wise and aesthetics:
{
"prompt": "nude beautiful woman, breasts, full body, dynamic studio lighting, Art Deco room",
"negative_prompt": "(worst quality:1.2). ugly hands. bad hands. bad fingers. bad anatomy. wrong anatomy. malformed. deformed. mutated. overexposed. underexposed. blown highlights. blurry. Asian face. extra arms. 3 arms. three arms. extra limbs",
"prompt_expansion": "",
"styles": "[]",
"performance": "Speed",
"steps": 25,
"resolution": "(1344, 1344)",
"guidance_scale": 5.5,
"sharpness": 4,
"adm_guidance": "(1.5, 0.8, 0.3)",
"base_model": "juggernautXL_v9Rundiffusionphoto2.safetensors",
"refiner_model": "None",
"refiner_switch": 0.5,
"adaptive_cfg": 6,
"sampler": "dpmpp_3m_sde_gpu",
"scheduler": "exponential",
"seed": "7956872909401832839",
"metadata_scheme": "fooocus",
"version": "Fooocus v2.3.1"
}
@deluxelex134 I would suggest creating a full prompt. This model is very precise and the more you add the better image you'll get.
Full body shot of a ((nude)) beautiful woman, ((breasts)) dynamic studio lighting highlighting sharp geometric patterns of an Art Deco room, dominant colors gold and black, style vintage glamour photography, mood sophisticated, lighting high contrast, perspective close-up, texture glossy surfaces and intricate fabric details.
@deluxelex134 I posted a sample with settings below in the samples :) Hope that helps you.
@Colorblind_Adam Thanks for the prompting suggestion, I'll try it out.
Just had a few questions, and thanks for releasing this!
How do you feel about captioning with the new internlm-xcomposer2-vl-7b-4bit and CogVLM v7?
I did a 50/50 split locally to train 120,000 images and it was completely uncensored captions. (Using 4090) I just had to use some traditional Chinese characters at the end of the english context prompt and it became uncensored somehow. But it was extremely accurate in the captioning, these were very nsfw images with multiple people and pretty much has the "boring realism" aesthetic. I am getting best results I've ever had this way.
I'm just wondering why you guys decided on the captioning model you chose to use? Because if it's that good I may try it out.
Btw, not sure if you guys trained with Onetrainer but would recommend converting the diffusers in comfyui instead if you did, there is an issue when using their .safetensor outputs https://github.com/Nerogar/OneTrainer/issues/224
PS. I will be releasing this nsfw model in the future. It's still training and will probably take quite a long time to complete, I resume it every night with Onetrainer, but already after 1 epoch it's looking great.
Thanks for releasing a slightly more nsfw version of jugg.
I tried CogVLM v7 and it was good, but GPT4 (for SFW) was on a whole different Level. The Difference was so big that i decided to pay for the GPT4 Vision Access and i feel like it was worth it :)
Didnt try the other one tho...
I used OneTrainer in the past for my old SD 1.5 Checkpoints. Since SDXL came out i used a Custom Kohya_ss Notebook
Awesome work once again - thank you!
Wonderfully versatile and still at the top of my list as default model...
Soon to start all over again with sd3? Either way - this one ain't going anywhere! 💥👌
SD3 is naturally in focus, and we're all eagerly awaiting the weights to finally get started with training ;) But the journey from Juggernaut on SDXL is not over yet; there's still plenty of room for improvements, and I'd like to make use of that.
at the moment it is raw and unfinished, but the potential is great, some things have become better!! but something is doubtful, there are problems with landscapes, blur, clarity...in general I like the work done!
Version X is ultimately just the rework of the base set from V1. After that, several versions have been released, and I will gradually add those datasets with GPT4 Captions to the base. So, it's essentially like a fresh start, which undoubtedly comes with its fair share of issues.
I recommend using the Prompting Guide. The two images in the guide should be a great help in navigating certain areas better (both in the prompt and in the negative prompt). For example, it helps to include classic blur effects in the negative prompts (Blur, DOF, etc.) and to include desired effects towards the beginning of the prompt, such as "Sharp focus, high-resolution image, high resolution," etc.
great work,look forward the lightning version
I never get good images with the recommended settings. I need to use default settings in advanced tab or dpmpp_3m_sde for good images. Are the recommended settings for fooocus? The model itself is insane 👍
Thank you for mentioning our work in the document. I'm glad that the GPT4V tagger could help with the improvements to Juggernaut. Look forward to more open-source models using multimodal models to enhance the accuracy of training image descriptions!😃
I thank you for the development of your captioning tool. I've been searching for a good all-in-one solution for ages and couldn't be happier to have found your tool :) And then getting appropriate graphics that represent the most common keywords/tokens in the dataset was like the cherry on top for me :) In fact, I've been recommending your tool to everyone I talk to about it for months :D
Leosam we appreciate your hardwork with the GPT4V tagger. And Kandoo really put it to good use! :)
I don't know how you did it, but you managed to get CLIP G ((weights)) to work. Every other SDXL checkpoint has very little effect whenever I try to (((push))) a CLIP G detail, only CLIP L (older, less dynamic, SD 1.5 style), but yours does. Thank you so much for this great model. The NSFW in CLIP G is great too. Creating nudes has been a struggle in XL because of the blank anatomy. Please continue to include this trigger in your updates.
Surprised that the models follows the clothing colors in details, text is not always accurate, it gets a bit hard though, as you have to detail every aspect.
I think the goal here is to really give artists complete autonomy. But yes you do have to be very specific with what you want. Talk about the object, textures lighting, colors, etc.
Love your work but I think this should not be called X (ten). More Alpha or beta test. Love your work and will follow your success further.
1st of all thanks for version X. I want to know from below which one is official Juggernaut inpainting model. If any of those not official inpainting model then which one you are recommend to use or any others SDXL inpainting model you recommend. Please suggest best SDXL inpainting model🙏🙏🙏.
model 1:
https://civitai.com/models/245423/juggerxlinpaint
model 2:
https://civitai.com/models/403361/juggernaut-xl-inpainting
The First one is for V7 and 8. The Second one for V9 and the current version. All of them are fine :)
@KandooAI can you recommend any others SDXL inpainting
If you use Fooocus for inpainting/outpainting you can use any SDXL model. No inpainting version needed because Fooocus uses its own DPMPP algorithm.
What can I say you always Blow Me Away with your beautiful work I'm so happy all my Laura's work with yours
Is there a way to get pruned model maybe till 3-4 Gb's anytime soon?
It's already fp16
do i need to use a SDXL refiner for this model
No you dont need that
Amazing model, I love version X!
The whole team is happy so many people are enjoying this model.
With such a diverse model how do you accommodate different categories (cars, people, etc)into your training or dataset preparation?
1. Do you separate categories into different folders or do you put all the images into one dataset? And how important is captioning?
2. What's the secret to training hands well?
1.The dataset has indeed been restructured during the rebuild. Here, I was significantly more precise than with my previous datasets. Cars, Architecture, Wildlife, Portrait (Man/Woman), and a bunch of other folders. After that, you can caption folder by folder and then curate. This is significantly more comfortable than having to caption the entire dataset at once and then having to review it all over again.
2.Hands are a good example. Ultimately, a large amount of material is needed for hands because there are simply an unbelievable number of positions/poses for hands... whether it's gestures (middle finger, thumbs up, waving, peace sign, etc.) or completely natural things (hand on face, hand on table/couch, holding objects). GPT-4 V helps well with captioning. Just make sure to use a prompt for GPT beforehand that focuses on the hands and specifically considers them during captioning.
@KandooAI Quality response right there! I have up to 31 different categories (outdoor, city, people, etc), all separated into individual folders. Here's the issue, every time I train them (with dreambooth) all at once I come across a huge problem of having a set of categories over-trained, others under-trained and others well trained. However, when I train them individually with the same settings that normally works. Any insights regarding this?
Regarding hands, would you suggest allocating a folder specifically for hands? Like a folder with only close-ups of hands in different poses. Or it's better to just combine the hands images with images of people into one folder than separating them. Am bringing this to your attention because I managed to collect 300 images of hands only. I then separated them into two folders. One has hands holding something and the other has hands in different gestures. But that approach hasn't proven to work
@sphiwe
There's a difference between Dreambooth and "Full" finetuning. Dreambooth can handle multiple concepts, but there's a higher risk of overtraining certain areas more quickly. Finetuning seems to manage this better, at least it doesn't overtrain as fast as Dreambooth. However, it makes sense to prepare the dataset accordingly. Some areas require less material than others—a good example would be close-up portraits versus full-body shots. You'll need fewer close-up portraits in your dataset compared to full-body shots. I'd like to help more, but in the end, it often comes down to trial and error. One tip: make sure you create regular checkpoints during training and check them, so you can counteract overtraining if necessary.
Regarding hands, it's hard to say which approach is best. We all struggle with hands, whether small creators like myself or big companies like MidJourney, DALL-E, etc. I only started to really tackle hands with versions 7-8, and I structured it similarly to you. However, during the work on version X, I heard about another interesting possibility that I'd like to try in future versions. The basic idea is to first train hands as we've done before. Then create a Bad Hand LoRA with the model and add it with a negative value (-0.3, for example) to remove bad hands from the model and give more focus to the well-trained hands. I've only tested this on a small scale so far and noticed an improvement. I'll look into it more after the release of the Lightning Model, so I can't say much more about it yet.
@KandooAI I need to give this full fine-tuning approach a try as I have been utilizing Dreambooth only. I dropped you a discord message. There's amazing work I'd like to show you
What are the secret training settings?
If I train base sdxl, everything works fine. If, with the same settings, I try to train Juggernaut, the entire thing collapses within seconds.
Are you hiding whatever settings you used for the model training on purpose so no-one can train a lora on this? Like I get technically it works to train on base SDXL and use the lora here, but it'd work way better if we could train with the correct settings on the model directly.
I'd appreciate if any of the exotic training settings that deviate from SDXL1.0 were added to the description.
From what I understand. Training on SDXL base will work with Juggernaut X :)
@Colorblind_Adam What? He already said EXACTLY that in the OP: "Like I get technically it works to train on base SDXL and use the lora here, but it'd work way better if we could train with the correct settings on the model directly."
There is no big Secret, only real Difference is the VAE i used in Training (FP16-Fix) and Multires Noise (Iteration: 6) (Discount: 0.6) Everything else is pretty much normal
Training against Juggernaut-X works fine using kohya_ss. If it is not working for you I suspect it has nothing to do with this model.
Previous versions are checkpoint merge of Rundiffusion model. Just wonder is the X version a single trained checkpoint or a merged one?
I noticed you removed the description of the tecommended settings for the lighting model.
Would it be possible to upload them somewhere? (In this comment section is perfect)
When u click on the Lightning Model you can find the recommended Settings under "About this Version"
Sampler: DPM++ SDE
Steps: 4-6
CFG: 1-2
HiRes:
Sampler: Siax_200k
Steps: 2
Denoise: 0.35
Upscale: 1.5
Whenever I am creating a picture without a negative prompt, this skin looks pretty realistic. But when I just add one simple negative prompt the skin looks like an instagram filter. does someone has a solution for this?
You may have to turn off the styles in fooocus for jugg10 to follow the prompt properly, if the prompt is complicated/detailed.
suddenly I cant generate anything.. im getting this error after downloading this tensor file and trying to use it in auto1111 - AttributeError: 'UiControlNetUnit' object has no attribute 'enable'
Looks like the token limit of 75 is literally killing the use of this version for me. Sadly.
What are you talking about? There is no difference between Juggernaut and any other SDXL model when it comes to the token limit. You can obviously go beyond 75, but you get another chunk of tokens with weights reset. So token 76 has a higher weight than token 75 because the 76th is the first token in the next batch. You should be aware of that when you are writing prompts but there is absolutely no difference between this checkpoint and any other when it comes to this.
just for my understanding: what are tokens? the count of words?
@binauralhealing100139
It's not quite that simple. Words can consist of multiple tokens. Everything you write in the prompt, including punctuation, is part of this.
For example, when you enter a text prompt into Stable Diffusion, each punctuation mark might be recognized and processed individually. This helps the model to understand sentence boundaries and the organization of ideas within the prompt, which can influence the generated images.
In Auto1111 (which I mostly use), the number of tokens is displayed right next to the prompt, which is quite helpful.
@KandooAI okay thanks
@joeuser12 You may want to read the models actual description.
Ihr habt im Prompting guide noch was vergessen zu übersetzen 😉😂
Ja ist mir bereits aufgefallen :D Werde das noch im Laufe der Woche ändern :)
lass es drin. "keywörter" statt "schlüsselwörter" liest man sonst nirgends xD
würde es auch drin lassen 😂
Hyper or Lightning ? Both is possible with Juggernaut, so what do you guys prefer ?
what is hyper?? 😮
A lightning version would be nice!
@KandooAI damn! with more steps like max 8 the hyper would be give more accurate and better results 😃
Hyper seems to have less artifacts and better prompt adherence. There is no reason to stick to antique months old methods.
@Weltleere is that like a new model? like different to sdxl? im kinda new tho
@kd231002573 No, it's a LoRA that creators will bake into their models to make them faster. With models that have Turbo, Lightning or Hyper baked in, you can use only 1 to 8 steps and still generate good images. Hyper is the newest and can be seen as an improvement, I guess. But they are all worse than normal models, because they trade quality for speed.
Hyper looks neat. Go with that one.
@KandooAI Thank you for version X great model overall, but like another comment said there is a lot of unprompted nudity. (For example specifying a guy should have blond hair often causes them to be nude).
Could you also please release the weights for the SFW version?
You can avoid 90 % of Nudity when u put the Triggerwords in the negative.
You can find those Words in the Prompt Guide:
Nude
Naked
Vagina
Pussy
Breasts
Penis
Testicles
@KandooAI Thanks for the response, but it doesn't fully solve the problem. There seem to be certain keywords that cause nudity even with those triggerwords in the negative (e.g. some ethnicities are associated with nudity). I can give you some example prompts if it would help debug.
@testuser1111 yeah a prompt would absolutely help :)
I struggle a lil bit with the X Version, always get too saturated / overexposure / too sharp pictures. I didnt have this Problem with the same prompt with the other Version. Still love this Model! :)
With the rework of Juggernaut, the behavior during prompting has changed slightly. Prompts that worked well in Version 9 or in other models might not necessarily work in the new version - and vice versa. I have created a small prompting guide for Version X. It also includes 2 images that should be helpful for prompting. Among other things, it contains 200-250 quality tokens/keywords that work excellently in Version X.
https://storage.googleapis.com/run-diffusion-public-assets/Prompting_Juggernaut_X.pdf
use something like DPM++ SDE, not 2m or 3m version, also lower CFG scale
It is definitely different that v9, but once you adapt your prompt style and find your settings it works really really well. A little heavy on the NSFW so make sure you add both positive and negative prompts to stop it if you don't want it. But it is also nice to have the ability to use easily if you want to.
This is exactly my experience with it when I started testing it. Well said.
This is one of the best checkpoints on this website!
What settings are you guys using in fooocus for best results? 🤔
Hey, im new to this :) Can anyone tell me which size i have to set when using inpaint?
X is too horny 😂 creates nsfw even with neutral prompts
You can avoid NSFW 90% of the times when you put the Triggerwords for it in the negative prompt (You can find it in the Prompting Guide)
Prompt for Clothes + Nude/Naked in negative and you should be good to go. :)
Version X should rather be called XXX. You get vaginas even when putting "nude" and "vagina" in the negatives. Results from V9 turn out better across the board especially with male subjects. You should probably put more RunDiffusion Photo and less NSFW into your mix.
Really? If you actively try to make porn with this model you will get nightmarish monstrosities 100% of the time, it actually knows very little about NSFW in that sense as far as I can tell
This model is not really suitable for XXX work we didn't train it on porn. If you want to avoid nudity remember to prompt for clothes and also put Nude/Naked in the negatives. If you still get an errant boob try adding breast as well. HOpe that helps you
awesome model
all of my results look terrible. any tips?
Well it would be helpful if you show some examples with a prompt. Otherwise it will be hard to help you ;)
You can try the new guide we put out today.
https://learn.rundiffusion.com/prompting-guide-for-juggernaut-x/
Don't use a LoRA.
Just curious as im new, im using 2060 nvidia, why does the render take 30-40mins? Am i doing something wrong
Hey bro, i had the same cg before, it's just too weak. I upgraded to 16go 4060 TI (500e), génération now takes 30 secondes. If you dont have the money try foocus, it's working with 2060 but résult are less good than A1111 or comfy ui
Render times for SD1.5 models would probably be fast, but with SDXL models it would be too slow as they're quite large and would exceed your GPU's VRAM
your vram is too small
There is definitely something wrong.
Even a GTX 1050 TI (4GB) only needs about 3 or 4 minutes, for SDXL, on Comfy UI.
And an RTX 2070 (mobile), needs less than 30 seconds.
The RTX 4060 TI (16GB), only needs about 12 seconds, on Comfy UI.
And the RTX 4090, only needs 3 seconds !
Here are some benchmarks:
https://github.com/comfyanonymous/ComfyUI/discussions/2970
If you're using Comfy UI, maybe you are generating in "CPU mode".
You need to open Comfy UI with the [ run_nvidia_gpu.bat ] file instead for [ run_cpu.bat ]
@jr81 thanks guys, so it should not take so long. I have 16g Ram. I used another model, realisticV60 and i takes a few seconds, realvis40 takes much longer as well...
I am new so I have not tried Comfy UI, nor have i checked if i am using GPU or CPU mode...
I'm just using Stable Difussion
@br00xxx Oh dear okay so i see where i am going wrong.. this XL would mean it is more heavy on the ram usage. I should stick to the normal SD Models
The RTX 2060 only has 6GB VRAM.
https://www.techpowerup.com/gpu-specs/geforce-rtx-2060.c3310
It's VRAM that matters for image generation, not RAM.
Realistic Vision V6.0 B1 is SD1.5 ( https://civitai.com/models/4201/realistic-vision-v60-b1 )
RealVisXL V4.0 is SDXL ( https://civitai.com/models/139562/realvisxl-v40 )
You can use the SDXL Lightning versions, with only 4 to 6 steps, so it's faster.
But all SDXL models are larger than 6GB.
With lower VRAM, Comfy UI should be faster than A1111, if you want to use SDXL.
https://github.com/comfyanonymous/ComfyUI
It should only take a couple of minutes, at most.
@jr81 that's the information i needed... thank you, I will have to take a look at Comfy soon
hi bro, try to remove --no-half command line in you webui-user.bat. I had a same problem and now I can generate 1024*1024 pics within 10 second with hyber version, with my 4060ti
I have been using v6, v7 and v9 EXTENSIVELY with amazing results. Do I have to change anything to get v10 to work at least on the same level?
Version X has been newly captioned. Prompts that you used in V6-9 might not necessarily work in Version X. The guide is a small aid and provides a lot of tokens that you can use to improve your results. Alternatively, it's worth taking a look at the community gallery in the comments section. Here you will find some really great pictures with their corresponding prompts :)
@KandooAI Yeah... I'm good.
@KandooAI Thank you for your reply and amazing contributions!
Hey sir Thanks for your work.
PROBLEM SOLVED : It was face restoration, code former exactly, i just disabled any face restoration and everything is working like a charm
I'm checking if everything is working in my setup annnndd it's not.
*When i try to reproduce your picture i'm loosing all détails in last step.
I already try : Disable VAE, Change High res fix, command -nohalf, upscaling (with 4x-NMKD) instead of high rez, refiner SDXL or not.
Any advice ? Here is the freckles girls version for me "out of box", i mean original (so your picture) in png info+send to text to img+ generate
IMPORTANT EDIT : I tried to save pics from each step, only the very last one has erased detailed. Before they looks like exactly yours, i guess it's the high res fix ? I still can't resolve the problem. It's litteraly the same upscaler than you
4060 ti oc 16go; stable diffusion 1.9
I checked the Metadata for your Image....It still shows "Face restoration: CodeFormer" and that u used the SDXL Refiner . I personally didnt use any Face Restoration or Refiner, so i would turn them off
lol didnt saw the message that you already solved it :D
@KandooAI Aahha yeah that was it, again thanks for everything. I let this post online, maybe usefull for someone.
Update:
After testing last week, we have decided to release a Hyper version of Version X. This should help those who are working with slightly less powerful PCs :) At the same time, there will be a new guide written by Adam that is significantly more detailed than my own. Some advice for using the Hyper version has also been added there :) So definitely check out the new guide at the release :)
Release will be in the next 12 hours :)
Thank You.
V9 Lightning is working great for me, with IPAdapter FaceID + InstantID ( + 5 image batch ), at 6 steps.
With a GTX 1070 (8GB), I can generate in about 30 seconds, in Comfy UI, with this complex workflow.
There will also be a Hyper version that facilitates negative prompts and larger guidance scales (e.g. 5~10).
If it's different enough, maybe you can also make a version for this.
@jr81 Thanks for the Info about that upcoming Hyper Version with a higher possible cfg...That was new to me :) Can´t wait to try that out
@KandooAI Could you please share how you have made a Hyper version of your model? Was that a simple merge with Hyper lora?
@Quan_Chi Yeah just with the 4 Steps LoRA. Pretty much the same i did with V9 (Lightning)
@KandooAI Thanks!
Hey what settings do you recommend for v9 (lightning merge) on comfyui?
I tried:
4-6 steps
dpmpp_2m_sde, karras
and the results are really bad.
I can get slightly better results with euler, normal but nothing great...
Try the DPM++ SDE Sampler, not the 2m Version :)
You can also try dpmpp_2m ( non SDE ), karras, at 6 steps.
It's working very well for me, in Comfy UI, and It will be faster than DPM++ SDE
Sometimes Lightning and Turbo models don't work well with dpmpp_2m ( and DPM++ SDE is needed ).
But V9 Lightning seems to be the model that gives the best dpmpp_2m results, as far as I know.
@KandooAI Thanks! I realised i forgot the clip skip in my workflow, it's working great now!
Details
Files
juggernautXL_juggXByRundiffusion.safetensors
Mirrors
juggernautXL_juggXByRundiffusion.safetensors
juggernautXL_juggXByRundiffusion.safetensors
REAL_XL_Juggernaut_Rundiffusion.safetensors
Juggernaut_X_RunDiffusion.safetensors
Juggernaut-X-RunDiffusion-NSFW.safetensors
fabric-designs.safetensors
MyBack_SDXL_Juggernaut_XL_VAE_V10(version_X).safetensors
juggernautXL_juggernautX.safetensors
juggernautXL_juggernautX.safetensors
Juggernaut_X_RunDiffusion.safetensors
juggernautXL_juggernautX.safetensors
Juggernaut-X-RunDiffusion-NSFW.safetensors
juggernautXL_juggernautX.safetensors
SDXL1.0_JuggernautXL_10.safetensors
Juggernaut_X_RunDiffusion.safetensors
diffusion_pytorch_model.safetensors
juggernautXL_juggernautX.safetensors
Juggernaut-X-RunDiffusion-NSFW.safetensors
Juggernaut-X-RunDiffusion-NSFW.safetensors
juggernautXL_juggernautX.safetensors
Juggernaut_X_RunDiffusion.safetensors
Juggernaut-X-RunDiffusion-NSFW.safetensors
Juggernaut_X_RunDiffusion.safetensors
juggernautXL_juggernautX.safetensors
Juggernaut_X_RunDiffusion.safetensors
juggernautXL_juggernautX.safetensors
juggernautXL_juggernautX.safetensors
70_juggernautXL_juggernautX.safetensors
J-X-N.safetensors
Juggernaut-X-RunDiffusion-NSFW.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.