NOTE: If using on the generator, I highly recommend ticking: Enhanced Compatibility
as prompt, the following is much better with this model.
ComfyUI workflows; (Save as .json)
Very basic beginner workflow: https://pastebin.com/q2dfZ8d6
Intermediate Level workflow with a 2nd stage img2img upscale: https://pastebin.com/U636LdnF
Complex UltimateSDUpscale workflow: https://pastebin.com/dhWcvAaz
My Realistic Pony model is available on the generator:
https://civarchive.com/models/515034/jib-mix-pony-realistic
And my Realistic Illustrious model
If you like this model, check out my New Photo Realistic Illustrious model on the generator: https://civarchive.com/models/1255024/jib-mix-illustrious-realistic
I will try to keep this model in the onsite generation list.
I try and look at every image you post and award 80-120 buzz to images I really like.
This is a Merge model of many of my favorite realistic SDXL Models (and some more artistic models put in for style and consistency) and I really like the output it makes.
Recommended settings:
Sampler: DPM++ 2M Karrass
Steps: 8-35
CFG: 2-5 (I Often use CFG 3)
with DPM++ SDE you can go as low as 4 steps. with a CFG of 2.2
Version 16 Aphrodite:
More photographic realistic default look.
Increased background details
Better Prompt following
Better tasteful NSFW nudes without hurting detail too much
Version 16 Aphrodite XXX:
Much Better NSFW capabilities at the cost of some loss of detail and art styles.
Version 14:
A modest increase in details and contrast.
Recommend
Version 13:
After the very poor launch of SD3 I decided SXDL still had more to give and bring you this improved merge I have been tweaking for a good while now. FP32 version coming soon!
This version seems to benifit from a lower CFG than previous versions 2-5 CFG seems good. (2.7 seems the sweet spot for portraits)
Version 12:
This version has better brighter colours and contrast plus better prompt following.
Don't forget to also check out my new Pony Model: https://civarchive.com/models/515034/jib-mix-pony-realistic?modelVersionId=598011
This is likely to be the last SDXL model merge/finetune I publish as SD3 will be out to tune Next week!
Version 11:
Bringing you another merge of great models I like and doing 100's of A/B tests to pick the model that gives the best images. This is not a lightning model as I prefer quality over speed.
This is likely be my 2nd to last SDXL model merge/finetune I publish as SD3 will be out to tune soon.
and also tested them quite a lot with the the new Hyper Lora (8 Step).
Version 10: Is merged with the SDXL Lighting Lora. (If you want to do non photographic artworks like Watercolours, you should stick with version 9 I think.)
You need to use a CFG of :CFG: 1-2
and can use steps a low as: Steps: 4-8
and DPM++ SDE Sampler is best.
Version 9: Merged with some good prompt following models, much better male bodies.
Version 8: I have merged with some models newer models, tried to work on skin texture (but it is still not where I want it to be) and increasing the background details in images.
Version 7: I have merged with some models that claim to have improved Prompt Adherence (you still needs some cherry-picking of images, it is no Dalle.3) and some other new models with styles I liked the look of.
Version 6 has better/more detailed backgrounds and higher image contrast as well as some better NSFW capabilities. It also seems to give a wider variety of poses and image variations.
Version 5 has been merged with a few different SDXL eye Improvement Loras and several NSFW Loras. then merged back with older versions of the model and some Anime and Artistic models to bring back the artistic capabilities.
Version 4 incorporates newer released flexible models and seems to improve composition, skin texture, and artistic styles.
Version 3 incorporates hassansdxl_v10 NSFW model at a low 0.15.
And improves NSFW generations without affecting (probably Improving) overall consistency and artistic flare.
Version 2 incorporates
RundiffusionXL_beta
Colossus Project XL
and improves skin texture and overall composition.
Version 1 is a Merge of:
JuggernautXL_version6Rundiffusion
realvisxlV20_v20Bakedvae
wowXLV2_v2
Thinkdiffusionxl_v10
nightvisionXLPhotorealisticPortrait
MOHAWK_v14
It is quite good at NSWF , prompting (boobs out) (naked) will usually be enough, but can be improved with the https://civarchive.com/models/117603/finenude
Lora to add better detail.
Finally, a few recommendations for the settings:
Sampler: I like the Heun Sampler but it is slower than most.
Steps: 30-40
CFG: 6-7 (Often use CFG 7 but 6 can look a bit pleasing/natural for photo-realistic images.)
Including: "35mm photograph, film, bokeh, professional, 4k, highly detailed" gives very good skin details in portraits.
I like to put the token (Bokeh) In the negative prompt to remove some blurring and improve the image.
Most of my sample Images have used the SXDL Styles Selector on the Photographic prompt injection setting.
I like to give all my output images at least a 2x upscale with img2img with high steps and a 0.15 - 0.6 Denoising strength.
Some times I the do a 4K upscale with Ultimate tiled upscale script extension.
Description
Better Skin details and photo realism.
FAQ
Comments (87)
This looks freaking great! Amazing work. blows other models out of the water being super realistic and also very flexible.
Great results J1B! Do you have any images posted with your workflow? I'm getting nothing even approaching the quality you're getting.
I mean images that are png with the embedded workflow.
Considering there is "BREAK" present in his prompts I suppose he is not using ComfyUI. I mean, sure you can use it in ComfyUI too with right nodes, just not that many ppl do.
@poondoggle I am going to post my Comfyui Workflow I use today, I am just trying to clean it up a bit as it has dozens of nodes in I never use.
@Mescalamba That is me just reusing my old prompts from Automatic 1111 days and Regional Prompter I suspect.
@J1B Thank you!
Very basic workflow with optimised settings https://pastebin.com/q2dfZ8d6
Intermediate Level workflow with a 2nd stage img2img upscale: https://pastebin.com/U636LdnF
Much bigger complex workflow to do a 2.5x upscale: https://pastebin.com/dhWcvAaz but it takes around 4-5 mins per image on my 3090.
@J1B You can actually install https://github.com/asagi4/comfyui-prompt-control and use majority of A1111. Basically its possible to get all A1111 inside ComfyUI and some extra, like LORA scheduling.
First of all, thank you for this checkpoint. I've been using it since early versions.
However, the latest v18 seems overcooked (I didn't get a chance to use the one before, so I cannot say whether the issue was present in v17). I could only make it work with Euler samplers and very low CFG. Using a DPM++ 2M or DPM++ 3M at any CFG deep fries the image regardless of CFG setting. It's almost like trying to use a DMD2 checkpoint in regular mode.
Seeing the mention of "with DPM++ SDE you can go as low as 4 steps. with a CFG of 2.2" in the model description kind of suggests that there might be parts of distilled/accelerated models in it, which is something I generally try to avoid as it breaks my workflow. If this is the case... May I ask for a non-accelerated version please? Thank you in advance!
Yes for V18 one out of two pictures are over cooked , I didn't have the problem with version17
Right. I downloaded V17, and there is no such problem there. So it seems V18 has some accelerated model merged in, which basically ruins it.
Strange, you will have to post an image with workflow or settings as I have been unable to recreate it in ComfyUI at least.
All Jib Mix versions have had some form speed loras mixed in since at least version 10.
the skin game is far superior from last version, but the pose and general lighting game suffered a bit, thanks for what you do anyways
@J1B I notice the problem came from I used Euler Simple instead of Euler Automatic.
the result is a lot better now, also too much things in negative prompt was causing some overcooking.
Anyway thanks for you great models !!!
@tta_mementomori Thanks, yes I noticed long prompts in the negative produce more lightness in the image and less darker.
https://www.reddit.com/r/StableDiffusion/comments/1lr3qkm/comment/n1apc6i/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
I have been using a conditioning node at 30-50% for the negative in ComfyUI to get a better balance.
i think it has a dmd2 lora merged into it, so LCM, EUL_A, DPMPP_2s_A should still work fine with cfg 1, and 6-8 steps. if wanting to use the dpmpp_sde variants, even lower cfg is required.
menyan1381 Its a bit different, my V18 model doesn't behave the same as other dmd2 only models like Big Love, for many versions my models have preferred a low CFG around 2.5 but I actually think I prefer this version around CFG 4 for stability.
Outstanding prompt adherence.
Stunning results thanks!
If this mode allow us to easily add text to the images, it would be the best model ever!.
I love it. It's my favorite model and it helps me to create the best images!
Thanks.
I do have a Flux model that specialises in that: https://civitai.com/models/686814?modelVersionId=1355240
But SDXL models are not anywhere as good at text as Flux, I have seen some loras and SDXL models that do improve text some what, so I will look into that.
edit: i've used v17 and 18. its definitely different from v16 which i use quite a long time.
i'm totally impressed, cfg 4 only sometimes (rarely) overbake into artifact (cfg 3-3.5, 12 steps is my preferred value). What made me even more impressed is how deterministic sampler works as well. normally when i use dmd2 lora, i'm stuck with stochastic sampler, and very low cfg 1-1.5.
One of my favourite models. I noticed that v18 adds skin blemishes/freckles too intensely, to the point where i have to change the prompt to specifically generate smoother/skin. i use v17 when i want smoother skin.
outstanding, either way.
tried at lower step? i've got good result with 6-8 steps
Yes I have noticed this, putting "freckles" in the negative will eliminate them though and leave you with smoother skin.
This only generates nsfw content. Is there a way to disable this? I tried negative prompts but that didn't work.
No, it should be fine, are you sure there isn't something in your prompt that is alluding to NSWF? Try describing their clothes ect that might help.
I don't think that it's a problem with your model. Even the base sdxl model generates nsfw content even for seemingly innocent prompts.
I have the same problem 98% of generation is NSFW, i tried to use negative but dosen't work. BTW it's a great checkpoint. Amazing quality
@fedezupi500 If you could provide me with the prompt please? And I can investigate.
Positive prompt:
(upper body portrait, (photography of a 20 year old babe:1.2), huge breast, wearing a yellow red short dress, belt, boots, crochet), beautiful, (no teeth visible:1.2), perfect body, highly detailed skin, professional photography, (((photo realistic,ultra realistic,photorealistic,8k uhd)), photo-realistic, highly detailed, professional lighting, sharp focus, intricate details, realistic, gorgeous, bokah, jewellery, diamond, intricate detail, delicate pattern, necklace, earrings, bracelet, armlet,looking at camera,(classroom, teacher,full body, sitting near table)
Negative:
lowres, bad anatomy, deformed, mutated, asymmetrical face, ugly, blurry, text, watermark, logo, oversmoothed skin, plastic skin, cartoon, drawing, extra limbs, duplicate face, distorted eyes, unnatural teeth, multiple pupils, closed eyes, noise, artifacts, nsfw, nipples, breasts, naked, nude breast, breast exposed, areola, areola slip, wet, sperm, cum
This is one example
I generate most of my work with J1bs model, I have no problem making sfw images
This is the best realistic model I've found. It is incredibly good and fast. It is superb with the texture of the fabrics and materials. My only nitpick is that it struggles to generate people holding swords. I am trying to make a lora to fix that.
When i try to use this for img2img i get
NansException: A tensor with NaNs was produced in Unet. This could be either because there's not enough precision to represent the picture, or because your video card does not support half type. Try setting the "Upcast cross attention layer to float32" option in Settings > Stable Diffusion or using the --no-half commandline argument to fix this. Use --disable-nan-check commandline argument to disable this check.But only with JIB; does anyone know how to fix this? I read something about VAE but i'm not sure how to do it...?
Are you using it in ComfyUI or something else?
Go to your installation folder and look for a file named "webui-user.bat". Right-click it and select Edit. Find the line that says:
set COMMANDLINE_ARGS=
and add the following (including the hyphens):
--disable-nan-check --precision full --no-half --no-half-vae
Save the file and restart. That should fix the NaN errors. Just note that image generation may be a bit slower with these settings.
Edit: Done thanks to J1B
SDParameterGenerator
SDPromptReader
I tried everything to find these world most hidden nodes who don't want to install in my python 3.10.11 if this python compatibility issue you guys must mention that.
This post on there Github helped me lately when I had to reinstall and it wouldn't work: https://github.com/receyuki/comfyui-prompt-reader-node/issues/123
git clone --recursive <repo>
Hopefully that helps, I find them to be very useful nodes, but if you cannot get it working change them to standard load and save images nodes.
@J1B Yes this helped thank you so much.
Best for realism and skin detail. What needs to improve is emotions(badly). Beside smile and neutral, it barely knows anything.
Yeah SDXL has limited vocabulary and prompt following because of the the older Clip text encoder and model architecture. My Flux or Wan model should be more nuanced.
The most favorite model for creating skin, body proportions, and especially hands. It works very well to improve the condition of the skin and hands in Upscale and Inpaint. This works great in Fooocus. Sincerely, thank you for your work.
Not just for NSFW images though it does excel at those also I have used it for several fun science fiction and other renderings and its versatility in many genres makes it probably my favorite checkpoint for use in comfy. Also works great when you want a more realistic non fashion model look to your render characters.
Thanks, yes a flexible checkpoint is always what I aim for with Jib Mix SDXL.
Petition to bring this model back to the on-site generator!!
I am
Yeah, sorry I got distracted last night and forgot to increase the bids in the auction.
I wish there was a way to just pay to get it put back on instead of waiting a week
@J1B 100% also great work my man, you definitely have the best models on the platform for photorealism
I am joining : Petition to bring this model back to the on-site generator!
Yeah, sorry I got distracted last night and forgot to increase the bids in the auction.
I wish there was a way to just pay to get it put back on instead of waiting a week.
@J1B You should no pay/bid to keep such good model in. I really hate this policy. Thank you for the good job.
I am joining : Petition to bring this model back to the on-site generator!
It is back on again this week, but I am scraping the bottom of my Buzz barrel, so I am not sure if it will always be available unless lots of people use it.
@J1B I've been generating like mad on it, hoping it catches fire again!
Yeah what the heck. This was my favorite model! I can't run comfy locally plzzz bring back onsite!!!
DAWW NNAWWWWWWW
HASHTAG JIBMIXMATTERSSSSS
Ya GOT to be JIBBIN ME
It is back now!
@J1B thats what I'm jibbin about
I am joining : Petition to bring this model back to the on-site generator!
hey man thanks or your service let us know if you will ever migrate to a free platform for nsfw
What platform is that?
Damn, this model (v180) is excellent. Thank you very much for sharing the model.
Beautiful model with beautiful output. A gentleman's finetune.
Love the attention to detail in the prompts. very realistic and accurate generations around (CFG 7/30 Steps on DPM Kara)
BRAVO!!!!!!!!!!!!!!!!!!!!!
I opted out for the Jab and got the Jib instead!
Those who got the jab(clearly those whom down thumb'd), have you gone to the Dr. and get checked for cancer (cancer blood test)? You might receive some bad news. Should have gone for the Jib instead!
one of the best models ive ever used. this was my main go-to model before zimage came out, this is the one sdxl model ill keep forever :)
You're a great author, and every output is social media-worthy. The only issue is with the images are specifically the fingers and toes.
Yeah, that is just the SDXL model being SDXL, it is old tech now , z-image turbo models fix this nowadays but my ZIT model is not supported to be on the Civitai generator yet, but my lora can get you part of the way there if used with z-image base: https://civitai.com/models/2194714/jib-mix-realistic-z-image-lora
@J1B yeah it is working good in sdxl after using lora but i have to learn about z image turbo model i dont know that thing
Hey! On Comfy, when I set clipskip to -1, which is default, v18 gives a black image.
BUT, since a few days ago I've been enjoying clipslip set -2, and then v18 works? What's going on?
This is only for v18. I re-downloaded v17 to check whether I get into the same issue, but it works fine.
Hmm I am honestly not sure, I never really play with clip skip as I think it is recommended more for anime models, as it throws away the last x layers of the model that can add details to the image.
Disabling the node is the same as skip=1, I also have that issue on many models.
heey, is the fp32 full model still coming for v18? :)) i can see you´ve uploaded them for older versions.
I could... I will do some testing and if it is actually better in fp32 I will upload it , the changes are usually only very very slight.
I have posted a fp32 version now: https://civitai.com/api/download/models/1966530?type=Model&format=SafeTensor&size=full&fp=fp32
It quite often makes slightly more coherent images (maybe with slightly better/differenet prompt following and often smaller fine grain details but unless you are comparing them side by side you cannot usually pick out much difference to the fp16: https://civitai.com/posts/26540673
Is the full version has baked Text Encoder & VAE?? Why is it like 12 GB..
No, no Text Encoders or VAE baked in, most people that post SDXL models to Civitai post fp16 versions because that quant is "good enough" but the original SDXL release was in full fp32 weight, it is very slightly better at generating images and might be more consistent with hands/limbs, but if you do not have a lot of VRAM just use the "standard" fp16 6GB model.
@J1B i am already using it and you trained a very well performing model, it can do some of the things that even new Flux.2 Klein can't do properly.. back pose & POV..
Downloaded all 3 workflows today. basic works ok, Jib_Mix_SDXL_V18_Intermediate_Work_Flow moans about missing addons on, installed but when restart same moan? sdpromptreader and tensroRTloader
found your old post about doing the custom node thing
Really like this model. Any hint on how to train LORAS for it? Any special requrements?
No you should be fine to even train loras on the SDXL base model and use them on my one I think, it is usually fine.
Best i got so far are LORAs made with Juggernaut. I tested JIB MIX to create loras but they are not close to similar. Base SDXL are not as good as Juggernaut. Was hoping that there is any other way.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.