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
A new version of my SDXL model with some more detail in skin and backgrounds and my first release of a 32fp version of my model that has slightly better (almost sub-pixel) sharpness.
FAQ
Comments (19)
keep up the good work!
Super cool there is a new SDXL version with v15. Didn't expect this post-Flux, can't wait to try it out! 😎🙏
Face Loras end up with stretched and elongated faces with this v150 checkpoint. never had that before. Any suggestion why such a thing could happen with only this one checkpoint?
Yes , I did notice that in testing, if the aspect ratio is set quite long. no idea what has caused it! it is annoying, try and use a more square aspect ration seems to fix it. I will have a think about what might have caused it and if I can correct it in a future release.
@J1B i redid the face loras, and yes square is the only option right now ... also it think a list of which trigger prompts work mostly the best would be helpful, i am not getting such great result with prompts right now. the woman all get fat wit faec loras as well, which is something i have never encountered before. i encountered too slim and anorexic woman body shapes, but never have i encountered fat necks and fat shoulders and body's. i am excited though that background and environment are a tiny bit better than most checkpoints....
@akos I think I have found the issue, I merged in the SDXL Add Detail Lora <lora:add-detail-xl:1> and it is making the faces long especially if you use it again, I have found using it with a negative weight of -2 helps get the face shape better <lora:add-detail-xl:-2>
https://civitai.com/models/122359
You can even see the issue in the sample images provided by the lora.
You would think that after over 36 million generations with this lora someone else would have noticed that!
@J1B cool, thanks...i will test... what an odd thing to occur
@J1B did you train your checkpoint on 512x512 or on 768x768 or on 1024x1024 images?
@J1B If you look at the safetensors metadata for Detail Tweaker XL, it was in fact trained on literally thousands of images with the SOLE resolution bucket being 1024x1024. So it doesn't support any other aspect ratio at all, technically.
@akos the issue is definitely just what Jib said, Detail Tweaker was trained on a really insufficiently diverse dataset in terms of resolutions
@diffusionfanatic1173 where , or how can i see such checkpoint metadata? Detail tweaker XL is a LORA.... and this here is a checkpoint ....
@akos The "Checkpoint Merger tab in Automatic1111 has a "Read metadata from selected checkpoints" button, that is where I would go to read that info. Kohya_ss is the tool I use to merge Checkpoints and LoRa's
@J1B so cool, thanks.... i was wondering if thats possible to merge checkpoints with face loras. and if the result would be then better, compared to add the face loras into a prompt, or even use adetailer face yolo to do face swapping but using inside of the adetailer the face lora in the prompt of a detailer. i was doing adetailer for each time face loras were part of the main prompt the checkpoint would return really bad results ...not follow the prompt on background or even crop.
@akos No, I don't think it makes the quality any better and you cannot change the weights so there is less tweaking.
4 step lightning version please
This is the best Checkpoint which I used till date. Images ae of superior quality. Only issue which i am facing is not able to generate full body images.
Thanks. Probably the most effective way to get further away full body shot is to use a zoom slider Lora with a negative value. (Or the Negative version with positive value if using the onsite generator):
https://civitai.com/models/213307?modelVersionId=949771
You can sometimes get away with not using a lora if you descried the feet or shoes of the person, but the Zoom Lora is best.
@J1B Thanks a ton bro.
Jib Mix Realistic XL stands out! In my opinion the perfect checkpoint. Incredibly versatile and varied.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.