This is an early version!

When IllustXL gets to a more mature stage and the best practices for training and prompting are more clearly defined, this model will then be updated to a full release.

Changed from PonyV6 to base IllustriousXL
Improved dataset/tagging should make the model even more flexible
New version should be more receptive of different art styles while also able to keep the original style if desired
New prompting structure with extra key words to accommodate model flexibility
Trigger words: womanmri, brown hair,green eyes,low ponytail,large breasts
Optional additional words: long hair, narrowed eyes, terasu mc
Additional negative trigger/key words: sketch, flat color, alt style,3d koikatsu,
Weight: 1
Trained on Illustrious-XL but I'm prompting it on WAI-NSFW-illustrious-SDXL
Illustrious-XL works best with Euler a.
Clip Skip 1.
For resolutions lower than 1280x1280 Hires Fix is recommended for full body shots.

My personal preference for prompting her:

main prompt:

masterpiece,best quality,amazing quality, womanmri, brown hair, low ponytail, green eyes,hair scrunchie,large breasts

negatives:

lowres, (bad quality, worst quality:1.2), bad anatomy, sketch, flat color,jpeg artifacts, ugly, poorly drawn,muscular female,

Greetings!

After playing around with IllustXL, I've decided to post my best attempt on each of the KNT characters, and so far even at an early stage results have been very promising, but I still see some more room for improvement.

About the model itself, one major thing to point out is that using "3d koikatsu" on negatives yields interesting results that are closer in style to the original doujinshi, so please feel free to post your results using this method!

Also, note that I posted some image grids using different epochs and tags. The epoch on this version is epoch 8, which I believe gave the best balance between flexibility and fidelity, but if you would like to try different epochs feel free to reach out to me and I'll send them to you.

Here's my socials in case you want to keep in touch with what I'm up to: Discord X.com Pixiv

If you want to support me: Patreon Ko-Fi

////////////////////////////////////////////////////////////////////////////////////////////////////////

Update PonyXL v2

Changed from autismMix to base PonyV6
Improved dataset should make the model more flexible
New version should be more receptive of different art styles while also able to keep the original style if desired
New prompting structure with extra key words to accommodate model flexibility
Changed trigger word to " womanmri "
Additional positive trigger/key words: green eyes, brown hair, low ponytail, motherly
Additional negative trigger/key words: sketch, flat color, alt style,bold lines

My personal preference for prompting her:

main prompt:

score_7_up,womanmri,green eyes,low ponytail,brown hair,motherly,large breasts

negatives:

bold lines, censored,sketch,flat color, alt style,

About the tags 'motherly' and 'alt style'

In order to improve model flexibility fanarts were used, so to accommodate the deviation in style, a trigger word was added to the dataset so as to not let the model dilute her original style. So if you would like to keep the style as close as possible I'd recommend putting the tag 'alt style' on the negative prompt.

As for why using 'motherly' instead of something like 'mature female', it simply boils down to Pony not being able to decouple concepts from art style very well, so I've opted for a tag that would still retain some meaning from training while also not influencing the art style too much.

As always thanks for the support and stay tuned for the next two updates (Nao and Kanako)!

////////////////////////////////////////////////////////////////////////////////////////////////////////

Update PonyXL v1

PonyXL version added!
Trained on autismmixSDXL_autismmixPony
Thanks to PonyXL, the model is now more flexible compared with the 1.5 ver
Lora Block weights is no longer necessary for this model in PonyXL

Update v3

Added new images to the training dataset
Changed base model to AnyLoRA
Important: To avoid overfitting I recommend using the extension sd-webui-lora-block-weight and setting the weight to: OUTALL
Also check out the other characters of this pack!

Model Characteristics

There are 4 different LoRA's available, one for each of the characters and each with its own trigger word.

Trigger words: mori_kaede_\(terasu_mc\)
Use weight of 0.7 to 1. See what works best for you.
I'm using DivineEleganceMix and Cartunafied and got good results, so I'd recommend using checkpoints that have an anime style.
Play around with the samplers, but I got best results with DPM++ 2M Karras and DPM++ SDE Karras.
Clip Skip 2 works best.
The Ayumu LoRA can be a bit tricky to get good results due to limited dataset.
Hires Fix works like a charm, I recommend using Lanczos or R-ESRGAN 4x+ Anime6B as the upscaler.
Although not fully tested, I got some interesting results when using dramaticgecko1600 TerasuMC Art style, so I'd recommend to give it a try and see what results you get ;)

Training

I'm relatively new to LoRA trainning so any tips for improvement are much appreciated!

Dataset size: 80 (Combined)
Model: AnyLoRA
Buckets enabled
Epoch: 1
Repeats: Good dataset - 100 & Bad dataset - 50
Total steps: 12000 (Combined), Around 3k each.

////////////////////////////////////////////////////////////////////////////////////////////////////////

Update v2

Added new images to the training dataset
Changed base model to Hardcore-Hentai for better r18 scenes

Like my stuff? Buy me a Monster!

All my models and images will always be 100% free, don't feel obligated to donate!

////////////////////////////////////////////////////////////////////////////////////////////////////////

Greetings!

My first LoRA model here, and starting off with an absolute banger.

The model works well with weights between 0.7-1 + Euler a @ 25 steps, with the AnyLora checkpoint, but it should work well with pretty much anything anime related.

You can also use TerasuMC Art Style to really bring up TerasuMC's art style to the image!

I'd be happy if you guys could give me any character suggestions for my next LoRA!

Cheers.

Like my stuff? Buy me a Monster!

All my models and images will always be 100% free, don't feel obligated to donate!

Search AI Models & Addons

very few results