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.<ul><li>Changed from PonyV6 to base IllustriousXL</li><li>Improved dataset/tagging should make the model even more flexible</li><li>New version should be more receptive of different art styles while also able to keep the original style if desired</li><li>New prompting structure with extra key words to accommodate model flexibility</li><li>Trigger words: womanmri, brown hair,green eyes,low ponytail,large breasts</li><li>Optional additional words: long hair, narrowed eyes, terasu mc</li><li>Additional negative trigger/key words: sketch, flat color, alt style,3d koikatsu,</li><li>Weight: 1</li><li>Trained on <a target="_blank" rel="ugc" href="https://civarchive.com/models/795765/illustrious-xl">Illustrious-XL </a>but I'm prompting it on <a target="_blank" rel="ugc" href="https://civarchive.com/models/827184/wai-nsfw-illustrious-sdxl">WAI-NSFW-illustrious-SDXL</a></li><li>Illustrious-XL works best with Euler a.</li><li>Clip Skip 1.</li><li>For resolutions lower than 1280x1280 Hires Fix is recommended for full body shots.</li></ul>My personal preference for prompting her:<ul><li>main prompt:</li></ul>masterpiece,best quality,amazing quality, womanmri, brown hair, low ponytail, green eyes,hair scrunchie,large breasts<ul><li>negatives:</li></ul>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: <a target="_blank" rel="ugc" href="https://discord.gg/DGAVSyfxPH">Discord</a> <a target="_blank" rel="ugc" href="http://X.com">X.com</a> <a target="_blank" rel="ugc" href="https://www.pixiv.net/en/users/106341858">Pixiv</a>If you want to support me: <a target="_blank" rel="ugc" href="https://patreon.com/lizarb?utm_medium=clipboard_copy&amp;utm_source=copyLink&amp;utm_campaign=creatorshare_creator&amp;utm_content=join_link">Patreon</a> <a target="_blank" rel="ugc" href="https://ko-fi.com/lizarb">Ko-Fi</a>////////////////////////////////////////////////////////////////////////////////////////////////////////Update PonyXL v2<ul><li>Changed from autismMix to base PonyV6</li><li>Improved dataset should make the model more flexible</li><li>New version should be more receptive of different art styles while also able to keep the original style if desired</li><li>New prompting structure with extra key words to accommodate model flexibility</li><li>Changed trigger word to " womanmri "</li><li>Additional positive trigger/key words: green eyes, brown hair, low ponytail, motherly</li><li>Additional negative trigger/key words: sketch, flat color, alt style,bold lines</li></ul>My personal preference for prompting her:<ul><li>main prompt:</li></ul>score_7_up,womanmri,green eyes,low ponytail,brown hair,motherly,large breasts<ul><li>negatives:</li></ul>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<ul><li>PonyXL version added!</li><li>Trained on <a target="_blank" rel="ugc" href="https://civarchive.com/models/288584/autismmix-sdxl">autismmixSDXL_autismmixPony</a></li><li>Thanks to PonyXL, the model is now more flexible compared with the 1.5 ver</li><li>Lora Block weights is no longer necessary for this model in PonyXL</li></ul>Update v3<ul><li>Added new images to the training dataset</li><li>Changed base model to <a target="_blank" rel="ugc" href="https://civarchive.com/models/23900/anylora-checkpoint">AnyLoRA</a></li><li>Important: To avoid overfitting I recommend using the extension <a target="_blank" rel="ugc" href="https://github.com/hako-mikan/sd-webui-lora-block-weight">sd-webui-lora-block-weight</a> and setting the weight to: OUTALL</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/user/lizarb">Also check out the other characters of this pack!</a></li></ul><h1 id="model-characteristics-ilciby48a">Model Characteristics</h1>There are 4 different LoRA's available, one for each of the characters and each with its own trigger word.<ul><li>Trigger words: mori_kaede_\(terasu_mc\)</li><li>Use weight of 0.7 to 1. See what works best for you.</li><li>I'm using <a target="_blank" rel="ugc" href="https://civarchive.com/models/6174?modelVersionId=238656">DivineEleganceMix</a><a target="_blank" rel="ugc" href="https://civarchive.com/models/58431"> </a>and <a target="_blank" rel="ugc" href="https://civarchive.com/models/54204/cartunafied">Cartunafied</a> and got good results, so I'd recommend using checkpoints that have an anime style.</li><li>Play around with the samplers, but I got best results with DPM++ 2M Karras and DPM++ SDE Karras.</li><li>Clip Skip 2 works best.</li><li>The Ayumu LoRA can be a bit tricky to get good results due to limited dataset.</li><li>Hires Fix works like a charm, I recommend using Lanczos or R-ESRGAN 4x+ Anime6B as the upscaler.</li><li>Although not fully tested, I got some interesting results when using <a target="_blank" rel="ugc" href="https://civarchive.com/models/105510/terasumc-art-style">dramaticgecko1600 TerasuMC Art style</a>, so I'd recommend to give it a try and see what results you get ;)</li></ul><h1 id="training-yid8fe7p6">Training</h1>I'm relatively new to LoRA trainning so any tips for improvement are much appreciated!<ul><li>Dataset size: 80 (Combined)</li><li>Model: <a target="_blank" rel="ugc" href="https://civarchive.com/models/23900/anylora-checkpoint">AnyLoRA</a></li><li>Buckets enabled</li><li>Epoch: 1</li><li>Repeats: Good dataset - 100 &amp; Bad dataset - 50</li><li>Total steps: 12000 (Combined), Around 3k each.</li></ul>////////////////////////////////////////////////////////////////////////////////////////////////////////Update v2<ul><li>Added new images to the training dataset</li><li>Changed base model to <a target="_blank" rel="ugc" href="https://civarchive.com/models/136054/hardcore-hentai">Hardcore-Hentai</a> for better r18 scenes</li></ul>Like my stuff? Buy me a Monster!<a target="_blank" rel="ugc" href="https://ko-fi.com/lizarb"><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/0a22fdbf-afd8-4917-b830-5c33b5f86940/width=525/0a22fdbf-afd8-4917-b830-5c33b5f86940.jpeg" /></a>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 <a target="_blank" rel="ugc" href="https://civarchive.com/models/23900/anylora-checkpoint">checkpoint</a>, but it should work well with pretty much anything anime related.You can also use <a target="_blank" rel="ugc" href="https://civarchive.com/models/105510/terasumc-art-style">TerasuMC Art Style</a> to really bring up TerasuMC's art style to the image!<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e201f736-c4ed-4f7e-bfb4-c97f4715b99f/width=525/e201f736-c4ed-4f7e-bfb4-c97f4715b99f.jpeg" />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!<a target="_blank" rel="ugc" href="https://ko-fi.com/lizarb"><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/0a22fdbf-afd8-4917-b830-5c33b5f86940/width=525/0a22fdbf-afd8-4917-b830-5c33b5f86940.jpeg" /></a>All my models and images will always be 100% free, don't feel obligated to donate!

Mori Kaede [TerasuMC] [黒人の転校生に幼馴染をNTRる] - Lora

<ul><li>More samples added to training dataset</li><li>Changed base model from Anylora to <a target="_blank" rel="ugc" href="https://civitai.com/models/136054">Hardcore - Hentai v1.2</a> for better r18 scenes</li></ul>

Files with SHA256: 5341dc491cd7f9287167c9d8cfebca13b8c0d5fefac312afaf151fa0d3df6883