V2This model was created to refine Canopy: The Digital Alchemist to better suit my personal workflow. My goal was to ensure zero errors with my existing WebUI A1111 ADetailer settings, while pursuing a specific style: distinct outlines, exaggerated body proportions, and unique facial deformation (deformé).For the past few days, I’ve worked hard to improve style consistency and eliminate AD issues without sacrificing prompt adherence, all while maintaining the original vibe I loved in the previous version.These past few days have been a realization that to truly achieve the results I want, I need a deeper understanding of the NoobAI model structure and the technical ability to actually train the models myself. (I also learned the hard way that experimenting with different settings is incredibly time-consuming and that my current computing resources are quite limited.)As a result, I’ve reached a compromise with v2 and v2-Dash. To be honest, they’ve drifted away from the v1 style I initially aimed for, and I can’t guarantee an increase in prompt adherence. However, I have made significant progress in improving style consistency and noticeably reducing AD issues (though it's not perfect).Interestingly, I suspect these issues persist partly because the prompt adherence is actually higher than before, causing it to behave differently from the model that I used to. If you don't apply specific prompts to ADetailer, the issues should be much less noticeable. (I personally use wildcards for emotions and facial features to maximize randomness). I’ve also improved the unintended 'over-detailing' issue.While it's not without its lingering regrets, I think this is as far as I can take this model.Have fun~!V1This model was made for my own use and is compatible with WebUI A1111.The overall quality is slightly lower than typical NoobAI models, but it still outperforms typical Illustrious XL models.Exaggerated Stylized illustration(comics) style.Please note that style consistency is a bit lacking and occasional artifacts may occur.As for updates to fix these weaknesses, it's currently undecided.I noticed that the style consistency is significantly lacking compared to my expectationsI uploaded the wrong file.Apologies for wasting your time due to my oversight.<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:14px">I've finished re-uploading it.

PL3rd_ExgStlzNoob

pl3rdExgstlznoob_pl3rdExgstlznoobV10.safetensors

PL3rd_ExgStlzNoob_v1.0

PL3rd_ExgStIz_v2

The recommended settings are as followsBaked in VAEcfg 1.5~3(A high CFG scale increases saturation and makes it more 2D)steps 7~10(A high steps make it more 2D style)Sampler: Euler aV1,1Improvements:Improved facial style consistency.Prevents unnecessary background details when applying hires.fix (very Slight). (when simple background or abstract background).Reduced frequency of noticeable issues with After Detailer hands.(imperfection)Slight increase in prompt understanding.Bit more saturationchanging body proportionDisadvantages:Slight decrease in realistic textures.Minor changes in art style and facial style compared to v1.Note:Lora is not working well, especially those related to art or artist stylesThis update is for v1 users and minor updates, so if you are satisfied with v1, there is no need to upgrade to v1.1 .Remaining Tasks: Unwanted patterned pantyhose (or stockings), the level of realistic sheer pantyhose (stocking) texture.However, whether this is within my capabilities is uncertain.If you want to save time and achieve images closer to a realistic feel, try using the settings from the example images.A high CFG and high steps will create 2D style images, so they will be completely different from the example images I providedI've just completed my first attempt at making ILXL model, just merging not actual traning.The base model is the uncannyValley_ilxl10Noob version.I applied mbw to the lobotomyMix_v1 merging and then applied mbw again to the illustrioussatIOR merging.I adjusted the brightness, contrast, and detail/noise appropriately.The goal was to make the base model slightly darker and the textures more realistic, aiming for something close to realism while still maintaining a high level of prompt comprehension.I aimed to achieve two things: increasing the texture detail while adjusting the facial style to match my desired look. However, it’s not perfect.There was a very slight loss in prompt comprehension, but there’s a small improvement in background or texture expression.(More attempts bring more frustration. The more I merge, the more the results deviate from what I want)For now, this is the best I can do.Please refer to the EXIF data of the example images.(I used a low CFG value to emphasize the realistic feel.)Feel free to test the prompt as you like. However, a quality prompt is not necessarily required.I highly recommend using extensions like After Detailer(ADetailer).It has great benefits for generating full-body images<s>The ultimate goal is to combine the realistic textures of the SD1.5 model, the sentence comprehension of ILXL, and the facial style of my ideal type.</s><s>I tried merging Lora into the checkpoint, but The Supermerger isn't working as expected.</s>V1Weaknesses:1. Although very subtle, facial consistency is slightly lacking.2. The overall anatomical accuracy, especially in the hands, is slightly unstable.P.SI thought I had tested enough while testing the merge.After uploading to civitai, I'm continuing to try additional generating.Character recognition lost a little more than expected.However, the realistic textures gained more than expected on some of the horror feel images.I can communicate in Korean and English.私は日本語が少しできます我不会说中文，但可以借助 AI 帮助Thank you, and I hope you enjoy!V1.1lora가 잘 안먹히는 모델이다. 특히 art 나 artist style 로라가 잘 안 먹힌다.변경점:얼굴 스타일 일관성 향상.hires적용시 배경에 불필요한 디테일이 늘어나는 것을 방지. (아주 약간). (simple background 나 abstract background 일 때)손 ad 티나는 이슈 빈도 감소. (불완전)prompt이해력 미세한 증가채도 약간 증가몸매 비율 변화단점:리얼리스틱한 텍스처 소폭 감소.v1과 그림체 약간 변화, 얼굴 스타일 약간 변화.참고:v1 사용자들을 위한 작은 업데이트이므로 v1에 만족한다면 굳이 v1.1로 버전업 안 해도 됨.남은 과제 : 원하지 않는 패턴스타킹, 리얼리스틱 투명스타킹 질감 정도. 하지만 내 능력으로 가능할 지는 미지수.드디어 ilxl모델 제작에 도전하게 되었습니다. 비록 학습이 아닌 merging이긴하지만요.베이스모델은 uncannyValley_ilxl10Noob 버전입니다.lobotomyMix_v1를 mbw로 병합하고illustrioussatIOR 모델을 다시 mbw 로 병합하고밝기, 대비와 detail/noise를 적절히 조절해주었습니다.목표는 베이스모델을 살짝 어둡게, 텍스처는 리얼리스틱하게 해서,리얼리스틱에 가깝지만 프롬프트 이해력은 여전히 높게 만들고자 했습니다.텍스처 디테일은 높이면서 얼굴스타일을 제가 원하는 스타일데로 맞추는 것, 두 가지 토끼를 잡고자 했는데 완벽하진 못했습니다.프롬프트 이해력을 아주 정말 미세하게 손해보았지만, 배경표현 혹은 텍스처에 미세한 증가가 있습니다.더 많은 시도가 더 많은 좌절을 불러옵니다. 병합 할 수록 결과가 원하는 방향과 다르게 나옵니다.현재로선 이것이 최선입니다.예시 이미지들의 exif를 참고하세요.(저는 실사느낌을 강조하기 위해 낮은 cfg값을 사용했습니다.)권장 세팅값은 다음과 같습니다.VAE가 포함되어 있습니다.cfg 1.5~7(높은 cfg 스케일은 채도를 높이고 2d화 됩니다.)steps 7~20(높은 steps는 2d화 됩니다.)Sampler: Euler a시간을 절약하고, 실사느낌에 가까운 이미지를 원하신다면 예시 이미지의 세팅값을 사용해보세요.높은 cfg, 높은 steps는 2d 스타일 이미지를 만들기 때문에, 제가 제시한 예시이미지들과는 완전히 다를 것입니다.프롬프트는 자유롭게 테스트해보시길 바랍니다. 하지만 퀄리티 프롬프트가 꼭 필요하진 않습니다.after detailer 류의 확장기능 사용을 강력하게 권장합니다.full body 이미지생성에 굉장한 메리트가 있습니다.<s>최종 목표는 sd1.5의 실사모델 텍스처와 ilxl의 문장이해력과 제 이상형의 얼굴스타일을 갖추는 것입니다.</s><s>로라를 check point에 병합시도해 보려고 하는데 supermerger가 작동하질 않는군요.</s>V1약점1. 아주 미세하지만 약간 얼굴 일관성이 떨어진다.2. 인체 완성도(특히 손부분)이 아주 조금 불안정하다P.S병합 테스트를 하면서 충분히 테스트 했다고 생각했습니다.civitai에 업로드후 추가로 생성을 계속 해보고 있습니다.캐릭터 인식율에선 기대보다 살짝 더 손해를 보았습니다.하지만 리얼리스틱한 텍스처는 일부 horror 느낌 이미지에서는 기대보다 더 이득이 있었습니다.한국어, 영어 가능합니다.私は日本語が少しできます我不会说中文，但可以借助 AI 帮助감사합니다. 즐겨주시길<pre><code></code></pre>문단 구분 왜 적용 안되나?

PL3rdILXLsemireal

Improvements:Improved facial style consistency.Prevents unnecessary background details when applying hires.fix (very Slight). (when simple background or abstract background).Reduced frequency of noticeable issues with After Detailer hands.(imperfection)Slight increase in prompt understanding.Bit more saturationchanging body proportionDisadvantages:Slight decrease in realistic textures.Minor changes in art style and facial style compared to v1.

Files with SHA256: d104d01d75511ab3a6cfc08c86f92d5768a089d48e8acca0119b69282f85ca77