Illustrious Renaissance v6_NS Update
First things first: all the preview images you see here are pure Text-to-Image (t2i). No LoRAs, no Hires.fix, and no Face Detailer were used. What you see is exactly what the model can do out of the box.
Sorry for the silence! This update took a bit longer than I expected, but it’s because v6_NS is a total game-changer for the model's future.
I’ve been spending a ton of time in the "lab" experimenting and training text encoders from various models. The goal? To bake in enough flexibility so you can use your favorite LoRAs without headaches.
What does this mean for you? The model is now a bit of a hybrid. I’ve tuned it to be compatible with almost everything: SDXL, Illustrious, and Pony. I’ve also started adding support for Noob—some things already work, and full compatibility is next on my list.
It’s a work in progress: Is it 100% perfect? Not yet. But it’s finally at a point where it plays nice with a huge variety of LoRAs, and I wanted to get it into your hands as soon as possible.
I need your eyes on this! Try it out, break it, and please share your feedback and images. I really need to know how it’s performing for you so I can keep polishing it in the next updates.
Enjoy the new versatility! I hope you like it. —DeViLDoNia
⭐ Illustrious Renaissance
Quick Start: For v5 and earlier versions
Recommended Sampler: euler, euler_ancestral, dpm++ 2m
Steps: 30+
CFG: 3.5–7 (5 recommended for most cases)
Clip Skip: 2
Hires Fix / FaceDetailer: Strongly recommended
Resolution: 1152x1536 / 1024x1536 / 832x1216 / 896x1152
Recommended Upscaler: 4x_NomosWebPhoto_RealPLKSR.pth
Upscale Factor: ×1.5–×4 (depending on your hardware)
Trigger Words: For v5 and earlier versions
realistic_skin_texture, realistic, realistic_detail, detailed_background
Positive Prompt Example:
masterpiece, best quality, YOUR PROMPT, realistic_skin_texture, realistic_detail, detailed_background
Negative Prompt:
worst quality, low quality
Enjoy and give me all your love: like, follow, review, post, it helps a lot.
~ DeViLDoNia
📧 Business Contact:
For commercial inquiries, collaborations or licensing, please contact:
📩 email: [email protected] / Discord: hades5826
Description
Improved color rendering for richer tones and natural balance
Slightly better prompt adherence, especially in complex compositions
Overall output quality improved subtly but noticeably
FAQ
Comments (3)
V1.D15
1. Age adherence: confirmed, works well with my ``adult, grown-up, mid-20s`` ! :D
2. Image quality feels "flux like" (for an IL checkpoint ; havent used flux in almost a year though)
3. Sadly, hands/feets gets way worse with 21:9 (quite common)
4.1. Good overall adherence (i do have complex prompts!)
4.2. Despite my PG-13/R prompts, there are occasional "slips" of R/X/XXX in the image results, not bad per se, just fact.
4.3. I mean these (scene/skimpy prompts) - links contain -> NSFW / XXX:
- https://civitai.com/images/84634267
- https://civitai.com/images/84644549
Thank you so much for such a well-structured and insightful comment 🙏 — seriously, it's incredibly helpful. When you're trying to identify real issues in a model, detailed feedback like this from actual users is absolutely essential.
About the hands and feet issue, yes, I'm aware of it. I'm still unsure whether it's due to limitations in my dataset or inherent to the model architecture itself. I suspect it might not be fully solvable — even with higher resolution datasets and more wide-format training images. That said, I do plan to experiment with those in upcoming training rounds, so we'll see if anything improves. But honestly, I think it's more of an XL limitation, though I’m hopeful I can at least minimize it as much as possible.
Regarding the NSFW "slips", I’m fairly confident that in the V2 they won’t happen nearly as often — or maybe not at all. The more balanced and dense the dataset becomes, the less it tends to occur. The challenge, however, is that models tend to "forget" NSFW content during training — and worse, it often affects deepdanbooru tagging, which is one of the key reasons to train these kinds of models (because of their adherence). It's a balancing act and definitely makes the process more complex.
Also — I absolutely loved the images you linked on the wall. Really impressive stuff! 🔥
That’s all for now. I'm taking things slow and carefully with this model, because I genuinely plan to use it for my own creative work. At the same time, it’s a learning experience that I hope will lead to something positive for the future.
Thanks again!
@DeViLDoNia You're welcome.
I'm not sure if you're saying the NSFW slips are due my current use of Danbooru tags, or that the "slips" shouldnt (or no longer) occour once you "updated" the next dataset with them?
Either way, dont worry too much about it, it is also quite common, I havent met a single checkpoint where this didnt happen at all - alas - it does happen slightly more often with this one. (as i stated before, its not bad, and I would blame it to be due of a pony/illustrious source)
It FEELS like my use of "see-through_clothes" increased the chance of removing clothings "alltogether" - at times.
Also, I've figured that my "unsex" job isnt completed, meaning, I noticed that 2 or 3 images still had some explicit (R/X/XXX) trigger words in their supposidly "skimpy" (PG-13/R) prompt.
That said, I do have "applied" (random) clothing/(predefined-outfit-sets) too all my scenes, with low weight - as in [blabla]
Same about hands, i just noticed because recently I was doing images with checkpoints that have a very (exeptionaly) high success rate with hands/feets, no matter the aspect ratio, but i'm also fully aware, that this still is a very common issue among all checkpoints - in the end, its still "solvable/fixale" using Adetailer hand-fix, so here as well, dont worry too much! :)