Join The Tinkerer on Whop to download this model and many more — all for one monthly payment. Plus, get early releases, private tools and a lot more. 👉 Join on Whop
CyberIllustrious Semi-Realistic bridges the painterly elegance of Illustrious with the crisp, lifelike look of CyberRealistic.
Think: clean realism with a soft artistic touch - strong anatomy, polished skin and fabric detail, cinematic light, and smooth tag understanding for Illustrious-style prompts.
If you liked CyberIllustrious and CyberIllustrious Catalyst, this version goes further: clearer tag comprehension, steadier faces, better skin/detail handling, and more consistent semi-realistic renders - especially when paired with my Danbooru Prompt Crafter GPT.
What it’s great at
Semi-realistic portraits & fashion/editorial (soft glam → polished photoreal)
Character work that keeps a natural look while honoring Illustrious tag style
Cinematic lighting & moody color without over-baking effects
Cleaner hands, eyes, and hair with fewer artifacts at low CFG
Recommended Settings
Sampler: DPM++ SDE Karras / DPM++ 2M Karras / Euler a
Steps: 30–40
CFG: 4–5
Resolution: 896×1152 or 832×1216 (portrait-leaning works best)
VAE: Baked-in
Optional (face refiners):
ADetailer
face_yolov9c.ptIf you only want the main face refined, set “Mask only the top k largest” to 1.
Danbooru Prompt Crafter (GPT)
I’ve updated the Danbooru Prompt Crafter with a larger Danbooru tag list for more accurate, detailed prompts.
It’s built to write, clean, and optimize prompts for models that use Danbooru-style tags.
What it does
Converts descriptions to Danbooru tags: paste a natural sentence, messy tag list, or even an image → get a clean, copy-ready prompt.
Cleans up messy prompts: reorders tags, removes duplicates, and replaces vague terms with precise, valid Danbooru tags.
Multiple modes:
Strict Mode → accurate, clean Danbooru tagging.
Creative Mode (default) → adds aesthetic polish for more visually appealing results.
Output types:
Realism (default) → for photorealistic images.
Anime/Art → for stylized or illustrated outputs.
Tag validation & correction: fixes misspelled/invalid tags automatically.
Flexible formatting: choose commas, underscores, or spaces.
Understands images: upload an image; it analyzes character, pose, fashion, background, and lighting to generate a Danbooru-style prompt.
Notes & Etiquette
This model can generate mature content. Use responsibly and follow platform rules.
No explicit deepfakes of real people. Don’t be that person. 🙃
Description
This is a smooth polish update, not a radical overhaul.
Several improvements and behavioral refinements from CyberRealistic CyberIllustrious Anime have been carefully carried over, resulting in more consistent outputs and noticeably better prompt adherence. Danbooru tags are now followed far more reliably, making results easier to control and predict.
Most example images for this release were generated using the Cyberdelia Ralston (RK2) Sampler for Forge Neo, which pairs extremely well with this model and further improves overall consistency.
FAQ
Comments (22)
It looks absolutely gorgeous. What a masterpiece.
Congrats.
Where can I download the cyberdelia sampler?
Click the link :)
Or: https://www.patreon.com/posts/cyberdelia-rk2-147409612
Did you download the checkpoint? I accidentally uploaded the SDXL inpaint version. The correct version is almost uploaded.
Update: correct version is uploaded!
@Cyberdelia Damn, that's why nothing works imao. also I'm using Stability Matrix and I cann't find your sampler anywhere on it.
Im assuming RK2 will work in ComfyUI then too. Thats linear sampler, if I remember right? Kinda like midpoint, but will give it a try.
10/10 ! Good job as always !
v1.5 absolutely rocks! Awesome work, well done. What scheduler would do recommend for your CR (RK2) sampler? Thank you so much for making and sharing. ❤️
Auto
I really like this checkpoint version.
the old versions are awesome, also
Just a little request : The cyberdelia_ralston.py seems created for forge, so, are you planning to code similar module for comfyUi?
for advance, thank you.
No, probably not, because I don’t use it myself.
ok, thank you for your feeback
what is that and where can i find it?
v1.5 produces characters less faithful to the LoRA material than v1.1. It seems to force a slim, adult physique. Furthermore, contrary to what is stated in the "About this version" section, it doesn't appear to follow your prompt any better than v1.1. I ll stick to v1.1 after testing.
What I meant is that it follows Danbooru prompts better. I don’t know if you’ve tested those? But that’s what I was referring to.
I only use Danbooru tags. Using natural language with ILXL is like flipping a coin and hoping for a good result. I did about a hundred renders before posting. Therefore, the probability that these points are solely related to the seed is practically zero at this stage. Thus why i ll stick to v1.1. But this isn't a bad checkpoint at all. It just doesn't render stuff the way i expected after having been using 1.1.
Edit: here is a raw example using the same seed and no inpainting which illustrates that v1.5 does not follow the prompt as well as v1.1 and represents the character less accurately compared to v1.1 despite the LoRA. As you can see, v1.5 struggles with nails, produces a crowd, does not generate the spellcasting stance and does not respect the framing (it isn't a cowboy shot, as the field of vision extends well below the character's thighs).
v1.1: https://i.ibb.co/Ps7v1mgW/00001-773555297-v1-1.jpg
v1.5: https://i.ibb.co/sv4YMwJm/00001-773555297-v1-5.jpg
I'm a little conflicted - v1.5 and 1.1 are both very good! Compared to 1.1, the new version:
- does seem to follow the prompts a little better, especially with camera angles (eg from side)
- produces less jank (eg extra limbs)
- has better nonhuman understanding, especially characters
- produces cleaner backgrounds (less random noise)
On the other hand, 1.1 has better composition - the arrangement of the scene looks more real-world and it does a better job of avoiding the "AI-slop" look (less symmetrical, fewer repeating details, etc)
Like I said, they're both good though!
It’s not that you’re only allowed to use one version. You can simply use version 1.5 alongside 1.1 :)
Yep! I usually try to limit the number of similar models though, so I'll need to do a bit more testing to see if the differences are worth keeping both or picking just one. Mostly I just wanted to provide some feedback about what I liked from each.
@Mutaclone Yes, I understand and I really appreciate that! It was more of a joke.
Would you ever consider releasing a LLM tuned for prompt crafting, to download for offline use?, while travelling some people don't always have internet access. It would be a huge bonus to not have to plan ahead, prep a bunch of prompts, and only generate while travelling.
I’ve released a tool on Patreon that uses a local LLM to generate prompts. It connects with LM Studio and KoboldCPP, though you can also use LM Studio on its own if you prefer off course.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.