You can get this model, along with many others, by joining The Tinkerer on Whop. It’s one monthly membership, and you’ll also get early releases, private tools, and a bunch of extra stuff.
CyberIllustrious blends the artistic flair of Illustrious-XL with a big dose of CyberRealistic’s style. You still get all those cinematic, dreamy strokes, but now they come through with crisp, realistic detail.
If you want that painterly, elegant look but still want your images to feel sharp and lifelike, this is your model. It’s great for editorial vibes, soft glam, or even fantasy art that looks like a real photo. Basically, it just nails it.
It’s powerful, easy to use, and honestly? Prompting it is seriously fun.
📘 Get the most out of this model with the Illustrious / NoobAI Prompting: Master Guide.
⚙️ Recommended Settings
Sampling method: DPM++ SDE Karras / DPM++ 2M Karras - SGM Uniform / Euler a
VAE: is already Baked In
Sampling steps: 30+ Steps
Resolution: 896x1152 / 832x1216
CFG: 4-5About (DMD2)
The special DMD2 edition of CyberIllustrious is created for faster, cleaner generations while requiring fewer steps.
DMD2 takes the balanced improvements of the standard release and optimizes them for speed, precision, and responsiveness, making it ideal for both SFW and NSFW workflows.
Expect tighter results, smoother control, and a model that reacts beautifully to the classic Illustrious prompting style.
Recommended Settings
Sampler: LCM
Schedule Type: Exponential
Sampling Steps: 8 (12 for extra fine detail)
CFG Scale: 1.5
⚡ These settings are optimized for DMD2 and are different from my standard Illustrious workflow!
🧠 Prompting Made Easy
Positive Prompt:
masterpiece, best quality, high quality, ultra-detailed, realistic, ... rest of promptNegative Prompt examples (optional):
cartoon, illustration, anime, painting, CGI, 3D render, low quality, watermark, logo, labeldeformed, deformed face, low quality, bad quality, worst quality, (drawn, furry, illustration, cartoon, anime, comic:1.5), 3d, cgi, extra fingers, deformed teeth, watermark, logo🩺 Optional ADetailer Settings
Adetailer model: face_yolov9c.pt
If you only want the main face being refined set 'Mask only the top k largest' to 1.🔗 Resources
Backup location: HuggingFace
☕ Support the Madness
Liking what you see? Grab me a digital espresso (or a couple!). It keeps the creative energy going and helps bring more cool prompts and models your way.
⚠️ Heads Up
This model can spit out some mature stuff. Use it wisely, and please, no cursed creations—we’ve really seen enough of those.
💬 Want Better Prompts?
Use this custom ChatGPT to auto-generate polished prompts for CyberIllustrious:
🔗 [Try it now on ChatGPT]
Description
The DMD2 edition of CyberIllustrious V7.0 has been rebuilt from the ground up, designed to generate faster, ultra-realistic results with fewer steps.
Recommended Settings
Sampler: LCM
Schedule Type: Exponential
Sampling Steps: 8 (or 12 for extra detail)
CFG Scale: 1.5
FAQ
Comments (32)
I typically NEVER comment on anything, but this one time I'm gonna due to the necessity of telling you something. You absolutely knocked this one out of the park, the prompt adherence is great, the quality is great v7.0 is FANTASTIC, and feels just great to work with, your efforts absolutely were worth while, and this is simply a masterpiece imho. Well done @Cyberdelia
Thanks, much appreciated.
What are the differences between v7 and v7 DMD2?
DMD2 has DMD2 LoRA merged in, allowing it to create image in about 8 steps (altho I would suggest going for 12-16). Its like old Turbo/Hyper, just better.
@Mescalamba I see. What would you recommend? With or without DMD2?
@Lopodelas Thats a hard question.
Honestly I pick without DMD2 and if I want DMD2, I have DMD2 LoRA, so I simply enable that and have both.
https://huggingface.co/tianweiy/DMD2/tree/main
You want to get dmd2_sdxl_4step_lora.safetensors
Its big, but works much better than fp16.
You can use it on any SDXL, ILLU, PONY model you find. Use 1.0 strenght, if you are using ComfyUI, use LoraLoaderModelOnly node and add it after your other LoRAs.
Just remember that if you are using DMD2 LoRA, you must use CFG between 1.0 and 2.0, or it just wont work (there are workarounds, but it would be long to explain). Also not all samplers work for it, but for example LCM will work every time, or Euler/Euler_A.
Basically all DMD2 models are just this LoRA fused into model, nothing else. It can sometimes mess with model knowledge and thus prompt follow. It doesnt happen if you use it added as LoRA. Keep in mind, you do need more VRAM then (cause its 800MB LoRA).
@Mescalamba Thank you for your thorough reply
@Mescalamba the fp16 version is just 375 mb. It has nothing to do with vram anyway. You can apply 20 loras, the speed is the same, atleast in Forge.
Ok, i tried fp32. The difference is so small, any general detailer lora has more and better impact.
@gregoribanner970 Ive mentioned VRAM, cause each lora adds to VRAM. You would need to do a bit more extensive research to find out if fp32 and fp16 is equal. From my experience with many (think 50+) models its better.
That extra 500MB is worth it for me. Doesnt need to be worth for everyone, just my personal experience, thats all.
@Mescalamba thank you that solved my question
4 step DMD2 seems to work better with the LCM sampler/scheduler. 6 steps is better. (I'm using Fooocus)
I am sorry to be naysayer but v7 still appears to lost a lot of illustrious knowledge. Just try "crotchless_pantyhose", instant fail. And this is 1k+ tag on danbooru. Prompt following in general is sketchy at best. The texture is great though.
Can try if it can do that tag, might need extra nudge.
Thing is, most realistic ILLU is done by merging with either realistic LoRA or directly with SDXL (or optionally PONY real models.. which often are created exactly same way).
Issue with "real" is that it somehow takes considerably more space in model and reduces model knowledge about "not real" stuff (eg. stuff thats in base ILLU/PONY).
Probably tied to amount of knowledge needed to make highly detailed photo compared to detailed anime picture. Only solution would be bigger models.
@Mescalamba It is definitely possible. For example "Rillusm" model understands this tag (stupid censorship). While Cyberrealistic always resorts to simple "thighhighs"
@somedoby @Mescalamba I know what's happening and why (I think). Will take a look/investigate. Thanks for the useful feedback!
@Cyberdelia Thank! I can't even type the damn tag in a different language, civitai is just wow.
@somedoby @Mescalamba https://civitai.com/models/1986798/cyberrealistic-cyberillustrious-catalyst
What is the 'Civitai Browser'?
@Cyberdelia we can't find it in civitai's search either. Or tell me what search to do to find it ;-)
@LittleZomb @LittleZomb yes there was a problem releasing this model and I think your problems are related. Support knows about this. Thanks for letting me know.
@LittleZomb and the problem was technical... To be clear
Fp32 ?
v7 IS a step up from v6.
but it's still a step down from v5, imho
I feel the same. somehow v5 is still the best in prompt adherence and overall generation
@REALMZEROFIVE so do you feel v7 any good?
I dont think I liked v7, maybe its just certain prompt setups or Loras but I didnt like some results i got with same setups I used on v6/5
do you get same results with v6 & v5? I feel v5 still produces the best results
@Danielfdo_13 yes I was getting similar feel in v5 and v6, very authentic/real looking, and great at amateur. v7 things went in a different direction.
My legend. I'm already in love with v60 but I'll give this one a try <3
I love that model, but I struggle with inpainting. With other (PonyXL) models it is really easy. But with illustrious models the inpaintet areas are not content aware. What do I do wrong?