Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.
👉 Join on Whop
💬 Join the community for support, free tools and early news on Discord
CyberRealistic XL Catalyst V2.0 (DMD2)
This is a special DMD2 edition of CyberRealistic XL Catalyst, rebuilt from the ground up and designed for faster, ultra-realistic results using fewer steps.
It’s a direct evolution of CyberRealistic V7.0, delivering next-level quality for both SFW and NSFW creations.
Recommended Settings:
Sampler: LCM
Schedule Type: Exponential
Sampling steps: 8 (or 12 for extra detail)
CFG: 1.5
⚡ These settings are not the same as my regular SDXL workflow — be sure to adjust your pipeline!
Hires.fix Settings:
ou can try using fewer Hires steps than your main Sampling steps.
Denoising Strength: Recommended range is 0.2 – 0.5.
Hires CFG Scale: You can keep it the same as your main CFG scale, or set it to 1.
CyberRealistic XL Catalyst is where SDXL gets weird in the best possible way. This isn’t some spotless showroom model. Think of it more like a wild test garage. Sparks fly, bolts get overtightened, and suddenly something totally unexpected comes roaring to life. It’s built on the same bones as CyberRealistic XL, but this version is packed with experimental tweaks. Catalyst is for anyone who actually likes a little chaos mixed in with their clarity.
Here, realism crashes right into raw unpredictability. Sometimes it’s brilliant, sometimes it’s just plain bizarre, and every now and then, it’s so off it actually loops back around to being awesome. Welcome to the lab.
🧪 What’s New in Catalyst?
Realism is tuned in weird ways (it’s a bit less censored, and sometimes you’ll wonder what just happened)
Always changing, always mutating (no changelog, just new vibes and glitches every update)
Output can get wild, but when it nails it, it really nails it
⚙️ Recommended Settings
Sampling method: DPM++ 2M SDE Karras
VAE: is already Baked In
Sampling steps: 30+ Steps
Resolution: 832x1216 / 896x1152
CFG: 3-5
Clip Skip: 1
Upscale: 1.5 upscale (0.3 - 0.35 denoise)
Upscalers: 4x_NickelbackFS_72000_G / 4x_foolhardy_Remacri / 4x_NMKD-Siax_200k🧠 Prompting
Example negative prompt:
cartoon, illustration, anime, painting, CGI, 3D render, low quality, watermark, logo, labelor use CyberRealistic Negative
🔧 ADetailer (Optional)
Adetailer model: face_yolov9c.pt
If you only want the main face being refined set 'Mask only the top k largest' to 1.⚠️ Keep It Ethical
Yes, Catalyst can be uncensored. That’s not a free pass — it’s a responsibility. Don’t misuse it.
Description
A special DMD2 edition built for faster, ultra-realistic results with fewer steps - perfect for both SFW and NSFW creations.
An evolution of CyberRealistic V7.0, rebuilt from the ground up.
Recommended Settings:
Sampler: LCM
Schedule Type: Exponential
Steps: 8 (or 12 for extra detail)
CFG: 1.5
⚡ These settings are different from standard SDXL workflows – adjust for best results!
FAQ
Comments (43)
DMD2 has been always the superior format for XL. Better aesthetics, better anatomy, more realistic, fast, best in everything. Sorry, somebody had to say it. DMD2 is the future. It always has been.
I'm glad Cyberdelia has seen the light of the dmd2 format, finally.
Except some models that incorporated it went pretty dumb, losing a lot of prompt follow (altho one could argue, some of those didnt have it to start with, it just became even worse).
Its good thing, but entirely depends how well it was used. Cant tell, only testing or author can say that.
@Mescalamba prompt following is not related to dmd2 sorry. There are a lot of dmd2 checkpoints with good prompt adherence. It helps to use Rouwei text encoders, too.
@Mescalamba @ElectricDreams The DMD2 version was developed from the FP32 version of CyberRealistic XL Catalyst 2.0 (available for download via Patreon). It took some experimenting, especially with the settings you can see in the sample images. Of course, there are both pros and cons - and the standard CyberRealistic versions will always remain non-DMD2. Luckily, Catalyst gives you the freedom to experiment with this.
In any case, i recommend using 12 steps always with dmd2 checkpoints. Using 8 steps in a cyberrealistic checkpoint is a disservice. I use 12 steps by default and sometimes 14 steps for complex compositions. The difference in image quality between 8 and 12 is astronomical and it only takes a few seconds more.
@ElectricDreams true, I have some images with 12 steps. Not always better but just experiment
@Mescalamba
With dmd every steps and cfg matter. Between 1cfg 8steps and 1.5cfg 15 steps there are extreme huge difference. Everybody needs to find, what he likes the most.
If somebody just want photoreal nudes 1-8 is the best. More cfg, more adherence and less photolook. Somewhere inbetween(1.25) is the truth.
You are right about prompt following, but it depends on cfg, not dmd.
@salmanfulton696 I also sometime use steps 8 - and steps 4 for upscale with lower CFG. It's indeed really experimenting and because it's fast - that is not a issue.
@Cyberdelia What upscaling method do you use? Did you find any good settings for hiresfix(if you use it)?
@salmanfulton696 Upscaler: 4x_NickelbackFS_72000_G (this is just my personal preference). I usually do a 1.5x upscale, but you’ll need to experiment with lower CFG and step counts or just keep them the same, depending on your workflow.
@Cyberdelia I meant highres cfg, steps, and samplers, schedulers. With dmd i only found euler a karras with 1 cfg and 4 steps max acceptable. Lcm is not good for upscaling, more steps with euler a mess with skin very badly.
@salmanfulton696 It entirely depends how DMD2 is merged in model. But Im pretty sure thats above scope of everyone here apart Cyberdelia. And there Im sure, she is aware.
CFG, its method, steps, sigmas.. even noise. All-so-many variables have impact on results for any model. That wasnt my point.
@Mescalamba @salmanfulton696 I have added a Forge screenshot :)
@Mescalamba Yep, lots of variables. Since the CFG is low, the settings have a big impact on everything.
@Cyberdelia Are you girl? Thats totally new info for me:)
@ayanfinbarr286 I identify as "whatever keeps the conversation interesting" 😂
@ayanfinbarr286 you didn't know? and yes! she's gorgeous too!
@ElectricDreams Sincere doubt. Not saying its not possible, just unlikely.
@Mescalamba Oh, sweetie… are you really questioning my gorgeousness? Mmm, risky move, darling. Because if I have to explain it, you couldn’t handle the syllabus, let alone the final exam.
@Cyberdelia Its my nature to question everything and believe nothing. Apart that, its like that old saying. "Guys online are girls, girls online are guys, except little girls, those are FBI agents."
@Mescalamba Ha, seems like your instincts didn’t fail you this time.
Trust them... they’ll serve you well… especially around me ;)
@Mescalamba Cyberdelia, besides being one of the most gifted creators in this platform, is a fine dame. But i really don't think this is important. We should rate her works intead of her person. The genius behind her works. It's all that matters in the end: her legacy. If people wants to think she is a him, it's totally irrelevant. I don't think this is important at all!
@ElectricDreams At the end of the day, I’m just a dude here to create a little chaos :) still not know where this rumour started
Did you mean LCM instead of LMS? I tested it and LMS produces broken results, while LCM works as expected.
Oops, it needs to be LCM. Thanks!!
I even got this error on my Patreon.... stupid stupid :)
@Cyberdelia More like brilliant, brilliant. This model is fantastic. In a world filled with acronyms, we're bound to get a few mixed up. But most people using DMD2 style models already know LCM is the way to go, so no worries!
I don't want to oversell this checkpoint, but Cyberrealistic XL Catalyst DMD2 has changed my life.
I'm really impressed with how well this DMD2 model handles tag adherence. The results feel like a whole new level compared to the regular SDXL models. I might even consider switching to this one over the others I currently use. Amazing work!
Quick question: what are the recommended negative prompts or embeddings to use here? It seems there’s no real difference between specifying something versus leaving the negative prompt empty.
I keep it simple: " lowres, bad anatomy, bad hands, ugly, deformed".
I had 'Ignore negative prompt during early sampling' set to 0.25 (Forge) and that's for this checkpoint not really working. So I have disabled it.
This model is fast, detailed, good prompt adherence and plays well with LoRA's. What more can you ask for with a model that cranks out gorgeous images in 12 steps or less.
i noticed this also. I use 8-12 steps, and it gives me superior to models that require 20-30+ steps. Not sure how this is pulled off, but it makes using the less efficient models feel weird when you can make an image in 3-4x less time with this one, and often it looks better.
I dont know, if its intended, but according to user pics, people dont understand, how this works.
I know, right? It’s not like the instructions aren’t clear.
It’s the same with my TI models that don’t work in the Civitai generator. I’ve put up big warnings, but people still use them that way.
I love your Checkpoint Model!!!
Hi, first fantastic work! This might be the best model I ever saw.
I have a question, I use your CyberRealistic Workflow with your DMD2 Version and used the values you provided in the description for Input and Upscale, but there are still some values I don't know which to put in UltimateSD Upscale, the Face and Hand detailer?(steps, cfg, sampler,..)
Every preview looks fine until the end result which looks blurry and low definition, eyes don't look good. :(
Thank you for your work!
"Here, realism turns into gross unpredictability." In fact, it's brilliant, I thank the author for an excellent tool for creativity.
Definitely the best model I ever saw.
what is a release candidate version ?
what differences with "normal" version?
Do you mean the difference between this model and, for example, CyberRealistic XL/Desire?
This is/was more of an experimental version, but many of the same elements are already included in those other models.
Or did you mean something else?
@Cyberdelia no, juste release candidate version, what is it ? is it a special version ?
@Suzanne Release Candidate (RC) is a potentially the final version, released for testing before the official stable launch (CyberRealistic XL).
Think of it as "this is probably it unless someone finds a showstopper bug." It's further along than a beta: feature-complete, with no known critical issues, just waiting for final validation.
@Cyberdelia ah, ok thanks
Details
Files
cyberrealisticXL_catalystXLV20DMD2.safetensors
Mirrors
cyberrealisticXL_catalystXLV20DMD2.safetensors
cyberrealisticXL_catalystXLV20DMD2.safetensors
cyberrealisticXL_catalystXLV20DMD2.safetensors
CyberRealisticXL_CatalystXL_V2.0_DMD2.safetensors
SDXL-cyberrealisticXL_catalystXLV20DMD2.safetensors
cyberrealisticXL_catalystXLV20DMD2.safetensors
cyberrealisticXL_catalystXLV20DMD2.safetensors
cyberrealisticXL_catalystXLV20DMD2.safetensors