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
This isn’t just another version of CyberRealistic XL. Desire is a special edition, fine-tuned for artists who want to create stunning NSFW (and SFW) content with speed, realism, and flexibility.
So why go with Desire?
While the original CyberRealistic XL aims for pure realism across all types of work, Desire turns up the creativity. It’s designed for those moments when you want more freedom, more play, and a smoother workflow.
Desire captures the spirit of exploration. Whether you're pushing boundaries or just having fun with your art, it’s here to help you move fast and create without limits.
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_200kRecommended Settings DMD2
Sampler: LCM
Schedule Type: Exponential
Sampling Steps: 8/12 (12 for extra detail)
CFG Scale: 1-1.5
⚡ These settings are not the same as the standard SDXL workflow — be sure to adjust your pipeline!Description
V2.0 fixes a ton of the underlying issues, making prompting smoother, anatomy more consistent, and NSFW output a lot more reliable overall.
This version’s also more permissive and predictable, with much better proportions and posing - especially for tricky erotic setups. Overall it's a LOT more NSFW.
Most of the sample images were generated using DMD2 but there are also samples made without DMD2 so you can clearly see how the model behaves on its own.
Personally, I use two custom DMD2 LoRAs (dmd2_CyberRealistic_part1 and part2) with different weights. They're not public (yet?), but for now, just stick with the standard DMD2 LoRA or run the model without it. Both work just fine.
FAQ
Comments (15)
Tbh it sometimes feels like this model doesn't care about what I write in my prompt. "Woman sunbathing naked" model decides to throw in cowgirl scene for good measure lol.
The phrase "woman sunbathing naked" can technically work in SDXL, since the model does have some internal understanding of what "sunbathing" implies. But here's the catch: the results often depend heavily on the rest of the prompt.
If the supporting details aren't helping to visually anchor the scene, like pose, environment, lighting, or setting - the model may just default to something vague, generic, or fully ignore the “sunbathing” part.
Think of it like this: SDXL knows the word, but it doesn't always act on it unless it's backed up with clear visual cues. So if a prompt seems like it's not working, it’s often not about the main subject, but how it's framed around it.
@Cyberdelia True that, That is a good catch ✌🏽😉
Amazing.
Excellent model. I was running some analysis to compare the CLIP models embedded in different SDXL checkpoints from CivitAI and I found a corruption on layer 11 of the CLIP-L model in V2.0, where several tensors have infinite values. Not sure if this is relevant for the final quality of the generated images, though. This is something I haven't noticed on other checkpoints, anyway.
You can check it with the following Python code provided you have PyTorch installed
import torch sd = load_file("cyberrealisticXL_desireV20_clip_l.safetensors") for k, v in sd.items(): if not torch.isfinite(v.float()).all(): n_bad = (~torch.isfinite(v.float())).sum().item() print(f"{k}: {n_bad} non-finite values — shape {v.shape}")
If you run this code it shows the following
Pardon my ignorance. I'm new at this.
I load the workflow in ComfyUI from one of the images. But hit a snag.
Where do I get dmd2_CyberRealistic_part1 and dmd2_CyberRealistic_part2? I can't find them. I'm very confused, how do I get this to work?
Just use standard DMD2 lora
@Cyberdelia Ok. So I have dmd2_sdxl_4step_lora.safetensors installed. Just use that for both nodes?
@Reterg I'm sorry. I don't use ComfyUI so not an expert on this. I guess you need to load it as LoRa in your workflow
@Cyberdelia My bad. I loaded an image from the page, which loads the workflow in ComfyUI and it had two Lora Nodes at 0.5 each, totalling it to 1. I loaded one of the other pics and it had 1 Lora node at 1.0.
What do you use if not ComfyUI, just out of curiosity?
@Reterg I'm a Forge Neo addict ;)
@Cyberdelia I'll check it out. ComfyUI is a bit annoying sometimes and it's been hard to work out. The workflow loaded from the image, makes a blank screen image. 🤷♂️
@Cyberdelia Corretomundo. LoRAs are loaded via the "Load LoRA" or "Load LoRA (Model and Clip", separetly from the checkpoint.
@Reterg For LoRAs the path is super easy. Add the node "Load LoRA (Clip and model)". Checkpoint model and VAE as input, clip for prompts and model for ksampler as output. Make sure LoRAs are in the proper directory (\ComfyUI\models\loras).
I also recommend installing this to index LoRA, models and general management (if you use many): https://github.com/willmiao/ComfyUI-Lora-Manager
Like having a mini civitai on your setup =)
LoRAs can acquire much more complex roles; some people love them, others don't. I don't use them personally, proper prompting (avoiding negatives as much as posible) combined with great models -like those of Cyberdelia's- make WONDERS. The things I've created with just prompting, fine tuning ksampler and using the right model tuned -> blew my own mind more than once. A great time to be alive.
@gershu Yeah it's been a while since that post and I got it now. Thanks though