Septum Hoop Nose Ring (Krea2 Concept LoKR)

If you've spent any time prompting Krea2 for simple hoop septum piercings, you likely know the struggle. Natively, the base model almost always tries to generate horseshoe rings (the ones with the balls and gap), adds multiple extra rings, introduces unwanted facial jewelry, or just fails entirely. In my experience, generating a clean, continuous hoop septum ring natively has a success rate of about 0.001%.

I trained this concept LoKR specifically to fix that issue, giving you consistent, clean, and highly customizable hoop septum rings.

💡 Capabilities & Flexibility

• Trigger Word: SeptumHoopNoseRing

• Highly Flexible: Works beautifully across different art styles, aspect ratios, camera distances, and angles. Applies seamlessly to both male and female subjects.

• Customizable Styles: Fully supports descriptive prompting. You can specify materials and colors (e.g., gold, silver, black) and sizes ranging from small and thin, to large and thick, or even large and thin.

• Emergent Details: Despite being trained exclusively on simple seamless hoops, the model learned the placement and concept so well that you can successfully prompt for spiked septum rings or jewel-adorned septum rings.

• Bonus Feature: While it won't generate them unprompted, I’ve noticed this LoKR has the side-effect of noticeably enhancing the coherency of ear gauge plugs when explicitly included in your prompt.

⚙️ Recommended Usage & Settings

• Strength: The model holds up incredibly well at lower strengths. 0.45 - 0.75 is the sweet spot for a highly realistic look. You can push it to 1.0, though the piercing itself may take on a slightly "less realistic" appearance at full strength.

• Samplers: Tested and performs best using Euler Simple or ddim with a ddim_uniform scheduler.

• Cropped Head Workaround: If you are using this LoKR at full strength (1.0), it occasionally tries to crop the top of the head. Pro-tip: Briefly describe the subject's eyes and hair in your prompt to anchor the facial framing and easily prevent this.

• Using with Character LoRAs: This mixes very well with character LoRAs and doesn't bleed into or alter faces. However, the character LoRA will often fight for influence, so you will likely need to increase the weight of this LoKR to maintain the septum ring's coherency.

🛠️ Under the Hood (Training Details)

For the technical artists interested in the backend process:

• Dataset: 100 carefully curated images.

• Captions: Vision Language (VL) 1st pass, meticulously refined by hand.

• Steps/Resolution: 2000 steps at 768px (optimized for my current hardware limit).

• Trainer: FIzgig trainer utilizing the ultra-fast preset.

• Base Models: Trained on the FP8 Krea2 Raw model. Generations and testing were done in ComfyUI using the Krea2 Turbo default NVP4 model.

• Hardware: Trained locally on Windows 11, AMD ROCm 7900 XT (20GB VRAM), 32GB RAM, Ryzen 3900x CPU.

☕ Support & Future Plans

Right now, this model is locked behind paid Buzz access. I'm doing this temporarily just to help recoup the hardware and compute costs associated with testing and training, as soon as costs are recouped I'll change it to free download.

If you find this LoKR useful and want to support my work, tipping Yellow Buzz here on Civitai or supporting me on Patreon is hugely appreciated. While I don't currently have exclusive models on my Patreon, I am planning to introduce exclusive content there in the future.

For those familiar with my previous work across SDXL, Pony, Illustrious, and FLUX, I am currently in the process of re-training and releasing Krea2 versions of my entire back catalog of characters, styles, and concepts (with the exception of celebrity LoRAs, as I've permanently moved away from training on IRL people).

Commissions: I am currently open for custom LoRA / LoKR training commissions! Feel free to reach out if you have a specific concept, style, or character in mind.

Thank you for the support, and happy generating!