# OurSecret v2 (Krea 2 Foundation)
Hi everyone,
Following the amazing support and feedback on v1.0, here is OurSecret v2.0 !
For this update, I focused on refining what could be improved based on Output and community : better anatomical consistency in complex poses, sharper prompt adherence, and cleaner background details, while keeping that natural skin texture and high-fashion lighting.
Thank you again to everyone who tested, reviewed, and posted images with the first version !
### What is OurSecret?
OurSecret is designed as a versatile, true all-rounder foundation model based on Krea 2 (Turbo). It handles high-fashion portraits, cinematic storytelling, urban architecture, and tactile macro objects with equal ease.
And for those wondering: yes, the model can generate detailed NSFW content cleanly if you prompt for it.
### How to use in ComfyUI:
You can run this model directly using the official ComfyUI Krea 2 Turbo workflow:
1. Place the downloaded model file into your directory: ComfyUI/models/diffusion_models/
2. Select it inside the diffusion loader node.
(Note: The GGUF version requires a dedicated GGUF loader node in ComfyUI, such as "Unet Loader (GGUF)", to be used properly).
### Recommended Settings:
- Sampler: er_sde (recommended for organic filmic grain) or euler
- Scheduler: simple
- Steps: 8 to 12 steps
- CFG: 1.0 (keep it strictly at 1.0)
### Available Files (v2.0):
- BF16: The complete, uncompressed 23.8 GB master archive for maximum archival precision.
- FP8 (float8_e4m3fn): The 11.9 GB daily driver version with uncompromised quality for 12GB+ GPUs.
- INT8 TensorWise: 11.9 GB, specifically optimized for hardware matrix math on modern RTX cards.
- Q4_0 GGUF: 6.7 GB lightweight version for 6GB to 8GB VRAM setups, running smoothly without noticeable artifacts.
- NVFP4 (Experimental) - 6.7 GB) quality is slightly lower compared to the GGUF but generation speed is noticeably faster.
All showcase preview images and presentation visuals were generated directly on the FP8 version.
Feel free to post your generations and settings in the gallery below. Looking forward to your thoughts on v2!
Description
### What's New in OurSecret v2.0
OurSecret v2 is a major leap forward over v1.0, engineered for higher photographic realism, cleaner anatomy, and rock-solid generation stability.
*Enhanced Prompt Adherence: Significantly improved comprehension of complex prompts, nuanced lighting, and specific styling details.
*Superior Anatomical Coherence: Greatly improved body geometry and limb stability in challenging poses.
*Artifact & Glitch Reduction: Cleaned up unwanted background clutter, ghost elements, and visual noise for much cleaner, publication-grade outputs.
*Next-Level Realism & Skin Details: More organic skin micro-textures, natural eye reflections, refined specular highlights, and richer cinematic lighting depth.
*Overall Aesthetic Upgrade: A more refined, high-fashion editorial look straight out of the box.
Fully compatible with the Krea 2 ecosystem.
### 📦 Available Versions & Which One to Download:
To make sure everyone can enjoy OurSecret v2 regardless of their GPU hardware, 4 optimized formats are provided :
*BF16 (23.88 GB) — Full Master Precision:
The pure, uncompressed original weights. Perfect for high-end cards (RTX 3090, 4090 / 24GB+ VRAM) or cloud GPUs for 100% loss-free quality.
*FP8 (11.94 GB) — Recommended for Most Users:
The sweet spot between speed, quality, and memory. Visually identical to BF16 while cutting VRAM usage in half. Ideal for 12GB to 16GB VRAM GPUs.
*INT8 (11.94 GB) — ComfyUI TensorWise Quant:
Optimized quantized weights for native ComfyUI loaders. Ultra-stable and light on memory.
*Q4_0 GGUF (6.72 GB) — Low VRAM & Fast:
Ultra-compact build for 8GB–10GB GPUs or laptops using the ComfyUI-GGUF node. Fast download, minimal VRAM footprint, and remarkably sharp quality.
Quick tip : If you're not sure which one to choose, grab the *FP8** version !
*NVFP4 (Experimental / Speed Edition - 6.7 GB)
• Format : Native ComfyUI NVFP4 (E2M1 microscaled with E4M3 block scales). • Best for : Maximum generation speed & testing on RTX 40 & 50 series GPUs with 8GB VRAM. •
⚠️ Quality Note : Due to 4-bit float limitations, micro-skin textures are softer than BF16/FP8. • For best quality on 8GB/12GB GPUs : Use the GGUF Q4 version. • For ultimate photorealism & master textures : Use BF16 or FP8.
---
A quick word to the community :
A huge thank you to everyone who supported v1.0 and gave feedback! Your enthusiasm and positive reviews keep this project alive.
Feel free to post your generations in the gallery, leave a rating, and share your prompt, I can't wait to see what you create with v2!
FAQ
Comments (10)
I think v2 is a little overly horny compared to v1, but they're both quite capable! The horny quality does mean it distorts and biases the anatomy sometimes though, which is not ideal. But I hope this will continue to be developed, it's an excellent model.
@hushicho Hello, thank you for your feedback ! Which prompt caused you to get overly "horny" results ? I've tried to reduce this aspect in general, but it might not be quite right yet. I'll keep this in mind for v3 and continue to improve the template. Thanks again for your feedback ! ❤️💪
@Guillaume127 Any nudity is usually very horny! Male figures always have massive genitals and they are usually erect. I haven't figured out any way to try to balance that out. Most Krea2 checkpoints make male figures with giant genitalia, even v1's genitals are too large regularly. But v2 is much more frequent. The usual Krea2 output is a gigantic circumcised erection, everything else is a struggle.
Is it just me or are some of these krea 2 models difficult to get a full body and standing, one of the past techniques was to describe say shoes, or their legs that would help the camera zoom out a bit, but when i do that here I end up with the person kneeling or sitting such that their feet and their face are nearer to each other.
Any suggetstions?
You're definitely thinking about this in the right way - prompt around your subject to pull the camera back. Have them stand on a rug, or describe the flooring underneath them. Maybe there's a chandelier above them, or some art on the wall. The other part of the equation is the image dimensions. Make sure your width and height support the pose you are trying to create. If you're trying to get a standing pose on a 2048x768 image, you're not going to have a good time. As far as camera prompting goes, use cinematographic language to create wide shots. Play around with focal length and f-stops as well.
@wooba001466 @ornery54321 Hi I think Ornery has fully answered the question; I’ll make a note of it for the next version of the model, though, to see if this is a recurring issue. Thanks for your feedback, and enjoy generating !
great work,i like it .any idea for the nvfp4 or int4 convort.
@civitai000000 Thanks a lot for the feedback ! Just added the NVFP4 version to the download list. Keep in mind that quality is slightly lower compared to the GGUF version (due to 4-bit float limitations), but generation speed is noticeably faster. Enjoy !
@Guillaume127 tks.report a flaw,when it generates like "an obese man who wears a white tank top",the nipples sometimes appear or mix with the clothes.hope a fix.
You’ve really taken 100 steps forward with v2, but you’re still behind compared to certain models. v2 brings a very convincing level of realism (8.5/10), but I would focus more on prompt adherence (6.5/10), which isn’t convincing at the moment. Many sex positions are "bizarre", the semen looks like a torrent of water, and the anatomy sometimes does its own thing. That being said, your effort really shows—I truly appreciated it, and I hope a potential v3 fixes the flaws that v2 carries along. Keep up the good work!