# OurSecret v1 (Krea 2 Foundation Merge)
Hi everyone,
This is my very first model checkpoint based on Krea 2 (Turbo) ! I spent a lot of time fine-tuning this blend to get the aesthetic and skin texture right, but since it's my first public release, please be gentle with me. I'm really looking forward to the community's feedback, stress tests, and reviews to see what works best and what we can improve together for the next versions.
### What is OurSecret?
OurSecret is designed as a versatile, true all-rounder foundation model. 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:
- 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 RTX 5000 series cards.
- Q4_0 GGUF: 6.7 GB lightweight version for 6GB to 8GB VRAM setups, running smoothly without noticeable artifacts.
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!
Description
Initial public release of OurSecret v1.0. Master foundation merge focused on realistic skin textures, high-fashion editorial styling, and accurate prompt comprehension in 8-12 steps. Available in BF16, FP8, INT8, and Q4_0 GGUF.
FAQ
Comments (9)
This is a really good secret.
Thanks you !
Is the int8 convRot?
Hey! No, it doesn't use matrix rotations (like QuaRot).
It's standard native ComfyUI int8_tensorwise—only the linear weights are in INT8, while all norms and embeddings are kept in original BF16. It loads right out of the box in the standard ComfyUI loader with no extra setup needed, and runs super smooth (especially on RTX 4000/5000 cards)!
@Guillaume127 Thanks for the response. The reason I ask is that apparantly in8 is inferior output to normal fp8 w/o convrot, but superior to regular fp8 with convrot. So convrot is the decider for me as to whether int8 or fp8. Maybe you sidestepped the degradation with int8 by keeping the norms and embeddings in bf16?
@subterranian Yes ! Keeping the norms and embeddings in BF16 (with per-tensor scaling) indeed prevents the usual INT8 degradation.
Still, if your GPU handles it fine, Fp8 is the safest bet for pure 1:1 fidelity. INT8 is mainly there as a speed boost for RTX 5000 users.
I only have an RTX 2060, and so far all INT8 conversions perform roughly the same; yours is slower.
I'll look on my side to see if I can better optimize the conversion, I didn't have this problem on my computer. Thank you for your feedback, I can improve the INT8 version.
@pikkby476 Hey ! After running tests, pushing quantization any lower than 4-bit (like 2 or 3-bit) degrades image quality way too much on this 8-step architecture, so Q4 is the physical lower limit.
The INT8 file is still ~12 GB and keeps sensitive layers in BF16—which isn't natively accelerated on Turing GPUs (RTX 2000 series). On a 6 GB card, this forces massive PCIe swapping to system RAM on top of precision emulation, which explains why it's crawling.
I strongly recommend grabbing the OurSecret_v1_Q4_0.gguf file (~6.7 GB) from the download list instead! It's much lighter, fits way better on a 6 GB card, and generation times will be significantly faster (just swap your model loader node with the "Unet Loader (GGUF)" node in ComfyUI).
Thanks a lot for the feedback and happy generating !