This is the conversion of RawGirl - Krea2 v1.0 by Kyphix0. All credit goes to him (thanks for the amazing model, mate!)
✨ What it does
I suggest you take a look at the original model's page to get an idea what it is about in terms of overall style and purpose.
⚙️ My settings:
First pass: 10 steps - cfg 1.0
Face pass: 5 steps - cfg 1.0 - denoise 0.32
Upscale (tiled): 4 steps - cfg 1.0 - denoise 0.21
Face pass: 5 steps - cfg 1.0 - denoise 0.25
I encourage you to play around and try to find your best settings!
🎛️ samplers / schedulers recommendation
From my testing I find that these combinations of samplers / schedulers produce good images:
⭐⭐⭐⭐ dpmpp_2s_a* / beta
⭐⭐⭐ er_sde / beta
⭐⭐⭐ euler / simple
⭐⭐ dpmpp_2m_sde / beta
⭐ euler_a / ddim_u
⭐ sa_solver_pece* / ddim_u
*Be aware that dpmpp_2s_a and sa_solver_pece both are two-step samplers and hence take twice as long to generate (sometimes this is really worth it though).
Description
v1.0 INT8-ConvRot
FAQ
Comments (7)
Thanks for the model and workflow.
How can I reduce image smoothing after upscaling if I want to preserve a little more detail?
Is it possible to run prompt creation separately in this workflow without running the rest of the process?
Thanks for sharing this great model!
conversion based on bf16 ?
no, I used the native fp8. I don't think it's likely you will get bf16, according to Kyphix on his Discord :(
@fhaifhai yeah, now I get why it looks the way it does :(
I was hoping he’d make the INT8 version directly from the BF16 model.
Quantizing the already-quantized FP8 version means you’re compounding approximation error instead of starting from the clean source — a bit like recompressing an already compressed image.







