📋 Description
A merge based on KREA2 Turbo, optimized for fast and high-quality generation. This model combines the strengths of the KREA2 architecture with custom tuning for improved results across various styles.
Three versions have been posted: SAFE (clean art), BASE (nudity allowed), and NSFW (practically anything goes).⚙️ Recommended Generation Settings
Parameter
Value
Steps
8
Sampler
Euler a
Schedule type
Bong Tangent
CFG scale
1
Seed
3448636919
Resolution
1536×2048
Model
fascium_SFWKREA2_fp8
Model hash
98d064dc8f
Module 1
qwen_image_vae
Module 2
qwen3vl_4b_fp8_scaled
Diffusion in Low Bits
Automatic (fp16 LoRA)
RNG
CPU
Style Selector
Enabled (Style: base)
Krea2 Enhancer
✅ Enabled
💡 Tips & Notes
Low CFG (1.0) — This model is designed to work best with a very low CFG scale. Higher values may cause artifacts or oversaturation.
Few steps needed — Only 8 steps are required thanks to the Turbo architecture. Going higher won't improve quality and will just waste time.
Krea2 Enhancer — Make sure to enable the Krea2 Enhancer for best results.
Style Selector — Set to
baseby default. You can experiment with other styles, but disable randomization for consistent output.Resolution — Tested at 1536×2048 (portrait). The model supports various aspect ratios, but keep total pixel count around ~3MP for optimal performance.
fp8 quantized — This is an fp8 version for reduced VRAM usage while maintaining quality.
🖼️ Trigger Words / Prompt Guide
No specific trigger words required. Write natural, descriptive prompts. The model responds well to detailed style and composition descriptions.
⚠️ Known Limitations
SFW-focused merge
Best results with the exact settings above — deviating from recommended CFG/sampler may degrade output quality
Feel free to adjust the model name, add your own example images section, or tweak any details to better match your vision! 🎨
Description
work in progress, version 16/07/2026
FAQ
Comments (6)
What exactly is this a merge of? What was merged into the base model? A bunch of LORAs or what?
LORAs, Mostly my own. Far from all of them have been published.
The preview images are mostly nude women? Is that like 80% or more of the training data?
@ss9999 Bruh it clearly says the SAFE version is for those who don't want naked women, so why complain, just get that. The author is a great person for catering to everyone, most do not publish various versions.
@ferrrett33 It's not a complaint. The issue is that for those of us who aren't interested in accumulating a lot of mostly female and all-female content, it's helpful if creators put in the notes what percentage of attractive females vs. attractive males there are. It's not too much to ask at all, especially with checkpoints. I do have a complaint, though. I don't appreciate people misconstruing this polite reasonable request as a complaint or otherwise out of line. What is out of line is wasting people's time. Not everyone wants models that are 70%, 80%, 90%, or 100% female. That has nothing to do with so-called SFW vs NSFW. If people can't manage to spare the time to type a few characters (e.g. 80%AF, 20%AM) in the notes, then they could try to post more young attractive males in their example images. But, the percentages are really better because one could still produce example images with males in them despite a lora or checkpoint being 90% female. If creators were doing the opposite (posting what are labeled as general loras/checkpoints but only giving male examples and having 90% males in the training data without noting that anywhere) I can guarantee that there would be complaints, too. Imagine if anime would be 90% of the training data in what is demonstrated as a general realism lora or checkpoint? That's the situation people like me are frequently in when we don't want all-female or mostly female things.
@ss9999 That’s an interesting take. But there’s one nuance: I create these models for myself and just upload them as a place to store them—if people like them, great; if not, that’s fine too. The model description is detailed. My own interests in generation reflect the model's capabilities only partially.