Krealism 📷

A realism finetune of Krea 2 Turbo, built for ComfyUI. Film look, natural color, candid framing. No LoRAs needed.

---

✋ Which file do I grab?

V3.1 (latest):

File	Size	Quality (vs bf16)	Speed (RTX 5090, 1 MP)	Pick it if
Krealism_V3.1_bf16.safetensors	26.3 GB	the original	0.59 s/step	you have 32 GB of VRAM and want the reference
Krealism_V3.1_int8_convrot.safetensors	14.1 GB	0.007: nearly identical	0.30 s/step	💡 most people. bf16 quality at half the size
Krealism_V3.1_mxfp8.safetensors 🆕	14.5 GB	0.019: very close	0.33 s/step	RTX 50 series. fp8 quality, faster
Krealism_V3.1_w4a8.safetensors 🆕	8.8 GB	0.031: close	0.33 s/step	smaller cards. nvfp4 size, better than nvfp4
Krealism_V3.1_nvfp4.safetensors	8.8 GB	0.059: same person, details change	0.26 s/step	the fastest file (RTX 50 series runs nvfp4 natively)
Krealism_V3.1_fp8.safetensors 🔧	14.1 GB	0.016: very close	0.46 s/step	your setup runs fp8 but not int8 or mxfp8

How to read the quality score:

• LPIPS vs the bf16, one step from the same starting point, averaged over 3 prompts at 3 points of an 8 step render.
• 0 = identical. Under 0.01 you can't tell them apart, 0.01 - 0.03 is very close, 0.03 - 0.06 small details change, above 0.06 you can see it.

The two new files:

• mxfp8: 8-bit with a scale for every 32 values instead of one per layer. Fast on RTX 50 series; older cards load it, but get no speed gain.
• w4a8: 4-bit weights, 8-bit math. Half the size of the 8-bit files, and on an RTX 5090 just as fast as them. Fast on RTX 30 series and newer.
• ⚠️ Both need an up-to-date ComfyUI (with Comfy Kitchen). Older installs won't load them.
• On an 8 GB RTX 5060 laptop, both run about as fast as int8 (2.6 - 2.8 s/step at 1 MP) and twice as fast as bf16 (5.8 s/step).

V1.0 (older): Krealism_V1.0 (bf16), V1.0-fp8, V1.0_int8_convrot and krealism_v10Turbo-q8_0.gguf are still here if your workflows depend on them.

📂 All of them go in:

ComfyUI/
└── models/
    └── diffusion_models/

---

Settings

The same ones ComfyUI's Krea 2 template uses. Nothing fancy:

• Steps: 10
• cfg: 1.0
• Sampler: Euler
• Scheduler: Simple
• VAE: qwen_image_vae
• No shift node, no LoRAs

---

Instagram: synth.studio.models

Buy me a ☕ HERE

This represents many hours of work. If you enjoy it, please 👍 like, 💬 comment, and feel free to ⚡ tip 😉

"True Nothing is. Permitted Everything is"- Yoda Auditore, Assassin's Wars