https://civitai.red/models/2786499/realism-by-stable-yogi-krea2?modelVersionId=3329215 Realism by Stable Yogi Krea2 V3.0 Extended Civ INT8-ConvRot C85.safetensors SHA256: 66B392C27A About this version Real skin, a working range, and seven quants to run it. V1.5 through V2.5 were built for speed — sharp, fast, locked to one setting. **V3.0 Extended is the first in this line that gives you room to work.** It holds together across a real range, so you can direct the shot instead of taking what you are given, and it still runs fast when you are in a hurry. ⚙️ Settings - Fast: 8 steps · CFG 1 — drafts and batches. - Best: 12 steps · CFG 1.5 — the richest, most controlled version. - Range: 8–20 steps · CFG 1–2.5. ⚠️ Do not go above CFG 2.5. - Euler · Simple · native 1024²+. ✨ What is new in V3.0 - Real skin, not smoothed skin. Trained on photographs, not rendered from another model — pores, texture and light behave the way they do on a real face, and it holds up at 1:1. - Extended range. CFG 1–3 and 8–20 steps, instead of one fixed setting. Lower CFG for soft and natural, higher for punch. Fewer steps for speed, more for polish. - East Asian mode, on a word. Put *asian** in your prompt and the model moves to an East Asian face. Leave it out and nothing changes — it stays exactly the model you get without it. 📦 Downloads — seven formats | Q8_0 GGUF | 13.6 GB | Works on everything — NVIDIA, AMD, low VRAM. Pick this if unsure. | | fp8-scaled | 13.2 GB | ComfyUI-native. Runs on AMD too. | | INT8-ConvRot | 12.6 GB | NVIDIA only. Forge + ComfyUI native. | | NVFP4 | 8.2 GB | RTX 50 / Blackwell only. Smallest fast option on those cards. | | Q4_0 GGUF | 7.8 GB | Low VRAM, runs almost anywhere. | | INT4-ConvRot ⭐ NEW | 6.9 GB | NVIDIA RTX 20 and up — small and quick on older cards. | | Q2_K GGUF | 5.0 GB | 6 GB cards. Smallest. | (Model only — pair with the *Qwen3-VL text encoder + Qwen-Image VAE**, same as base Krea 2.)* Made by Stable Yogi 🧘 — drop a ❤️ and show me what you direct it to.