Krea 2 LoRA trained on 356 real portraits lets you prompt facial features directly
hell0nata · reddit · 2026-09-26
A fashion photographer tired of Krea2's cookie-cutter attractive faces trained a LoRA on 356 real snapshot portraits (174 men / 182 women, European, ages 20-40), captioning each with a structured multi-parameter face vocabulary covering face type, eyes, nose, lips, cheekbones, jaw, and chin.
- At strength 1.0 with no keywords, the default Krea face disappears and you get real human variation; adding feature tags reproduces them precisely, and tags stack well (e.g. long fleshy nose + hooded wide-set eyes)
- Author recommends strength 1.0-1.3, currently testing 1.3-1.5
- Eyebrow tags are buggy (too thick or missing) — avoid them
- The LoRA, "Realistic Custom Face", is on Civitai; feedback welcome on stacking with realism LoRAs
More from Multimodal
- Comment: China leads the AI video revolution while the West argued about AI art — TinfoilTricorn · 2026-11-11
- "QT your Geisha" meme takes over X with stunning porcelain-style AI portraits — LudovicCreator · 2026-11-11
- MiniMax H3-Image early access test: 4 images, one coherent cyberpunk world — Aiden_Tech_Ai · 2026-11-11
- MiniMax H3 Image generates a GTA6-style hurricane scene from a single prompt — socialwithaayan · 2026-11-11
- Dragon Ball reimagined as a 1971 movie: another hit AI retro trailer — techhalla · 2026-11-11
- ReCaVSR lands in ComfyUI: one-step Wan2.2 video upscaler hits 4K at 4.65 fps on an RTX 4090 — LatentSpacer · 2026-11-11