Krea 2 identity edit makes outputs look more "AI": reference image kills skin realism

ebubu03 · reddit · 2026-10-17

A ComfyUI user generating fictional characters with Krea 2 Turbo (fp8scaled, 8 steps, euler/simple, cfg 1) noticed that without a reference image the same prompt produces realistic skin (pores, freckles, phone-camera texture), but as soon as a face reference is added via the identity edit workflow, identity improves while skin becomes smoother, glossier and more "AI-looking" — even when the reference itself has detailed natural skin.

Setup: LoRA krea2identityeditv12 at 1.0, Krea2EditModelPatch (refboost 1.0, refboosta 1.0, fitmode "fit"), Krea2EditGroundedEncode (groundingpx 768), with realism/snapshot LoRAs stacked.

Current workaround: a second pass (latent upscale x1.5 + 0.80 denoise) restores realistic skin but drifts the face away from the reference identity unless the reference stays connected during that pass too. Open questions: whether the smoothing is expected from a LoRA trained on clean/synthetic pairs, whether lower refboost/LoRA strength or applying the reference only in early steps can keep identity without the AI look, and better nodes/workflows for "same face + keep realism LoRA texture".

Original post →

More from Multimodal

Multimodal channel →