# Bonkaiii SemiReal Skin — Krea 2 LoRA

Anime features with realistic skin and light. Softens Krea 2's default
photographic rendering into a semi-real look — smooth even skin, clean
highlights, anime-leaning faces — without flattening it into cel shading.

---

## ⚡ Settings

**No trigger word.** The style lives in the adapter. Load it and it works.

| | |
|---|---|
| **Weight** | **`STLLL TESTING`** — start around 0.5 and adjust |
| **Base** | Krea 2 (built on Krea 2 Raw, works on Turbo checkpoints) |
| **Sampler** | Euler (flow-matching) · scheduler linear |
| **Resolution** | 832x1216, 1216x832 or 1024x1024 |

**Stacks fine with other style LoRAs.** It was built to be dialled in, not run at
full strength — it changes rendering, not content, so it composes well.

---

## How it was made

A distillation LoRA. The look comes from **Bonkaiii SemiRealAnime v2**, an
Illustrious/SDXL merge, and this LoRA teaches that rendering to Krea 2.

| | |
|---|---|
| Dataset | 196 images, hand-culled from 503 candidates |
| Source | generated on Bonkaiii SemiRealAnime v2, then upscaled |
| Resolution | 2080x3040 originals, downsampled to 832x1216 |
| Rank / alpha | 32 / 32 |
| Learning rate | 1e-4, batch 1, 2000 steps |
| Captions | subject, pose and setting only — never the rendering |

Every training image is a 2.5x downsample of a 6.3 MP upscale rather than a
native 1 MP render. Six source pixels average into each output pixel, so the skin
texture and edges the LoRA learns from are supersampled rather than generated.

The dataset was deliberately varied in everything except the rendering — 196
distinct prompts, 69 hair descriptions, 12 settings, four skin tones, no repeated
character. A style LoRA learns whatever is invariant across its data, so the
rendering had to be the only constant.

---

## Credits

- **Bonkaiii SemiRealAnime v2** — the source of the look
- LoRAs active when the training images were generated, and therefore part of
  the style this distils: **prtyface v1**, **PowerPuffMix**, **ponyv6 noobV1.2**,
  **rimixO**

The result is a composite of that model and those LoRAs, not any one of them.

---

## Known limits

- **Hands** — Krea 2's general weakness; a style LoRA does not fix it.
- **At weight 1.0** the effect may be stronger than you want. This was built for
  0.4-0.6 territory.
- **Not a character or detail LoRA.** It changes how surfaces render. It will not
  add detail that the base model did not already produce.