# soyokaze FLUX — Ukiyo-e Bijin-ga Style (FLUX.1-dev LoRA)

A style LoRA trained **exclusively on copyright-free materials**, now for **FLUX.1-dev**. It reproduces a classic ukiyo-e bijin-ga (beauty portrait) look — narrow eyes, small pupils, small mouth, pale skin — with traditional woodblock-print linework and coloring.

Because FLUX is a single general-purpose base, this one LoRA covers three looks by prompt alone:

1. **Full woodblock print** — default. Just describe the scene with tags.
2. **Photoreal × ukiyo-e** — start the prompt with `photorealistic portrait photograph of a japanese woman, soyokaze, ...`. Photographic skin and light are kept while the Edo-period hairstyle, makeup and styling come from the LoRA (works even at weight 1.0).
   **Note:** at weight 1.0 the photoreal mode is seed-dependent by design — some seeds come out as photographs, others flip into a full woodblock print. Both are intended, usable results; generate a small batch and pick. If you want fewer print-flips, lower the weight to 0.85–0.9 and/or raise Distilled CFG to 3.5–4.
3. **Anime prompts** get converted into the print style (or pass through untouched at low weight).

Sister models for SDXL/Illustrious: **soyokaze** and **soyokaze hybrid**.

## Trigger words

- `soyokaze` — main trigger (always required)
- `kaze` — optional trigger for the **"woman in the wind"** motif (fluttering hair and kimono). Combine with `wind, floating hair` to strengthen it.

## Recommended usage

**Base model: FLUX.1-dev** (tested on flux1-dev-fp8). Do not use with SDXL checkpoints.

```
soyokaze, 1girl, solo, japanese clothes, kimono, pale skin, <scene tags> <lora:soyokaze_flux:1.0>
```

- Weight: **1.0**
- Sampler: Euler / Simple / Steps 20
- CFG 1 / **Distilled CFG 3.5**
- No negative prompt needed
- Size: 832x1216 (portrait) or 1024x1024 (face)

## 🔍 Community exploration wanted — help us stabilize the photoreal face

The photoreal mode is an **emergent behavior** — it lives on the boundary between FLUX's photo prior and the trained print style, so it is sensitive to prompt wording and seeds. We are publishing what we know and **inviting the community to find recipes** that reproduce the ideal painted face (see the first gallery image) in more poses and scenes.

**Known findings so far:**

- The face is triggered by wording, not weight: keep `traditional edo-period facial features` (add `white oshiroi makeup, red lips` to lock it further). Removing this phrase gives you a modern face even at weight 1.0.
- Words that kill the painted face: `RAW photo`, `real japanese woman`, `visible pores`, `cinematic photography` — they pull toward modern realism.
- Edo-themed scene words (teahouse, lanterns, veranda) raise the chance of flipping into a full woodblock print. Both outcomes are usable; batch and pick.
- The first three gallery images are confirmed successes of the **same recipe** (identical prompt, weight 1.0) on three different seeds: `1451267015`, `1142232125`, `4028250098`. Full parameters are embedded in each — start from one of them and change one thing at a time.

**If you find a prompt/setting that keeps this face stable in new poses or scenes, please post your image (parameters embed automatically). Great recipes will be featured and added to this description with credit.**

## Training data

Trained exclusively on copyright-free materials.

---

# 日本語

## 🔍 有志募集 — 実写モードの「化粧された江戸の顔」を安定させるレシピ探し

実写モードはFLUXの写真事前知識と版画様式の境界で成立している**創発的な挙動**で、プロンプトの言い回しとシードに敏感です。判明していることを全て公開し、**この理想の顔(ギャラリー1枚目)を様々なポーズ・場面で安定して出すレシピをコミュニティに募集**します。

現時点での知見:

- 顔のトリガーは重みではなく語句: `traditional edo-period facial features` を必ず入れる(`white oshiroi makeup, red lips` でさらに固定)。この句を削るとw1.0でも現代顔になります
- 顔を殺す語: `RAW photo` / `real japanese woman` / `visible pores` / `cinematic photography`(現代写実に引っ張る)
- 江戸系の場面語(茶屋・行燈・縁側)は完全版画への転落率を上げます。どちらも使える絵が出るのでバッチで選別を
- ギャラリー先頭3枚は**同一レシピ(同一プロンプト・重み1.0)の成功例**で、シードのみ違います(`1451267015` / `1142232125` / `4028250098`)。各画像に再現用パラメータを全て埋め込んであります。そこを起点に1要素ずつ変えるのが近道

**この顔を保ったまま新しいポーズ・場面が出せたら、ぜひ画像を投稿してください(パラメータは自動で埋め込まれます)。優れたレシピはクレジット付きで説明文に反映します。**

**著作権フリーの素材のみ**で学習した画風 LoRA の FLUX.1-dev 版です。切れ長の目・点のような瞳・小さな口・白い肌の浮世絵美人顔と、伝統的な木版画の線と彩色を再現します。

FLUX はベースモデル1本で全様式を扱うため、この LoRA 1本がプロンプト次第で3通りに使えます:

1. **完全版画調**(デフォルト)
2. **実写×浮世絵**: プロンプト先頭を `photorealistic portrait photograph of a japanese woman, soyokaze, ...` にすると、写真の質感のまま江戸の結い髪・装いが乗ります(重み1.0のままでOK)
   ※重み1.0の実写モードは**シードによって写真になるか完全版画に転ぶかが揺れます(仕様)**。どちらも使える絵が出るので、数枚バッチで生成して選んでください。版画落ちを減らしたい場合は重み0.85〜0.9、Distilled CFG 3.5〜4に
3. アニメ調プロンプトは版画調に変換されます

- トリガー: `soyokaze`(必須)、`kaze`(風の中の女。`wind, floating hair` 併用で強化)
- 設定: 重み 1.0 / Euler / Simple / steps 20 / CFG 1 / Distilled CFG 3.5 / ネガティブ不要
- **FLUX.1-dev 専用**。SDXL 系チェックポイントでは使えません(SDXL 版は姉妹モデル soyokaze / soyokaze hybrid を参照)