<p>中文速览：唯美复古老照片风格 LoRA，底模 Krea 2。触发词写在提示词结尾：<code>AURORAL Retro</code>。权重建议 1.0（设计强度）或 1.3~1.5（复古感更浓）；提示词只写「内容 + 构图」，不要再写风格词，否则会把 LoRA 稀释。若画面偶尔出现 <code>AURORAL Retro</code> 字样，把 <code>text, watermark, AURORAL, Retro</code> 加进负向提示词即可。文件含主档与 ×2 预缩放档。</p>
<hr />
<p>AURORAL Retro is a soft, nostalgic vintage-aesthetic style LoRA for Krea 2. It washes a scene with warm faded colour, gentle contrast, fine film-like grain and a quiet, dreamy mood — while leaving the content of your prompt where you put it.</p>
<p>It is tuned for people and quiet everyday scenes: portraits, old rooms, rainy streets, courtyards, still life. It is a mood layer, not a subject changer.</p>
<h3>Trigger word</h3>
<p>Put <code>AURORAL Retro</code> at the end of the prompt (that is the form it was trained with).</p>
<h3>Recommended strength</h3>
<ul>
<li>1.0 — the designed strength. Balanced and clean.</li>
<li>1.3 – 1.5 — clearly stronger vintage / film feel. On our own test sets this scored best on average (1.5 was the top setting on 6 of 10 library prompts, and on all 3 anime test prompts).</li>
<li>0.8 and below — not recommended: the style fades and the base model's own texture comes through more.</li>
<li>On Turbo / quantized checkpoints (e.g. krea2 Turbo nvfp4) the effect is a little weaker than on the original raw bf16 base.</li>
<li>A ×2 pre-scaled file is included in this version: load it and set 1.0 to get the same result as 2.0 on the main file.</li>
</ul>
<h3>Example prompts</h3>
<pre>a photograph of a young woman sitting beside a large window in the morning, holding a ceramic cup, wearing a cream knit sweater, half body shot. AURORAL Retro</pre>
<pre>a photograph of a young woman turning her head over her shoulder on a wet street at night, wearing a dark coat, city lights behind her. AURORAL Retro</pre>
<pre>a quiet courtyard with a wooden bench and hanging paper lanterns at dusk. AURORAL Retro</pre>
<p>Suggested negative prompt (see the note below): <code>text, watermark, signature, letters, AURORAL, Retro</code></p>
<h3>What it does well</h3>
<ul>
<li>Portraits and half-body shots of a single person; skin, fabric and light get the vintage treatment.</li>
<li>Quiet interiors and street scenes — old cafes, bookshops, corridors, courtyards, wet asphalt.</li>
<li>Works on both photographic and 2D / anime prompts (see the anime samples in the gallery); on anime prompts it acts as a warm, faded colour grade rather than a redraw.</li>
</ul>
<h3>Honest limitations</h3>
<ul>
<li>Prompt style wins over the LoRA. If your prompt already declares a strong art style ("seinen manga illustration", "flat colour comic", "smartphone UI screenshot"), the LoRA only shifts texture and colour; it will not override that declared style.</li>
<li>Best with content-only prompts. Describe subject / scene / framing and let the LoRA supply the look. Writing your own style words dilutes it.</li>
<li>Occasional trigger text. With short prompts the model can sometimes render <code>AURORAL Retro</code> as lettering inside the image (about 20% of generations in our short-prompt tests; 0/30 on long third-party prompts). Adding <code>text, watermark, AURORAL, Retro</code> to the negative prompt removes it in our tests.</li>
<li>Weak on hard-styled or crowded scenes: decorative flat graphics, anime character sheets and busy multi-subject compositions keep their own look.</li>
<li>Extremes of the strength slider degrade: at 1.5+ a soft scene can lose micro-detail; if that happens, drop to 1.0 – 1.2.</li>
</ul>
<h3>Training details</h3>
<h3>Notes</h3>
<p>All sample images in the gallery were generated with this LoRA. Photographic samples: 1024×1536, 8 steps, CFG 1 (local krea2 Turbo) and upscaled 2× with RealESRGAN. Anime samples: prompt declared <code>anime style illustration, cel shading</code>. The last gallery image is a strength guide: no LoRA / 1.0 / 1.5 on the same prompt and seed.</p>
Base modelKrea 2 (original raw bf16)
TypeStyle LoRA
Networkrank 64 / alpha 64 (scale 1.0)
Resolution1024 (bucketed)
Steps3000 (batch 1)
Learning rate1e-4
Optimizeradamw8bit, gradient checkpointing
Schedulerflowmatch
Dataset135 images, captions: plain content description + trailing trigger phrase
Toolai-toolkit 0.13.19, single RTX 4090 D 48G
File sha256f76dd5cdef0e352c1e62adf2… (main) · f72a0af527188743f4b15f44… (×2)

AURORAL Retro — Vintage Aesthetic Style (Krea 2)

<p>v1.1 更新 · 2026-10-25</p>
<p>中文速览：v1.1 是对 AURORAL Retro 的一次定向优化：重新调校了复古色彩与光影层次，并修复了 v1.0 在特定服装提示词下容易出现的服饰偏移问题。整体风格气质与用法保持不变：柔和暖旧的胶片色调、细颗粒、安静氛围；触发词仍是 <code>AURORAL Retro</code>（放在提示词末尾），推荐权重 1.0（设计强度）或 1.3–1.5（风格更浓，画廊样图用 1.5）。</p>

v1.1 的优化（相对 v1.0）
<ul>
<li>服饰偏移修复：v1.0 在「白衬衫 + 裙子」类提示词下，服装容易被画成日式校服元素（水手领 / 百褶裙 / 过膝袜）。同提示词、同种子三组对照：v1.0 三组全部出现该类元素；v1.1 三组均为普通衬衫领 + 自然垂坠长裙，零校服元素。</li>
<li>摄影质感更稳：v1.1 的样图使用与 v1.0 相同的提示词与种子（8 步、CFG 1.0、短边 1024、权重 1.5），可以逐张对比——肤色、发丝、衣物纹理与窗光/暖灯的过渡更干净，复古胶片感更明确。</li>
<li>风格与用法不变：柔和暖旧的色调、细颗粒、安静的氛围；适合人像、旧房间、雨夜街道、庭院与静物。提示词只写「主体 + 构图 + 场景」效果最好，堆风格词会稀释它。</li>
</ul>

<p>English: v1.1 is a targeted optimisation of AURORAL Retro — the vintage colour and lighting have been retuned, and the outfit drift that v1.0 sometimes showed on certain clothing prompts is fixed. The mood and usage are unchanged: soft faded colour, fine grain, quiet atmosphere. Trigger word stays <code>AURORAL Retro</code> (at the end of the prompt), recommended strength 1.0 (designed) or 1.3–1.5 (stronger; gallery at 1.5).</p>

What v1.1 improves (vs v1.0)
<ul>
<li>Outfit drift fixed. v1.0 could drift plain "white shirt + skirt" prompts into Japanese school-uniform elements (sailor collar, pleated skirt, thigh-high socks). In a same-seed paired test (3 prompts): v1.0 showed those elements in 3/3; v1.1 in 0/3.</li>
<li>More stable photographic texture. The samples were generated with the same prompts and seeds as v1.0 (8 steps, CFG 1.0, 1024 px short side, strength 1.5) so they can be compared one by one — skin, hair, fabric texture and window/lamp light transitions are cleaner, and the vintage film grade reads more clearly.</li>
<li>Same mood and usage. Soft faded colour, fine grain, quiet atmosphere. Best for portraits, old rooms, rainy streets, courtyards and still life. Content-only prompts work best.</li>
</ul>

Trigger word &amp; settings
<ul>
<li>Trigger: <code>AURORAL Retro</code> at the end of the prompt.</li>
<li>Strength: 1.0 — designed strength; 1.3–1.5 — clearly stronger vintage / film feel (the gallery was made at 1.5).</li>
<li>Sampling: 8–10 steps, CFG 1.0, dpmpp_2m / simple or euler / simple, 1024 px on the short side.</li>
<li>On Turbo / quantized Krea 2 checkpoints the effect is a little weaker than on the raw bf16 base.</li>
</ul>

Honest limitations
<ul>
<li>A declared style in your prompt wins over the LoRA — content-only prompts are the way.</li>
<li>On short prompts the trigger words can occasionally appear as lettering — add <code>text, watermark, AURORAL, Retro</code> to the negative prompt.</li>
<li>At 1.5+ on very soft scenes some micro-detail is traded for mood; drop to 1.0–1.2 when you want the texture back.</li>
</ul>

Samples
<p>This version adds 10 new samples, generated with the same prompts and seeds as the v1.0 gallery for a direct comparison. The generation data is public — open any sample image to read its full prompt and settings.</p>

Files with SHA256: c81933cc8b5212e5e01f459194a929961c12ada46b0c1138d9cf36b362cdd941