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：AURORAL Retro。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 AURORAL Retro 字样，把 text, watermark, AURORAL, Retro 加进负向提示词即可。文件含主档与 ×2 预缩放档。

---

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.

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.

Trigger word

Put AURORAL Retro at the end of the prompt (that is the form it was trained with).

Recommended strength

• 1.0 — the designed strength. Balanced and clean.
• 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).
• 0.8 and below — not recommended: the style fades and the base model's own texture comes through more.
• On Turbo / quantized checkpoints (e.g. krea2 Turbo nvfp4) the effect is a little weaker than on the original raw bf16 base.
• 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.

Example prompts

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

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

a quiet courtyard with a wooden bench and hanging paper lanterns at dusk. AURORAL Retro

Suggested negative prompt (see the note below): text, watermark, signature, letters, AURORAL, Retro

What it does well

• Portraits and half-body shots of a single person; skin, fabric and light get the vintage treatment.
• Quiet interiors and street scenes — old cafes, bookshops, corridors, courtyards, wet asphalt.
• 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.

Honest limitations

• 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.
• Best with content-only prompts. Describe subject / scene / framing and let the LoRA supply the look. Writing your own style words dilutes it.
• Occasional trigger text. With short prompts the model can sometimes render AURORAL Retro as lettering inside the image (about 20% of generations in our short-prompt tests; 0/30 on long third-party prompts). Adding text, watermark, AURORAL, Retro to the negative prompt removes it in our tests.
• Weak on hard-styled or crowded scenes: decorative flat graphics, anime character sheets and busy multi-subject compositions keep their own look.
• 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.

Training details

Notes

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 anime style illustration, cel shading. The last gallery image is a strength guide: no LoRA / 1.0 / 1.5 on the same prompt and seed.

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)