AURORAL Nostalgia · 极光旧梦 v1.1 — AURORAL Nostalgia

AURORAL Nostalgia · 极光旧梦 - v1.1 — AURORAL Nostalgia

NSFW

中文速览：极光旧梦是一个电影感光影 + 画质提升 LoRA，底模 Krea 2，二次元与写实双风格通用。它给画面加一层电影感光照——暖调定向光、金色轮廓光与边缘辉光、柔和泛光、更丰富的中间调过渡，并收尾以更干净的细节与细腻胶片颗粒。二次元插画上能强化氛围与完成度；写实人像 / 摄影风画面同样适用，能做出自然的窗光、逆光、轮廓光与胶片质感。无需触发词，它不改变画面里的角色与构图，只改变光怎么打在画面上、以及成图有多完整。权重建议 1.0（设计强度）或 1.3~1.5（光效更浓，画廊样图即用 1.5）。提示词只写「主体 + 构图 + 场景」效果最好，堆叠风格词会稀释它。样图的生成参数全部公开，可直接点开查看。

文件档位 · File options（v1.1 起）

v1.1 版本提供 4 个大小档，内容与用法完全相同，只有 LoRA rank 不同：rank128 · 872 MB（默认、完整）、rank64 · 436 MB（保留 95.6% 更新能量）、rank32 · 218 MB（90.2%）、rank16 · 109 MB（85.3%）。三个小档是 SVD 最优低秩截断（不是删层 / 删通道），风格主调完整保留；实测风格统计与原档的差异小于「同档换 seed」的正常波动。无需触发词，权重照旧 1.0 或 1.3–1.5，切换文件不需要改参数。详见 v1.1 版本说明。

AURORAL Nostalgia is a cinematic lighting and quality-enhancement LoRA for Krea 2, working on both anime and realistic / photographic styles. It lays a cinematic light pass over the image — warm directional light, golden rim light and edge glow, soft bloom and halation, richer mid-tone gradients — and finishes the render with cleaner detail and a subtle analogue film grain, while leaving your subject and composition exactly where your prompt puts them. On anime it lifts mood and finish; on realistic portraits and photography-style scenes it adds natural window light, backlight, rim light and a film-like grade.

It does not change who or what is in the frame. It changes how the light falls on it, and how finished the result looks.

File options

Four file sizes are provided for both v1.0 and v1.1: rank128 (872 MB, default/full), rank64 (436 MB), rank32 (218 MB), rank16 (109 MB). The smaller files are SVD-truncated (optimal low-rank approximation), not layer skipping — same look, no trigger word, same strength. v1.0 sizes keep 96.0% / 91.4% / 87.1% of the update energy; v1.1 sizes keep 95.6% / 90.2% / 85.3%. v1.0 and v1.1 are different weights — do not mix sizes across versions. See each version's notes for details.

Trigger word

None. The effect applies automatically — just describe your subject and scene.

What it does

Lighting: warm directional light, golden rim light and edge glow, soft bloom / halation, stronger separation between lit areas and shadows.
Quality: cleaner rendering with finer line and detail retention, smoother tonal gradients, gentle depth of field and a fine film-grain finish — the "finished illustration" look rather than flat output.
Colour: warm–cool separation and a nostalgic film grade, richer without going oversaturated.

Recommended strength

1.0 — designed strength: a clear lighting and quality upgrade, content intact.
1.3 – 1.5 — noticeably stronger glow, bloom and film character. Our own gallery samples were made at 1.5.
0.8 and below — the effect fades quickly.
On Turbo / quantized Krea 2 checkpoints the effect is a little weaker than on the full-precision base.

Works best with

Anime and semi-realistic anime illustrations — single character, half-body or full-body.
Realistic / photographic portraits and scenes — the LoRA shapes the light naturally and adds a film grade on top of a photoreal base.
Scenes that profit from atmosphere: dusk, night streets, lantern light, interiors, gardens, window light.
Prompts that already describe the subject well — let this LoRA supply the light and the finish. Your own style words dilute it.

Honest limitations

A declared style in your prompt wins. If you ask for completely flat, unlit cel shading, the LoRA will mostly add colour and grain instead of repainting the light.
Best with content-only prompts (subject / framing / scene) rather than stacked style keywords.
At 1.5+ on very soft or low-detail scenes some micro-detail is traded for mood; drop to 1.0–1.2 when you want the texture back.

Notes

All gallery samples were generated with this LoRA on Krea 2, 10 steps, CFG 1.0, euler / simple, at 1024 px on the short side. The generation data is public — open any sample image to read its full prompt and settings.

Description

v1.1 更新 · 2026-10-25

中文速览：v1.1 是对极光旧梦的一次定向优化：重新调校了复古色彩、光影层次与服装细节的表现，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：无需触发词，推荐权重 1.0（设计强度）或 1.3–1.5（光效更浓，画廊样图用 1.5）。

文件档位 · File options（v1.1 起）

同一个 v1.1，4 个文件可选，内容与用法完全相同，只有 LoRA rank（体积 / 细节保留率）不同：

主档 — rank 128 · 872 MB（默认下载）。完整权重，设计与样图均以此生成。
Lite — rank 64 · 436 MB，保留 95.6% 更新能量（SVD 截断）。实测与原档的风格统计差异（色调 / 饱和 / 对比 / 亮度 / 频域细节）小于「同一文件换一个 seed」的正常波动，日常出图可视为与原档一致。
Mini — rank 32 · 218 MB，保留 90.2%，同样无可测的风格损失，适合批量出图 / 显存或带宽紧张。
Nano — rank 16 · 109 MB，保留 85.3%。色彩、光影、氛围统计仍与原档一致；但个别模块只剩约 40% 能量，极端吃细节的题材（复杂服饰、精细金属、大量细碎文字状元素）建议用 64 / 32 档。

三个小档是 SVD 低秩截断（Eckart–Young 最优近似，按奇异值保留主方向），不是删层、不是删通道，所以风格主调（复古色调、胶片颗粒、光感）完整保留。无需触发词，权重照旧填 1.0 或 1.3–1.5，切换文件不需要改任何参数。

注意：同一个 seed 换不同档位会得到另一张构图（权重不同 → 采样轨迹分叉），这是扩散模型的正常行为，不是错误；如果要复现某张旧图，请使用生成它的那个文件。

v1.1 的优化（相对 v1.0）

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。
摄影光感更强：本次样图（8–10 步、CFG 1.0、短边 1024、权重 1.5）里，窗光、丁达尔光柱、台灯暖光、逆光轮廓的层次与胶片质感比 v1.0 更明确，画面完成度更高。
风格与用法不变：给动漫 / 半写实 / 写实画面加一层电影感光照 + 复古胶片调（二次元与写实风格都生效，写实人像同样能拿到自然的窗光、逆光与胶片质感），保留人物与构图；提示词只写「主体 + 构图 + 场景」效果最好，堆风格词会稀释它。

English: v1.1 is a targeted optimisation of AURORAL Nostalgia — the vintage colour, lighting depth and outfit detail have been retuned, and the outfit drift that v1.0 sometimes showed on certain clothing prompts is fixed. The overall mood and usage are unchanged: no trigger word, recommended strength 1.0 (designed) or 1.3–1.5 (stronger; gallery at 1.5).

What v1.1 improves (vs v1.0)

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.
Stronger photographic light. The new samples (8–10 steps, CFG 1.0, 1024 px short side, strength 1.5) show clearer window light, dust beams, practical lamps and rim light, with a more finished, film-like look than v1.0.
Same mood and usage. A cinematic light pass plus a vintage film grade on anime, semi-realistic and realistic images alike — realistic portraits also get natural window light, backlight and a film-like finish — leaving subject and composition untouched. Content-only prompts work best.

File options

v1.1 ships 4 files with identical usage — only the LoRA rank differs. rank128 = 872 MB (default, full), rank64 = 436 MB (95.6% of update energy kept), rank32 = 218 MB (90.2%), rank16 = 109 MB (85.3%). The three smaller files are SVD-truncated (Eckart–Young optimal low-rank approximation — not layer or channel skipping), so the signature vintage look is intact. In our same-seed tests their style statistics differ from the full file by less than the normal seed-to-seed variation. No trigger word; keep strength at 1.0 (designed) or 1.3–1.5. Note that re-using a seed with a different size gives a different composition — expected behaviour for diffusion models.

Recommended settings

Strength: 1.0 — designed strength; 1.3–1.5 — clearly stronger glow and film character (the gallery was made at 1.5).
Sampling: 8–10 steps, CFG 1.0, euler / simple or dpmpp_2m / simple, 1024 px on the short side.
On Turbo / quantized Krea 2 checkpoints the effect is a little weaker than on the raw bf16 base.

Honest limitations

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

Samples

This version adds 18 new samples (cover first, then portrait 9:16, then landscape 16:9). The generation data is public — open any sample image to read its full prompt and settings.