黑神话：悟空 · 四妹 | Fourth Sister — Krea 2 LoRA

触发词 Trigger word：simei

中文说明

角色

《黑神话：悟空》第四章盘丝岭的蜘蛛精「四妹」。本 LoRA 以四妹 cosplay 实拍 + 游戏 CG 混合训练，走写实电影感、暗调玄幻路线：青绿与墨蓝叠穿长袍、黑色立领内裙、金线滚边、红色短靴、高盘发配金簪，画面常伴蓝光蛛网与烛火。

注意：训练素材以 cosplay 影像为主，出图的面容更接近真人 coser 的骨相，而非游戏引擎里的原设模型脸。想要纯 CG 感，可以降低权重，或在提示词里加强 3d render, game asset。

用法

最简提示词：simei

推荐结构（先景别机位，再服装，再场景光线）：

simei, full body, standing, looking at viewer, front view, green and dark blue traditional chinese robe, black high-collar dress, gold trim, red boots, ancient chinese cave, glowing blue spider webs, candlelight, photorealistic, cinematic lighting

推荐参数

• 权重 strength：0.9 ~ 1.0。1.0 最贴角色；0.75 ~ 0.85 更容易接受新服装、新构图

• 采样器 er_sde，CFG 2.6（2.4 ~ 3.0）

• 步数 20 ~ 24；二段修脸 12 步、denoise 0.35

• 基础分辨率 1024×1024，竖版 1024×1536，横版 1536×1024

• 放大走像素域 lanczos 到 1.5 倍再低 denoise 二段，别用 latent 放大（细节会糊成方块）

提示词小贴士

• 服装词要写全。只写 simei 不写衣服，长袍配色容易漂

• 想换装就写明目标服装，权重压到 0.75 ~ 0.85，角色脸仍然守得住

• 侧面、背身机位更容易崩手，负面加 extra limbs, bad hands 有效

• 少数出图会带训练素材遗留的签名残影（右下角手写体）。负面加 watermark, signature, text 并提权到 1.3 可明显抑制，或直接裁掉底部

训练信息

• 底模：Krea 2（krea2_raw）

• 素材：72 张（cosplay 实拍 57 + 游戏 CG 15），全为完整着装

• 1024×1024 开 bucket；2016 步 / 28 epoch

• 网络 dim 64 / alpha 64；学习率 1e-4，cosine，warmup 40

• timestep sampling = krea2_shift；fp8 base + fp8 scaled

授权

• 角色版权归游戏科学（Game Science），本 LoRA 为非官方同人作品

• 训练素材来自公开 cosplay 作品与游戏 CG，版权归各自原作者

• 请勿用于商业用途，请勿冒充官方，请勿用于真人身份冒用或违法用途

English

Character

The Fourth Sister (四妹), a spider demon from Chapter 4 of Black Myth: Wukong. This LoRA is trained on a mix of cosplay photography and in-game CG, aiming for a photoreal, cinematic, dark-fantasy look: layered teal-and-navy robe, black high-collar dress, gold trim, red boots, high bun with a gold hairpin, usually framed by blue-lit spider webs and candlelight.

Heads-up: since most of the dataset is cosplay photography, faces lean toward the real cosplayer bone structure rather than the in-game 3D model. For a pure CG look, lower the weight or push 3d render, game asset in the prompt.

Usage

Minimal prompt: simei

Recommended structure (framing, then outfit, then scene and light):

simei, full body, standing, looking at viewer, front view, green and dark blue traditional chinese robe, black high-collar dress, gold trim, red boots, ancient chinese cave, glowing blue spider webs, candlelight, photorealistic, cinematic lighting

Recommended settings

• LoRA weight 0.9 ~ 1.0. 1.0 gives the closest likeness; 0.75 ~ 0.85 leaves more room for new outfits and poses

• Sampler er_sde, CFG 2.6 (2.4 ~ 3.0)

• Steps 20 ~ 24; second pass for the face at 12 steps and 0.35 denoise

• Base resolution 1024×1024, portrait 1024×1536, landscape 1536×1024

• Upscale in pixel domain with lanczos to 1.5x, then a low-denoise second pass. Avoid latent upscaling

Prompt tips

• Always spell out the outfit. simei alone gives the model too much freedom and the robe colours drift

• For outfit swaps, state the new clothing and drop the weight to 0.75 ~ 0.85. The face still holds

• Side and back angles are harder on hands; extra limbs, bad hands in the negative helps

• Some outputs inherit a handwritten signature from the source photos (bottom right). Put watermark, signature, text in the negative at about 1.3 weight, or crop the bottom edge

Training details

• Base model: Krea 2 (krea2_raw)

• Dataset: 72 images (57 cosplay photos + 15 game CG), fully clothed

• 1024×1024 with bucketing; 2016 steps / 28 epochs

• Network dim 64 / alpha 64; learning rate 1e-4, cosine, 40 warmup steps

• timestep sampling = krea2_shift; fp8 base + fp8 scaled

Credits

• Character (c) Game Science. This is an unofficial fan-made LoRA

• Training images come from publicly available cosplay photography and game CG; copyright stays with their respective authors

• Non-commercial use only. Do not impersonate the official release, do not use it to impersonate real people, do not use it for unlawful purposes

•