蒂法·洛克哈特 | Tifa Lockhart (Krea 2 LoRA)

触发词 / Trigger word：tifa

中文说明

角色

《最终幻想 VII》的蒂法·洛克哈特。本 LoRA 以 cosplay 实拍为素材训练，走写实电影感路线：白色背心 + 黑色短裙 + 红色手套 + 黑色长靴的经典造型，长直黑发、红色瞳孔。

注意：因为训练素材是真人 cosplay 影像，出图的面容更接近真人 coser 的骨相，而不是游戏 CG 里的原设模型脸。想要更接近 CG 感可以降低权重，或在提示词里加强 3d render, game asset。

素材构成（这是本 LoRA 的特点）

训练集全部来自中文圈 coser 的实拍作品（12 位 coser，共 729 张原图 → 精筛 200 张），刻意避开了欧美 coser：

• 每个 coser 最多取 24 张，按来源均衡限流，避免模型被某一个人的脸主导；

• 只保留竖版成品图，剔除双格拼版、低清、重复；

• 打标时禁止写入发色 / 发长 / 瞳色 / 肤色 / 人种 / 年龄，角色固有特征只由触发词和图像本身承担。

所以这个 LoRA 学到的是「角色的共性」，而不是「某一个特定 coser 的脸」。

用法

最简：

tifa

推荐结构（先定景别机位 → 再定服装 → 再定场景光线）：

tifa, full body, standing, looking at viewer, front view,
white tank top, black mini skirt, red gloves, black boots,
industrial steel catwalk, dim warm light,
photorealistic, cinematic lighting

推荐参数（Krea 2）

项目建议权重 strength0.9 ~ 1.0（1.0 最贴角色，0.75~0.85 更容易接受新构图/新服装）采样器er_sdeCFG2.6（2.4~3.0 区间）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放大像素域 lanczos 放大到 1.5 倍后低 denoise 二段，别用 latent 放大（会糊成方块）

提示词小贴士

• 服装词要写全。只写 tifa 不写衣服，模型会自由发挥。经典造型就写 white tank top, black mini skirt, red gloves, black boots。

• 想换装（礼服、护士服、泳装）就明确写目标服装，权重压到 0.75~0.85，角色脸仍然守得住——训练时服装已做解耦，换装不会变成另一个人。

• 侧面／背身机位比正面更容易崩手，负面里加 extra limbs, bad hands 有效。

• 出图若带训练素材遗留的签名或水印残影，负面加 watermark, signature, text 并适当提权，或直接裁掉底部。

训练信息

底模Krea 2 (krea2_raw)素材200 张 cosplay 实拍（12 位 coser，中文圈）分辨率1024×1024，开启 bucket步数2000 步 / 10 epoch网络dim 64 / alpha 64学习率1e-4，cosine，warmup 40其他timestep sampling = krea2_shift，shift 1.0，8bit AdamW，fp8 base + fp8 scaled

授权

• 角色版权归 Square Enix 所有，本 LoRA 为非官方同人作品。

• 训练素材来自公开发布的 cosplay 摄影作品，素材版权归各自 coser / 摄影师所有。

• 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

Tifa Lockhart from *Final Fantasy VII*. Trained on live-action cosplay photography, aiming for a photoreal, cinematic look: the classic white tank top, black mini skirt, red gloves and black boots, with long straight dark hair and red eyes.

Heads-up: since the dataset is cosplay photography, faces lean toward the real cosplayers' 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.

About the dataset

All training images come from Chinese-speaking cosplayers (12 cosplayers, 729 raw images → 200 carefully picked):

• Capped at 24 images per cosplayer — balanced sampling by source — so the model doesn't collapse onto one person's face;

• Portrait-format finished shots only; two-panel collage pages, low-res and duplicates removed;

• Captioning explicitly forbids hair colour, hair length, eye colour, skin tone, ethnicity and age, so the character's fixed traits are carried only by the trigger word and the images themselves.

The result is a LoRA that learns the character's shared traits rather than one specific cosplayer's face.

Usage

Minimal:

tifa

Recommended structure (framing → outfit → scene/light):

tifa, full body, standing, looking at viewer, front view,
white tank top, black mini skirt, red gloves, black boots,
industrial steel catwalk, dim warm light,
photorealistic, cinematic lighting

Recommended settings (Krea 2)

LoRA weight0.9 – 1.0 (1.0 = closest likeness; 0.75–0.85 = more freedom for new outfits/poses)Samplerer_sdeCFG2.6 (2.4 – 3.0)Steps20–24 (second pass 12 steps at 0.35 denoise)Base resolution1024×1024, portrait 1024×1536, landscape 1536×1024Upscalingpixel-domain lanczos to 1.5× then a low-denoise second pass — avoid latent upscale (turns fine detail into 16–32px blocks)

Prompt tips

• Always spell out the outfit. tifa alone gives the model too much freedom. For the classic look use white tank top, black mini skirt, red gloves, black boots.

• For outfit swaps (evening gown, nurse uniform, swimsuit) state the new clothing explicitly and drop the weight to 0.75–0.85 — the face still holds, because outfits were decoupled during training.

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

• Some outputs inherit signatures or watermarks from the source photos — put watermark, signature, text in the negative, or just crop the bottom edge.

Training details

Base modelKrea 2 (krea2_raw)Dataset200 cosplay photos (12 Chinese-speaking cosplayers)Resolution1024×1024 with bucketingSteps2000 steps / 10 epochsNetworkdim 64 / alpha 64Learning rate1e-4, cosine, 40 warmup stepsMisctimestep sampling = krea2_shift, shift 1.0, 8bit AdamW, fp8 base + fp8 scaled

License / credits

• Character © Square Enix. This is an unofficial fan-made LoRA.

• Training images come from publicly released cosplay photography; copyright remains with the respective cosplayers and photographers.

• 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.