<h1 id="or-tifa-lockhart-(krea-2-lora)">蒂法·洛克哈特 | Tifa Lockhart (Krea 2 LoRA)</h1><blockquote><p>触发词 / Trigger word：<code>tifa</code></p></blockquote><hr /><h2>中文说明</h2><h3>角色</h3><p>《最终幻想 VII》的蒂法·洛克哈特。本 LoRA 以 <strong>cosplay 实拍</strong>为素材训练，走<strong>写实电影感</strong>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。</p><blockquote><p>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 <code>3d render, game asset</code>。</p></blockquote><h3 id="lora">素材构成（这是本 LoRA 的特点）</h3><p>训练集全部来自<strong>中文圈 coser 的实拍作品</strong>（12 位 coser，共 729 张原图 → 精筛 200 张），刻意避开了欧美 coser：</p><ul><li><p>每个 coser 最多取 24 张，<strong>按来源均衡限流</strong>，避免模型被某一个人的脸主导；</p></li><li><p>只保留竖版成品图，剔除双格拼版、低清、重复；</p></li><li><p>打标时<strong>禁止写入发色 / 发长 / 瞳色 / 肤色 / 人种 / 年龄</strong>，角色固有特征只由触发词和图像本身承担。</p></li></ul><p>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」。</p><h3>用法</h3><p>最简：</p><pre><code>tifa</code></pre><p>推荐结构（先定景别机位 → 再定服装 → 再定场景光线）：</p><pre><code>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</code></pre><h3 id="krea-2">推荐参数（Krea 2）</h3><p>项目建议权重 strength<strong>0.9 ~ 1.0</strong>（1.0 最贴角色，0.75~0.85 更容易接受新构图/新服装）采样器<code>er_sde</code>CFG2.6（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）</p><h3>提示词小贴士</h3><ul><li><p><strong>服装词要写全</strong>。只写 <code>tifa</code> 不写衣服，模型会自由发挥。经典造型就写 <code>white tank top, black mini skirt, red gloves, black boots</code>。</p></li><li><p>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。</p></li><li><p>侧面／背身机位比正面更容易崩手，负面里加 <code>extra limbs, bad hands</code> 有效。</p></li><li><p>出图若带训练素材遗留的签名或水印残影，负面加 <code>watermark, signature, text</code> 并适当提权，或直接裁掉底部。</p></li></ul><h3>训练信息</h3><p>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</p><h3>授权</h3><ul><li><p>角色版权归 <strong>Square Enix</strong> 所有，本 LoRA 为非官方同人作品。</p></li><li><p>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。</p></li><li><p>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。</p></li></ul><hr /><h2 id="english">English</h2><h3 id="character">Character</h3><p>Tifa Lockhart from *Final Fantasy VII*. Trained on <strong>live-action cosplay photography</strong>, aiming for a <strong>photoreal, cinematic</strong> look: the classic white tank top, black mini skirt, red gloves and black boots, with long straight dark hair and red eyes.</p><blockquote><p>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 <code>3d render, game asset</code> in the prompt.</p></blockquote><h3 id="about-the-dataset">About the dataset</h3><p>All training images come from <strong>Chinese-speaking cosplayers</strong> (12 cosplayers, 729 raw images → 200 carefully picked):</p><ul><li><p>Capped at 24 images per cosplayer — <strong>balanced sampling by source</strong> — so the model doesn't collapse onto one person's face;</p></li><li><p>Portrait-format finished shots only; two-panel collage pages, low-res and duplicates removed;</p></li><li><p>Captioning <strong>explicitly forbids</strong> 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.</p></li></ul><p>The result is a LoRA that learns the character's shared traits rather than one specific cosplayer's face.</p><h3 id="usage">Usage</h3><p>Minimal:</p><pre><code>tifa</code></pre><p>Recommended structure (framing → outfit → scene/light):</p><pre><code>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</code></pre><h3 id="recommended-settings-(krea-2)">Recommended settings (Krea 2)</h3><p>LoRA weight<strong>0.9 – 1.0</strong> (1.0 = closest likeness; 0.75–0.85 = more freedom for new outfits/poses)Sampler<code>er_sde</code>CFG2.6 (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)</p><h3 id="prompt-tips">Prompt tips</h3><ul><li><p><strong>Always spell out the outfit.</strong> <code>tifa</code> alone gives the model too much freedom. For the classic look use <code>white tank top, black mini skirt, red gloves, black boots</code>.</p></li><li><p>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.</p></li><li><p>Side and back angles are harder on hands; <code>extra limbs, bad hands</code> in the negative helps.</p></li><li><p>Some outputs inherit signatures or watermarks from the source photos — put <code>watermark, signature, text</code> in the negative, or just crop the bottom edge.</p></li></ul><h3 id="training-details">Training details</h3><p>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</p><h3 id="license-credits">License / credits</h3><ul><li><p>Character © <strong>Square Enix</strong>. This is an unofficial fan-made LoRA.</p></li><li><p>Training images come from publicly released cosplay photography; copyright remains with the respective cosplayers and photographers.</p></li><li><p>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.</p></li></ul><p></p>

Files with SHA256: cce318078ea5e3efdf55b77b7e80368fefb3e317220fe55c52dd4f8279a57bad