Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringclasses
6 values
ejie, a real Chinese woman in her fifties, three-quarter view turned to the right, half body shot
ejie, a real Chinese woman in her fifties, side profile facing right, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the left, half body shot
ejie, a real Chinese woman in her fifties, side profile facing left, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the left, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the right, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the left, half body shot
ejie, a real Chinese woman in her fifties, seen from behind, back view, half body shot, standing, back to camera
ejie, a real Chinese woman in her fifties, side profile facing left, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, side profile facing left, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, seen from behind, back view, half body shot, standing, back to camera
ejie, a real Chinese woman in her fifties, seen from behind, back view, half body shot, standing, back to camera
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the left, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the right, half body shot
ejie, a real Chinese woman in her fifties, facing the camera, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the right, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the left, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the left, half body shot
ejie, a real Chinese woman in her fifties, three-quarter view turned to the right, half body shot

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

ejie LoRA 训练数据集

店长形象专属 LoRA(SDXL)。触发词 **ejie**。 生成时间:本次会话实测产出。所有参数均基于本机实测数据。


1. 数据集概况

项目 值
触发词 ejie
图片总数 33 张
主图(近景 5 + 全身 14) 19 张
脸部特写(从全身图重裁) 14 张
图片尺寸 长边 1600 / 1400,训练时裁 1024
标注 33 个同名 .txt(kohya 标准)
来源 影楼原片 19 张(7008×4672 与约 3000×6100)

图片分组

编号 内容
ejie_01 ~ ejie_05 近景:正面 / 3-4 侧 / 正面笑 / 正侧面 / 3-4 侧
ejie_06 ~ ejie_19 全身:正面 / 3-4 侧 / 正侧 / 背面回头 / 走动
ejie_06face ~ ejie_19face 上述全身图重裁出的脸部特写,用于补强五官分辨率

2. 预处理做了什么(以及为什么)

✅ 做了:曝光归一

问题:素材分两组,曝光差 2~3 倍。

组 原始肤色亮度 原始画面亮度
近景组(#1–#5) 91 ~ 143 57.8 ~ 88.9
全身组(#6–#19) 122 ~ 143 99.7 ~ 161.4

处理:按肤色亮度归一到 132,带高光保护(>250 的像素占比上限 2%)。

结果:肤色亮度极差 52 → 6,高光溢出 0%。

❌ 没做:白平衡校正(重要)

原本以为:肤色 R−B 在 34~52,是灯光色偏,应该校正。

实测发现:画面中的墙面本身就是中性灰(R−B ≈ −4)。用肤色去反推白平衡,会把中性灰推成青色——实测验证过,力度 0.85 时墙面 R−B 变成 −55,画面明显发蓝。

结论:这个暖调是她本人的真实肤色,不是色偏。不校正,让 LoRA 把它作为稳定特征学进去。

⚠️ 这条结论对后续所有 SKU 拍摄都适用:先看画面里有没有已知的中性灰参照物,有的话以它为白平衡基准,不要用肤色。

✅ 做了:脸部特写重裁

问题:全身图里的脸只占画面高度 16%,缩到 1024 训练尺寸后脸部仅约 164px。

处理:从原图(最高 7008px)重新裁出脸部特写,脸部像素达 730 ~ 1325px,与近景组(451px)量级相当。

这是真实信息的重新取景,不是合成数据。


3. 关键发现:SDXL 底模会把"中年亚洲女性"画老

本机此前已实测(见 店长LoRA重训方案.md):

对照 结果
无 LoRA 纯提示词 ❌ 陌生人,且明显偏老(约 65 岁)
提示词写 50 year old, mature but not elderly ❌ 完全无效
旧 LoRA(proxy_owner_*) ❌ 也是另一个人

触发词里带 proxy 说明上次训练用的是代理替身,不是本人——这是上次失败的根本原因。

这次素材已逐张确认为本人。


4. 云 GPU 训练参数

4.1 训练器选择

方案 说明
kohya_ss(推荐) 社区标准,CUDA 原生支持,云 GPU 上开箱即用
SD-Trainer / ComfyUI 训练节点 图形界面,适合不想写命令

云 GPU 建议:RTX 4090 / A100 单卡,24GB 显存足够

4.2 核心参数

参数 值 说明
底模 RealVisXL_V5.0_fp16.safetensors 写实人像强
分辨率 1024 不要用 512,写实人脸不够细
network_dim (rank) 16 33 张素材用 16;若过拟合降到 8~12
network_alpha 8 = dim / 2
学习率(UNet) 1e-4
学习率(TE) 1e-5
优化器 Prodigy 或 AdamW8bit Prodigy 更省心
batch size 1 ~ 2
总步数 ≈ 3300 ~ 5000 33 张 × 100~150
保存频率 每 250 步存一版 ⭐ 关键,方便挑最佳版本
混合精度 bf16 / fp16
打标格式 已有 .txt,直接用,不要重新打标

4.3 ⭐ 最重要的一条:必须多存检查点

不要只保留最后一个 LoRA。 每 250 步存一版,训完拿 8~15 个版本逐个测,挑最像的那一版。

上次的 wyy_s250 / s750 / s1500 就是这个思路,做法是对的。


5. 验证流程(训练后必做)

5.1 固定对照组

同一 seed、同一提示词,分别跑:

  • 无 LoRA(A0 基准)
  • 各步数检查点(s250 / s500 / … / s3000)

5.2 测试提示词

ejie, 50 year old chinese woman, mature but not elderly,
natural skin texture with pores and fine lines,
wearing a white shirt, front view, half body, plain white background,
photorealistic, high detail

5.3 一致性压力测试

用最佳检查点,生成 4 个不同角度,检查:

  • 是不是同一个人
  • 是否随角度漂移
  • 年龄感是否稳定在 50 岁左右

5.4 与云端对照

拿同一张参考图,对比本机 LoRA 出图 vs 云端 gpt-image-2.5 出图(后者已验证能"像")。

判断标准:本地方案至少要达到云端的相似度,才值得切换到本地。


6. 负面提示词(写进 negative)

sunglasses, smooth plastic skin, airbrushed, beauty filter,
young, teenager, elderly, wrinkles exaggerated,
blurry, lowres, watermark, text, multiple people,
deformed face, extra fingers

7. 训练完成后的整合路径

① ejie LoRA(训练产出)
      ↓ 生成「模特基准图」——正面全身 / 45° / 侧面 / 背面
② IDM-VTON(本机已打通,52s/张)
      ↓ 把商品图的衣服像素级穿到基准图上
③ 成品图
      ↓ 喂给 Seedance 出视频

关键收益:

  • 脸 → 由 LoRA 锁定,不再漂移
  • 衣服 → 由 IDM-VTON 像素级对齐,连"商品图偏色 14.8°"的问题都绕过了(它不做色彩理解,只做像素对齐)
  • 成本 → 电费 / 少量云 GPU 费,边际成本趋近 0

8. 已知限制

限制 说明 影响
素材 33 张 偏少(目标区间下沿) 需依赖检查点筛选,rank 不宜过大
表情分布 微笑偏多,中性约 4 张 生成中性表情时可能仍带笑意
服装单一 全部黑色高领 换装依赖 IDM-VTON,不能靠 LoRA 换衣
妆发固定 同一次拍摄 生成其他妆发时可能不稳定
微信传输 文件名含"微信图片",可能二次压缩 当前分辨率足够,但非绝对原片

本文档基于本机实测数据生成。白平衡结论、曝光归一效果、脸部像素测算均为实测值。

Downloads last month
161