基于wan2.2 文字生成 视频的模型 训练的人物lora
使用DiffSynth-Studio 训练 25G显存
"--dataset_repeat", "50", # 每个 epoch 中数据集重复的次数。
"--num_epochs", "4", # 轮数（Epoch）。
"--learning_rate", "2e-5", # 学习率。
"--lora_rank", "32", # LoRA 的秩（Rank）。
请帮忙检查一下 这个训练是否成功了？
有那些地方是错误的配置（数据集或训练参数）？
我将我的lora 和 训练参数发布出来了。
但是我发现 我无法上传两个大小一样的模型文件
所以我将它放在了 zip 中。
目前发现问题： 腿会变粗、尾巴末端变白、衣服变外观
寻求帮助
---------------------------------
Based on wan2.2 text generation video model, trained character lora, using DiffSynth-Studio to train 25G video memory
Training parameters:
"--dataset_repeat", "50",
"--num_epochs", "4",
"--learning_rate", "2e-5",
"--lora_rank", "32",
Could someone please help me check whether this training was successful?
Are there any obvious issues or incorrect settings in the configuration above, or could the dataset itself be flawed?
I’ve published both my LoRA model and the full training parameters.
However, I couldn’t upload two model files of identical size to the platform, so I’ve packaged them together in a ZIP file.
Problems found so far: legs will become thicker, the end of the tail will turn white, and the appearance of clothes will change
Ask for help
Description
原始数据
Raw data
I've recently been attempting to train a character LORA model using 3000 images and encountered the following issues: 1. Some images in the dataset feature high-resolution bodies and heads, but hands exhibit motion blur. I'm unsure whether this motion blur might interfere with the model and cause negative effects.
2. Certain images depict physical interactions between characters (e.g., handshakes, head touches). I'm concerned these extraneous limbs might interfere with the model's learning of the main character's appearance.
I would greatly appreciate any insights or assistance on these points.
我最近在尝试 使用3000张图片训练人物lora，
遇到问题：
1、数据集中部分图片的身体和头部是高清的，但是手部有运动模糊，我不知道这种运动模糊是否会干扰模型 带来负面影响。
2、人物和另一个人物有肢体交互 如(握手、摸头)，我担心这种外来肢体 会干扰模型学习主角的外观。
如果有人能帮忙解答，我将十分感谢。