modelscope
Wan2.1 lora training loss
对于wan2.1，在一些开源video-caption数据集上，比如vidgen，koala36m，大概最终的runnning_mean_loss是多少呢 For wan2.1 trained on open-source datasets such as vidgen and koala36m using lora, what is the ultimate running_mean_loss?
对于wan2.1，在一些开源video-caption数据集上，比如vidgen，koala36m，大概最终的runnning_mean_loss是多少呢 For wan2.1 trained on open-source datasets such as vidgen and koala36m using lora, what is the ultimate running_mean_loss?
我训练lora 10多个视频那种 可以到0.03-0.06 数字人视频也是 大致在这个范围波动
对于Wan2.1-1.3B，frame=81, (h,w)=(480, 832), 如果不添加lora，额外在DiTBlock里面添加一些module，在开源数据集上训练，最终收敛的running_mean_loss大致是多少呢？ For Wan2.1-1.3B, set frame=81, (h,w)=(480, 832), if adding no lora, but add additional modules in each DiTBlock, while trained on open-source datasets, what will the running mean loss be on convergence?
对于wan2.1，在一些开源video-caption数据集上，比如vidgen，koala36m，大概最终的runnning_mean_loss是多少呢 For wan2.1 trained on open-source datasets such as vidgen and koala36m using lora, what is the ultimate running_mean_loss?
我训练lora 10多个视频那种 可以到0.03-0.06 数字人视频也是 大致在这个范围波动
感谢！您有没有尝试使用更大数据量，训练lora+wan1.3B呢
一样的 收敛都在这个数值附近 我用了一万个左右的样本
------------------ Original ------------------ From: AliothChen @.> Date: Wed,Sep 24,2025 10:17 PM To: modelscope/DiffSynth-Studio @.> Cc: quanwei zhang @.>, Comment @.> Subject: Re: [modelscope/DiffSynth-Studio] Wan2.1 lora training loss (Issue#943)
AliothChen left a comment (modelscope/DiffSynth-Studio#943)
对于wan2.1，在一些开源video-caption数据集上，比如vidgen，koala36m，大概最终的runnning_mean_loss是多少呢 For wan2.1 trained on open-source datasets such as vidgen and koala36m using lora, what is the ultimate running_mean_loss?
我训练lora 10多个视频那种 可以到0.03-0.06 数字人视频也是 大致在这个范围波动
感谢！您有没有尝试使用更大数据量，训练lora+wan1.3B呢
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>
没测过几个 大概2 3个就有效果了吧epoch 学习率2e-5
------------------ Original ------------------ From: AliothChen @.> Date: Wed,Sep 24,2025 10:52 PM To: modelscope/DiffSynth-Studio @.> Cc: quanwei zhang @.>, Comment @.> Subject: Re: [modelscope/DiffSynth-Studio] Wan2.1 lora training loss (Issue#943)
AliothChen left a comment (modelscope/DiffSynth-Studio#943)
一样的 收敛都在这个数值附近 我用了一万个左右的样本 …
感谢，请问您使用1w个视频大概训练多少个epoch可以收敛呢，学习率是1e-4吗
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>
我用的64卡 几个小时就可以了 分辨率480*720
------------------ Original ------------------ From: SensenGao @.> Date: Mon,Oct 13,2025 5:27 PM To: modelscope/DiffSynth-Studio @.> Cc: quanwei zhang @.>, Comment @.> Subject: Re: [modelscope/DiffSynth-Studio] Wan2.1 lora training loss (Issue#943)
14B模型的
------------------ Original ------------------ From: SensenGao @.> Date: Mon,Oct 13,2025 5:33 PM To: modelscope/DiffSynth-Studio @.> Cc: quanwei zhang @.>, Comment @.> Subject: Re: [modelscope/DiffSynth-Studio] Wan2.1 lora training loss (Issue#943)
SensenGao left a comment (modelscope/DiffSynth-Studio#943)
我用的64卡 几个小时就可以了 分辨率480*720 …
感谢回复！您那边用的是14B模型吗，我想评估一下我使用8卡 1.3B大概要多久
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>
对于wan2.1，在一些开源video-caption数据集上，比如vidgen，koala36m，大概最终的runnning_mean_loss是多少呢 For wan2.1 trained on open-source datasets such as vidgen and koala36m using lora, what is the ultimate running_mean_loss?
我训练lora 10多个视频那种 可以到0.03-0.06 数字人视频也是 大致在这个范围波动
@zhangquanwei962 你好！请问一下你训练过程中的loss曲线是稳定下降的吗，我试了14b模型lora或者sft都是有上升趋势的，单步的loss震荡明显，我的数据集只有300个视频
