LMAO, my bare ass got recognized as flesh-colored leggings.

It depends on how many lora you are using at once. One lora? 0.8-1.0. Multiple lora? The more you use, the more you will have to tone town each lora strength. Experimentation will be necessary. Some lora are trained more than others. Some combine well with others at 1.0 but most on here are a bit over trained and will need to be turned down.

@ivever 很遗憾，如果不使用lora，我生成的走路视频很糟糕，向前向后走路都生成不了正常的行走视频，不是怪异就是抽搐诡异的，4060ti16gb，128gb内存，m2固态，用的是万能君那个版本，有可能用工作流就不会有这种问题吗，郁闷了很久了

这里有向前走的，https://civitai.com/models/1427481/catwalk，https://civitai.com/models/1447298/wan21walking-to-viewer-i2v-480720p-paseer，https://civitai.com/models/1361346/bouncy-walk-wan-i2v-14b

@ivever 万能君是b站一个大佬，他懂编程，他把工作流和模型等打包成一个易用的windows程序，他做了很多这样的程序而且免费，可自愿赞助他。走路问题我已解决，我用了comfyui工作流，且工作流比万能君版本更高清，不能正确走路的问题是：1，他那个版本存在问题，我已向他反馈，2，图生,视频（480p-itv）尺寸最佳为竖屏480x720，用此尺寸，在工作流上，即使不用lora，跑10次能有8次正确走路姿态，之前用的480x832，很糟糕，抽卡很难，3，在改图做图时，有时候要注意构图，虽然ai自动构图基本完美，但有些图透视错误会造成wan无法判断人和路面的透视结构，生成的视频会很怪异，抽搐等问题。

@ivever 已经看到了，已下载，改天测试一下，谢谢

Got it, thanks!

@morozig I Train also Loras with Video Clips in 720p. On a mac Studio with m4max with 96GB RAM/vRam it takes around 40GB of vRam.

I am hoping to find someone here on civitai who has experience training a motion-focused wan lora using videos on civitai's lora trainer. I have tried a couple of times but haven't had any luck and I don't know if I have the wrong settings, not enough videos, etc. I have a few wan character models based on still images and they work very well but I can't seem to get the motion loras to work and I burn through buzz in a hurry with experimentation.
Does anyone reading these comments have any experience or advice on this? Thanks!

@darkroast175696 I hope you've found a solution or workaround since I am in the same situation.

@xXRogerXx Sadly I haven't made any progress. I did try an experiment to prove that the videos are really being used in the lora generation. I found a random tiktok dance video, cut it down to 512x512, and made 10 copies of it. I fed that into the generator to see what would come out. After several repeats, all the videos started to look like the tiktok video so that was proof that the videos ARE being used. That means it's down to user error (me) and I haven't figured out how to fix that yet.
Is it my captions? Is it the number of repeats/steps/epochs? Other settings like learning rate and whatnot?
It's a little frustrating to experiment on civitai, aside from the buzz cost, because of two limitations here. First you get a max of 20 epochs. So if you think you may need a bunch of steps, the only way to get it is to increase repeats which means the bouncy nature of the learning rate means you can't get an epoch for every repeat and you might miss the one or two epochs that are actually good.
Second there isn't any data provided about the output. There is an outstanding developer request to add tensorboard log output to the training results which would help a LOT with testing, but I don't know if they're planning to do that at all or how soon.
So for the time being I'm just doing stuff I can make from still images like characters and clothing and so on. If any of the motion creators out there feel like sharing some of their data set or training parameters, that would be really helpful.

@ivever  thank you

@ivever then my dude is about to become a lady

actually set at 0.5 strenght it helped a lot. I had a room full of people walking away, and none would move with promptiong. using your lora they now walk away as requested. everyyone is walking away instead of no-one. I think setting it low strength enables the prompt to act on it.

I use for men between 0.35/0.5
Beyond that, the walk is too... languid.
Without this lora it is very difficult to make the characters move.