@ivever 老铁，是谁的视频，青龙吗？

@ivever 是用480p训练的？可能是万相底子问题，细节还是容易变异。

@ivever Thanks for the response, but I'm not sure what you mean by "The original Wan model is unable to generate a dynamic video of this pose" as a reason why this won't work for T2V. Sure that might be why this particular lora isn't working right now, due to how you trained it, but that's no reason why it couldn't be made to do it, as there are other loras that do poses that aren't in base Wan. So I'm just a bit confused as to why this lora is limited in a way that others aren't. Surely it's no extra effort to make it work with T2V training-wise?

I2V is way more useful than T2V, besides, OP has his own motivations and is kind enough to share. Be nice.

@ivever Thanks for clarifying! It would be great if you were able to do that, as it would make this lora useful for those of us that don't use I2V (my card can barely even support I2V due to the VRAM requirements, and besides I prefer the ease of use of T2V, and Wan's natural realistic style rather than Flux plastic).

Not sure why everyone is downvoting me like crazy. I'm just trying to understand why there are limitations here that don't exist in other loras, and OP has kindly explained it.

@kangaru861 bruh everyone's downvoting because you are being an ungrateful kid whining your ass off for something op shared out of kindness.

normal human would be like “love all the videos! thank you for sharing! sadly I'm using t2v and couldn't get the results I was hoping for. is it possible to train this for t2v too? hope I'm not asking too much!”

grow up dude

@jknt "Kid" 🤣. I'm not "whining", I expressed frustration because I had wasted my time trying to get the lora to work (given Wan is so slow), only to find it doesn't work for T2V, and I wanted to understand why not, because I would like to be able to use this lora.

When I posted, it wasn't clear that this lora won't work for T2V (unless I somehow missed it, the creator added a clarification afterwards). It was especially frustrating because it was not clear to me at all why it shouldn't work, given other loras work fine with both. The creator has been very helpful, and has clarified the situation, and I'm very grateful for their efforts and explanations.

Could I have been more tactful in my first message? Sure, I was very blunt, especially with "Honestly I don't see the point in this lora" which does sound rude on reflection, but that was not my intention. I just don't get why so many people are apparently okay with what seems like a pretty severe and unnecessary limitation. Again, the creator has since clarified.

I personally think t2v is unnecessary anyway, best to create the still image according to the plan with what ever models and loras are appropriate (I do mine with Flux and Pony), use inpainting etc to get the details right, then i2v to make a video. Then extend the good ones until there's long nice video.

@ivever thank you for making a t2v version! the i2v is indeed superior.

@ivever Thanks! I'll give it a try anyway

You're not doing anything wrong, it's really meant for I2V. I can generate T2V with this model, but it forces the characters to have certain style. As you say, it would need to be retrained for T2V (in a way that works - it was not trained for motion alone) for being useful in that mode.

Only training on I2V seems like a waste of time. All the other T2V loras work on I2V. Dont see the point of training on I2V

@ivever Yeah, please don't take this the wrong way. We appreciate you spending your time and energy to make Loras. This is just feedback to make it better. There's something about the training on this one that doesn't allow it to work with most workflows. I've used several loras and never had to add the "Skip Layer Guidance" nodes on other Loras. This one when just executed out of the box on a standard I2V workflow, doesn't animate very well at all.

It might be worth it to check how other Loras have been trained.

@ivever Thanks, I'll check it later. It does work well with Kijai's workflow though.

Did you manage to get it to work because I have the same problem. No movement at all no matter what I do. Using I2V 480p GGUF with fp8

@Yc3k So, I checked OP's workflow. It relies on a node that allows you to skip layers. There's a native version of it called "Skip Layer Guidance WanVideo" that you need to attach after the Teacache node. Blocks have to be set to 10 , start percent at 0.00 and end at 0.60. However, even this doesn't result in the movement all the time, it's a roulette.

I've found better success using this in tandem with the WAN General NSFW lora (WAN General NSFW model (FIXED) - v.014 (FIXED) | Wan Video LoRA | Civitai).

I uploaded the image to the gallery, just download it for the workflow.

@ivever That was probably the problem cause I was using the ComfyUI. Thanks.

Reading this is so funny dude. Props to all of you 🥂