@f1985 Hopefully we will fix this one day. I tested on wan2.1 and 2.2 over 50gb checkpoints. But couldn't fix lips talking even using Loras. Maybe devs see problem. Thanks for your sleepless work. The quality is really good. 👍🏻

(I haven't tried this LoRA yet, but the phenomenon exists in the base model.) I've found using words like silent/silence/silently to work better than phrasing like "lips don't move". Also try adding 'talking' to negative prompt if you're CFG>1.

Tried to play a bit with strength and add other action with higher strength that could override it like smiling?

best counter to that I've used so far is to describe the woman as "silent" in the prompt, and say she's silently doing something

@sam7244211886 @imonaboat1 Thank you for the idea. I will try it

@haidensd58757 I've now had a chance to try it and it definitely is harder to tame in this LoRA than in the base model, but it does help.

@f1985 yeah makes sense!

All T2V loras work for me VERY WELL in I2V or even better than those trained in I2V.

@rafaaguiar990170 I found that when I do that, it works, but it starts to loose the resemblance to the original image. Maybe just me?

@rollpollution193 No, it isn’t just you. Every time when a t2v model is posted there is always someone who says that it works with i2v as well because of bla bla but they never tested it themselves, at least in the most cases.

@SD_Agamemnon yeah they work but very poor, most change faces and stuff like that.. a I2V version is always better

@SD_Agamemnon I use T2V all the time in my videos. you can feel free to look at them, unfortunately Civitai doesnt let me credit the T2V LoRA I use because once you select T2V or I2V it locks you out of including LoRA from the other type in the resources selection (basically every video that has boobs in it I use a T2V Breast Helper LoRA for example, and most of the handjob videos I basically have to use one as well). There are many reasons that I2V videos warp or lose coherence and usually it has far more to do with temporal and spatial attention. Spatial attention dictates how the model tracks physical coherence in a frame and temporal is how it tracks coherence through time. If these settings are either wrong or simply not set at all (and this can be extremely exaggerated by using self-forcing speed LoRA which require you to use a CFG of 1.0 which knocks your negative conditioning down to half or lower strength) the coherence and the ability of the model to 'hang on' to its initial image or physical shape is greatly impacted. Since T2V models are trained to do what T2V does, if you don't have the proper guardrails in place to prevent this warping and changing then it will most likely happen, particularly if you have the LoRA set to 1.0 strength.

I recommend implementing a NAG (Normalized Attention Guidance) node and a UnetTemporalAttentionMultiply node after your Model Sampling Shift nodes and looking into how these parameters work and how to work with them instead of against them or leaving it up to chance. This can go a LONG way just in terms of general video quality and it really makes a huge difference if you are using a T2V LoRA in I2V.

@f1985 i tried your example prompt but dosnt work i need help i use wan 2.2 on pinokio

@f1985 i use pinokio please help me

@f1985 I will do that and format the videos or do whatever you like, just let me know what needs to be done.

1. You should search for and download a POV LoRA. They're out there. You can't get a LoRA trained to do one thing, do something else.
2. it's the exact same process - just drag the video on to the canvas. Not everyone includes the metadata in the video files though, so that only works some of the time. I would say, at best, half the videos I download from CivitAI have metadata included. Also, some of the files that do include metadata don't even use the LoRA they are attached to. It's a box of chocolates, for sure.

i've also found which works for me is to add to your prompt simply "pov shot" no quotes, if using wan2.2 it's very well designed also for camera control