generally the T2V loras will work on I2V, but some will need better prompting than others. Even in the description it says it doesn't always work, so it's always good to try one without the lora and one with, and find a strength or better prompts along the way.

@makiaevelio543 It seems some creators have better luck training against the i2v model first. While training t2v does work for i2v to an extent, training against i2v works better for i2v and also works great for t2v.

Training bespoke versions for each has the best results for the respective model, but that is a lot to ask of creators. If a creator is going to train just one and it can apply to i2v, I think training i2v is the way to go.

@playnproto266 I believe the opposite, train on t2v first. There’s a thread on the diffusion-pipe GitHub where we’re discussing this topic. I generally train on t2v and it works on i2v. I haven’t had a Lora that didn’t work that way.

@jalenbrunson one example that comes to mind is the cumshot lora. It was trained on t2v and when trying to use it on i2v the effect was minimal and difficult to trigger. It was retrained with the same settings and dataset on i2v and it's transformational.

It's entirely possible some datasets work well for both with just t2v training and some work well for both with just i2v training, some need individual training. It may come down to the concept and wan's base knowledge of that concept.

@playnproto266 appreciate the info. The 3 loras I have posted seem to work well for both, but I’m going to take one of them and do some testing with I2V to see how much better it can get.

@playnproto266 The creator did say that the original data set was full of images that already had cum on it. Starting with a base image that isn't in the training set may make it harder to trigger (there are two, maybe Im thinking of the facials lora).

@makiaevelio543 That makes sense. An action that has an action and result might be more difficult to train than an acton by itself.