Question about prompting on static elements

#81
by leonnn1 - opened

I am trying to build the prompts based on the official articles from Minimax-H3 but I have a question, I have this scenario that has static characters and elements on the start and end frames and I want to the leaves to blow and ignore the end image but only for the leaves. how can I do that?

I tried to make the leaves a subject and ask the model to ignore them from the second picture but so far, regardless of how I rephrase it, these elements keep static, is there a way to circumvent that?

leonnn1 changed discussion title from Question about prompting to Question about prompting on static elements

One thing I would test here is whether the end-frame conditioning is unintentionally telling the model that the leaves are part of the state that must remain fixed.

The official guide describes I2VA as:

first-frame anchor β†’ action onset β†’ continuous development β†’ result

while FL2VA is more like:

first-frame state β†’ observable intermediate changes β†’ progressively narrowing differences β†’ last-frame state

So if both keyframes show the leaves in almost the same position, I would avoid describing them as a preserved/static subject.

Instead, make the motion path explicit:

The architecture, characters and camera framing remain stable.

The tree leaves are the only strongly animated environmental element.
A gust of wind pushes the leaves toward frame right during the first half of the shot.
Individual branches and leaf clusters move with different amplitudes rather than as one rigid object.

The wind gradually weakens during the second half.
The leaves continue with smaller residual motion and settle back toward the arrangement shown in the final reference frame.

The important distinction is:

  • preserve the scene / character / composition
  • animate the temporal state of the leaves

rather than asking the model to preserve every visible element from both reference frames.

If the final image is not actually necessary for the shot, I would also test the same concept as I2VA using only the first-frame anchor. That removes the second visual constraint and may give the foliage more freedom to move.

If the final frame is required, keep FL2VA but describe the leaves' intermediate displacement and their return path explicitly.

The official guide also suggests FL2VA generally works best as a continuous single shot, so I would avoid adding cuts while testing this.

Sign up or log in to comment