An image can tell a model who a character is, or show exactly where a clip should begin. Those are different requests. Before uploading a reference, write one sentence about what you want to preserve.
Identity or an opening composition?
Imagine a picture of Lin standing outside a bookshop. If the next shot should show the same adult character sitting behind a counter, the picture mainly supplies identity. If the clip should begin at the door and follow her inside, first-frame guidance fits the intent more closely.
A last frame describes a destination: Lin is seated and the paper bag is on the counter. It adds an endpoint to satisfy, but it does not explain how she gets there.
The MiniMax H3 input documentation separates first/last-frame inputs from multimodal references. Those modes are mutually exclusive in that interface. Always check the combination supported by your selected model and form.
Separate what stays from what changes
Use one sentence for the information the image supplies and another for the new action. These original examples have not been generated.
For an identity reference:
Image 1 supplies adult character Lin’s face, hair and clothing. In this new shot she sits behind the bookshop counter and looks up at a customer to frame-right. Use a frontal medium shot with bookshelves behind her. The doorway pose is not the opening composition.
For a first frame:
Begin with the selected doorway image. Lin keeps her clothing and the paper bag in her hand. She pushes the glass door open and takes two steps inside. The camera stays on the street side; the counter becomes visible through the open doorway.
The two versions ask for different shots. A difference between their outputs is not, by itself, a failure to follow instructions.
Make every reference useful
If you need a person, a room and a product, identify the role of each reference. Then look for conflicts: different costumes, two versions of a package or room images with the door in different places.
Resolve those conflicts before generating. A model’s reference limit is a capacity limit, not a recommendation to fill every slot. Add a reference when it supplies information that is missing.
Budget the action between endpoints
If the first frame shows someone at the door and the last shows them across the street, the movement needs time. Turning, speaking, closing an umbrella and crossing the road may be too much for one short clip.
Describe a continuous path between the two frames. If that path is hard to explain, split it into smaller shots with clear starting and ending states. That also gives the editor useful replacement points.
Check the request after switching modes
In JMovie, select the model, confirm the input mode, then add the media. Check that each file appears in the current reference area and that the prompt mentions the right file. After changing modes, review the selected inputs again.
The same distinction matters for video. Borrowing motion or style from a clip is different from extending its timeline. Uploading a previous clip does not automatically mean “continue this exact video”; confirm that an explicit extension operation is available before relying on it.
Return to the practical question: are you preserving information about the subject, or fixing the opening and closing pictures? Make that choice before designing the action.