Muse image-to-video: the first frame is a coin flip, aspect ratio follows the reference
Testing 22 Muse image-to-video generations against a character sheet, the author found 9 of 18 opened on the literal reference image even when the prompt asked for a different outfit and location, and regenerating the same prompt flipped 3 of 4 bad ones. The aspect ratio is decided by the reference image, so cropping or padding the reference to the target ratio before generating is the practical fix.