Spent the evening trying to get Leoma to follow an actual shot list ✨ The clips are genuinely beautiful 😍 but it keeps ignoring half of what I wrote. I ask for a slow push-in and get a static frame, I ask for two people and get one 🤔 And the more constraints I add, the worse the image gets. Is everyone just accepting whatever it hands back?
That tradeoff is real and it is not you doing anything wrong. Prompt adherence and aesthetic quality get tuned against each other, and most public models are tuned to win on a single gorgeous clip because that is what gets shared. Camera motion in particular is very weakly conditioned in most of them — the words are in the prompt and almost nothing in training tied them to actual movement.
for camera moves i stopped asking in words. generate a still you like then drive the motion outside the model, even a slow crop and zoom in an editor beats fighting the prompt for an hour
两个人这条我笑了 😂 我要三个人 它给我两个半 那半个在角落里飘着 🔥 后来学乖了 一个镜头只提一个要求 想要复杂的就分镜生 再拼
One requirement per shot and assemble afterwards 🙈 That reframes the whole thing as storyboarding rather than prompting, which honestly makes much more sense. Trying it that way tomorrow, thank you both 💛
