Can you make a Halloween animation without knowing how to draw or build 3D models? After seeing how multiple images could be connected into a video in the Kling 4.0 feature announcement, I wanted to try it. Cute AI shorts can make the process look easy. I was curious whether I could get one to tell the story I had in mind.
I made a Halloween short about a baby red panda sneaking chocolates meant for trick-or-treaters. I prepared images for each scene, generated the video with Kling 4.0 on Higgsfield, then brought it into VDIT to add music and sound effects and refine the sequence. The images became a moving short, but getting the story to flow took revisions and editing. Here are the images, part of the prompt, and what changed along the way.
A Halloween story that starts with “just one”
The short is called Just one candy. A baby red panda helps Mom wrap Halloween chocolates. When she steps away, the little one tries a caramel-filled chocolate. One turns into more, wrappers pile up on the table, and the doorbell rings. The baby hands the last two chocolates to the trick-or-treaters, but Mom returns to find the evidence on the table.
I went for a cute Halloween feel, with a witch hat and orange and purple wrapped chocolates. The aim was a short animation whose emotions would read without much dialogue: temptation, panic about getting caught, and finally being found out.
What images do you need for an AI Halloween animation?
I started by defining the characters and setting in images: a baby red panda in a purple witch hat, Mom in an apron, a warm kitchen, and chocolates in Halloween wrappers. These became the references for keeping the characters and props consistent from scene to scene.
Once I had an image I liked, I thought about how it would connect to the next one. Who is holding what, and what changes next? Those details needed to be visible.
In the eating scene, the baby’s mouth needed to touch bare brown chocolate, with the wrapper removed. For the surprised expression, both paws went on the cheeks so the reaction would read through the pose as well as the eyes and mouth. At the door, the number of chocolates left on the plate showed how far the handover had progressed.
For expressions, I chose references for the face and pose before generating the images. That helped me check whether the emotion was clear even before the character moved.



Can multiple images become a continuous story?
On Higgsfield, I placed the images in story order in Kling 4.0’s Frames input and added a prompt describing the actions in each scene. The first attempt used six images to generate a 20-second video. The broad sequence came through: eating in the kitchen, reacting to the doorbell, and meeting the children at the door. But it was unclear how each child received a chocolate, and the wrapping and eating scenes looked awkward in places.
For the revision, I broke the handover into three visible states:
- Before the handover: both the orange and purple wrapped chocolates are on the plate.
- After the first handover: the orange one is in the rabbit’s basket, and only the purple one remains on the plate.
- After the second handover: both children have a chocolate, and the plate is empty.
I also prepared a scene of Mom leaving the kitchen so she would not seem to disappear between shots. With revised images and prompts and a longer duration, I generated a 24-second video from nine images.
The order of the handovers was clearer in this version. Some moments, including the release of a chocolate and the final expression, still felt awkward. Since I changed the prompts and duration as well as the image count, I cannot attribute the improvement to using more images alone. The aim was to show the intermediate states the viewer needed to follow the story.




What did I put in the Kling 4.0 prompt?
The images established how each scene looked. The prompt described what should happen within it and in what order. This is part of the eating scene from the full prompt I used:
Baby holds ONLY bare brown chocolate; opened foil stays on plate. Hold cautious rightward glance briefly then look at chocolate, take ONE small bite, lower it to reveal caramel, eyes close with pleasure.
The baby holds an unwrapped chocolate while the empty wrapper stays on the plate. After a brief glance to the right, the baby takes one bite, lowers the chocolate to show the caramel, and enjoys the taste.
I specified the object in the hand, the wrapper’s location, the direction of the glance, and one action at a time. Even detailed instructions did not always appear exactly as intended. I still had to check the generated video to see what the character actually did.
How can you manage credits while testing?
I tested the combined story at 720p, focusing first on expressions and scene connections and checking the displayed credit cost before generating. Costs at different resolutions can vary with the model, video duration, and service terms.
If you want to increase the resolution of a result you like, Higgsfield’s video upscaler is an option to consider afterward. I have not upscaled this project. Include the upscaling cost when comparing options; generating at a lower resolution and enlarging it later is not necessarily cheaper overall.
How did I edit the AI video’s sound and scenes in VDIT?
I imported the generated 24-second video into VDIT on an iPhone. In the first edit, I adjusted the speed of the opening. Then I worked on the music, sound effects, and repeated scene.
I added The Marching Fairies as background music. There was a rustling sound at the beginning that I wanted to remove, so I cut that section into a separate clip and muted just that part. This let me remove the unwanted sound without muting the entire video.

I made the chocolate-eating scene play twice and added a transition. Repeating that moment is one way to show how “just one” turns into more. The order and repetition of existing footage can help develop a story alongside revisions to the generated scenes.

I also created and inserted a sound effect of children saying “Trick or treat.”
I made this with an early version of the sound effect creation feature planned for VDIT v4.5.0. After the update, I’ll cover how to create a sound you cannot find and confirm its availability in the public release.
To finish, I added a Happy Halloween greeting and a VDIT app end card. The final edited video is about 31 seconds long.

Was making an AI animation really easy?
I was able to turn prepared images into a short animation without building 3D models or drawing each frame. Telling the intended story still meant planning where the character was looking, where the props were, and what changed after each action.
If you are trying this for the first time, start with one small incident. A chocolate someone meant to eat just one of, a sudden doorbell, and wrappers left on the table can be enough to tell a story. I revised the input images and edited the generated footage with music and sound effects. This Halloween experiment helped me see which decisions I still needed to make myself.
