How Do You Make a 4K AI Short Film? NadouPro 4K Short-Film Production Breakdown: From Assets to Final Cut

What does it take to make a one-minute short film with Seedance 2.0’s 4K capability? The answer is three elements: props, characters, and scenes. Build them first as assets, and video generation becomes a natural next step. This article fully breaks down the production workflow for a short film, including prompts for each scene and key techniques.

Project Overview

Short-film concept: a young man sits on a sofa watching TV, completely immersed in the 4K image, while his mother keeps calling him to eat but he refuses to get up.

Overall workflow: the Canvas Agent writes the prompts, and NadouPro generates the content. Each scene first needs three types of assets—props, characters, and scenes—after which video generation is almost automatic.

Scene One: Character, Mother, Room, and Remote Control

Four assets are needed: the main character, the mother, the living room, and the remote control.

Character (new outfit). With an existing character reference image, simply have the Canvas Agent write an outfit-change prompt—Y2K style, a pink T-shirt, and a teddy bear hanging at the waist. Generate it with Qizhi-IGI 2 at 4K quality.

Two key points: first, a gray background—after extensive testing, character reference images on a gray background perform much better than those on white or black. Second, make sure there is one image with a clearly visible face and one full-body image—this way, Seedance does not need to guess anything.

Then click "Create Asset," name it, and save it. Do this for every character, scene, and prop—the Canvas Agent will automatically reference these assets in the prompts.

Mother. Attach the main character reference image and have the Canvas Agent write a reference image for a 40–55-year-old mother in casual homewear, with facial similarities to her son—three angles, on the same gray background.

Scene context: a mother character reference for the young man in reference image 1. She must read instantly as his mother—the same ethnicity and facial DNA, from an older generation. The three photos are shot in a clean light-gray cyclorama studio: facial close-up, full-body front view, and full-body back view.

Scene. A modern open-plan living room with a teenage atmosphere, wide angle from a three-quarter angle.

A modern open-plan living room with a teenage atmosphere, wide angle from a three-quarter angle. A spacious room that also functions as a teen activity area—a bit messy, not a showroom. A 65-inch wall-mounted TV is the main visual focus, screen off. A comfortable sofa with crooked cushions, beanbag chairs. Hoodies and several T-shirts are thrown around, sneakers kicked beside the rug. Fairy lights and LED light strips glow along the walls. Natural window light plus warm table lamps. Photorealistic, 2020–2022 interior style. 16:9.

The key phrase is "three-quarter angle"—it shows more room information, gives Seedance better depth perception, and keeps it from breaking anything during rendering. We always use the image generation module for scene images; it is designed for cinematic shots. Video quality mainly depends on this one image—check every detail, and regenerate it if anything feels inconsistent.

Remote control (prop). Three-view image of a minimalist smart TV remote control, gray background.

Scene Two: Fantasy Dragon Flight

A fully AI-generated fantasy segment—a woman rides a dragon through the jungle.

First build the dragon asset: use Seedream 5.0 Pro to generate the dragon’s reference images. The key is to generate two close-ups, one with the mouth open and one with the mouth closed, to prevent facial errors when expressions change. Then generate the female character asset. Finally, write the video prompt.

Prompt essentials: the dragon’s movement has three modes—when diving vertically, its wings tighten into a streamlined shape; when passing through the forest, its wings spread out and glide still; when climbing, its wings flap powerfully. The slow-motion butterfly close-up is the highlight: time smoothly decelerates from normal speed to ultra slow motion, holds for about two seconds, then snaps back to normal speed.

Iteration loop. Generate → find the part you are not satisfied with → tell the Canvas Agent exactly what to change → regenerate. Watch the waterfall mist and the wing position when the dragon touches the water—this is the iteration loop at work.

Scene Three: Immersed in the Screen

Back to the room. All assets already exist, so only the intention needs to be described: the character is locked onto the TV, his mother calls him from off-screen, he mutters "wait a second, Mom," and then changes the channel.

Scene context: evening, same living room. @ADIL sits on the sofa, completely lit up by the TV image. His mother calls him to dinner from the off-screen kitchen; he is too absorbed to leave, mutters a delayed reply, picks up the remote control on the table, changes the channel, and keeps watching—he does not stand up. Handheld shake throughout, low angle, fixed distance for each shot, with cropping changing only at edit points. Shot 1 (0–6 seconds): Medium Shot, the character leans forward, eyes locked on the off-screen screen, screen light flickering across his face. Shot 2 (6–11 seconds): facial close-up, still locked on the screen and smiling; the off-screen voice urges him again, and he chuckles softly while muttering a reply. Shot 3 (11–15 seconds): Medium Shot of the hands, picking up the remote and pressing it; the color of the screen light changes, and he leans back into the sofa to keep watching.

Zoom in on the face: pores, every eyelash, natural hair, even forehead wrinkles. It does not look AI-generated—it looks as if it was shot with an expensive professional camera.

Scene Four: Wildlife Documentary

An ultra-detailed nature documentary shot. First build the scene—a wide-angle snow mountain—then generate the video: a telephoto shot of a snow leopard descending a hillside.

Key Lessons

Assets first. Build the three types of assets needed for each scene—props, characters, and scenes—before generating the video. The Canvas Agent will automatically reference named assets.

Gray background. Always generate character reference images on a gray background; the results are much better than white or black.

Three-quarter angle. Describe scene images from a three-quarter angle. Compared with a frontal view, it shows more spatial information and helps the system understand depth more accurately.

Facial close-up and full-body image. Make sure each character has one image with a clearly visible face and one full-body image, so Seedance does not need to guess.

Iteration loop. Generate → find the problem → make a precise change → regenerate. Do not expect success in one try, but also do not regenerate blindly—tell the Canvas Agent exactly what to change.

Slow-motion rhythm. Insert slow-motion passages into continuous shots: time smoothly decelerates, holds for several seconds, then snaps back to normal speed. This rhythmic contrast is an effective way to attract attention.

FAQ

Q: How many assets are needed to make a 4K short film?

For a one-minute short film, about 10–15 assets are needed: 2–3 character reference images, 3–4 scene images, 2–3 prop images, and possible creature reference images. Assets first is the core principle—build the assets before generating the video.

Q: Why use a gray background for character reference images?

After extensive testing, character reference images on a gray background perform far better than those on white or black in facial matching and clothing extraction. Gray provides the model with a neutral reference and does not interfere with feature recognition.

Q: What angle is best for describing scene images?

A three-quarter angle. It shows more spatial information than a frontal view, gives the system better depth perception, and avoids breaking the spatial structure during rendering.

Q: What role does the Canvas Agent play in a short film?

The Canvas Agent converts natural-language intentions into detailed video generation prompts. You describe the scene concept and character actions, and the Canvas Agent generates a complete prompt that includes camera movement, lighting, physics, and audio descriptions. Named assets are automatically referenced.

Q: How do you create a slow-motion passage?

Describe the time change in the prompt: "time smoothly decelerates to ultra slow motion, holds for about two seconds, then snaps back to normal speed." Seedance 2.0 supports timeline control, and details are enlarged during slow motion.

Q: How do you iterate if the generated result is not satisfying?

Do not regenerate blindly. Find the specific part you are not satisfied with, such as "the dragon’s wing position is wrong when it touches the water," tell the Canvas Agent exactly what to change, and then regenerate. Change only one variable each time.

For more features, you are welcome to try NadouPro: NadouPro

Contains AI-generated information that may be incorrect. Please verify independently.