Build the World First: Cinematic AI Films With Seedance 2.5

Most people open a video model, type a description of a scene, and hope. Then they generate again, and again, and the character's face changes every time, the room changes every time, and nothing cuts together. The fix is not a better sentence. The fix is that a cinematic sequence is an assembly job, and you have to build the parts before you assemble them. Seedance 2.5 accepts up to 50 reference images and a single prompt up to 30 seconds long, which means the model will happily hold your world steady for half a minute. It just needs a world to hold.

This guide walks the same production in order. A story, then a cast, then the props, then the world, then one prompt that stitches all of it into fifteen shots. The order is the whole method. Every step you skip shows up later as a face that drifts or a room that changes color mid-scene.

Assets are expensive and story is free, so settle the story first. Ask a chat model for five one-line ideas for a 30 second cinematic short in your genre, with your constraints stated up front: how many characters, what the setting is, what the relationship is. Five is the right number. One idea gives you nothing to compare against and ten turns into a shopping trip.

Then pick on a specific criterion, not on which one sounds coolest. Pick the idea that gives your assets something to do. In the worked example the winning idea was a decades-long voyage where a man collects the same tray of food from the same robot every single day. That premise was chosen because it forces the hero to age, which turns one character into three character sheets and gives the film a visible arc without a single line of dialogue. A story that needs three versions of one face is a better story for this medium than a story that needs three different faces.

A character sheet is a single image of one person, isolated, lit evenly, shot against a plain neutral gray field. It is not a scene. It is a reference. That distinction is doing more work than it looks like it is doing.

The gray background is the tell. Video models treat an image with a background as a place, and they will try to reproduce that place, or blend it with the environment you actually asked for. An image with no place in it reads as a person, not a location, so the model takes the face, the build and the wardrobe and leaves everything else to the scene. Put your hero on a spaceship in the character sheet and you have just given the model two competing spaceships to reconcile.

What the sheet prompt actually needs from you is short. Age, build, hair, face, any distinguishing marks like scars or tattoos, and the clothing and footwear head to toe. Anything you leave blank gets filled in for you, and a blank you did not think about is a blank the model decides differently on every regeneration. Two settings matter on the image side: at least 2K resolution so the detail survives being used as a reference, and 16 by 9 so the frame is already cinematic.

• Age and build, because they set the silhouette and the model will not guess consistently.
• Hair and face, including eye color, because eyes are the first thing that drifts between generations.
• Distinguishing marks, because a scar or a tattoo is a free consistency anchor across shots.
• Clothing and footwear head to toe, because a half-described outfit is a wardrobe change waiting to happen.
• Nothing about the location, because location belongs to the environment plate.

The moment you need the same person at two ages, or in two outfits, you have a continuity problem. The answer is not to write a second description from scratch and hope the two faces match. Generate the base sheet, approve it, and then feed that approved image back in as the reference for the variant. Ask for the same man in his sixties, same man in his eighties, same wardrobe, and the outfit and the bone structure carry across.

In the worked example that produced three sheets of one hero, at 35, at 60 and at 80, all wearing the same silver reflective suit and boots. The film reads as one life because the suit never changed. If each age had been described independently the suit would have shifted shade, the boots would have changed shape, and the audience would have read them as three different men.

The same mechanic is how you put a real person in the film. Upload a few photos of the face you want alongside the sheet prompt, and the model builds the sheet around that likeness. From there every variant descends from the approved sheet, not from the photos.

Here is the failure that actually happened in the build. The brief called for a female artificial intelligence robot who works as a lunch lady. The first sheet came back beautiful and completely human. No panels, no seams, no metal. A person in an apron.

The cause was upstream of the image model. The written prompt had described a humanoid character and had never insisted on the material, and the prompt was approved without a close read. The image model did exactly what it was told. The image model is almost never the thing that is wrong. When an asset misses, the miss is nearly always a property you assumed was implied and never actually specified.

The fix took one sentence back to the chat: she reads as a human, make her a true robot, visible metal, separate panels, still beautiful. Note the last clause. When you correct a property, restate the properties you want kept, or the correction overshoots and you trade a human-looking robot for a menacing one.

Anything the camera sees repeatedly across the film is a character, whatever it is made of. In the worked example the recurring object is a tray of futuristic cafeteria food that the hero collects every day for forty years. It appears in most of the fifteen shots. If it is not pinned down it becomes a different meal every time, and the joke of the film, that nothing changes but him, stops working.

A product sheet is built exactly like a character sheet: one object, isolated, neutral background, close up, described in terms of materials, colors, arrangement and scale. The rule for when to build one is simple. If an object appears in more than one shot and the audience is meant to recognize it as the same object, it needs a sheet. If it is set dressing that passes once, let the environment produce it.

This is the part worth taking away even if you never touch this exact toolchain. Once you move into the video model, your characters and your props do not keep the lighting they were generated with. They take on the light, the color and the texture of the environment you dropped them into. The plate is not a backdrop, it is the light source for everything else in the film.

That inverts the usual priority order. People spend an hour perfecting a hero's face and thirty seconds on the room, and then wonder why the film looks cheap. The face was never going to carry it. A merely fine character inside an exceptional plate looks cinematic. An exceptional character inside a flat plate looks like a cutout on a background.

So the plate gets described harder than anything else you make. Architecture and layout. Materials and surfaces, down to whether the metal is brushed or polished. What is on the walls. The quality and direction and color of the light. What the camera can see from where it stands. And the camera and lens itself, because the lens choice is what makes it read as a film rather than a render.

• Give the model the story before you ask for the plate. A room written to serve a premise beats a room written to a description.
• Build the interiors that host the action, and an establishing exterior if the film needs to say where you are.
• Increase the batch size on plates specifically. This is the one asset where picking a favorite from four is worth the credits.
• Name the camera and lens. A 40mm large format look was the choice in the worked example, and it is a reliable default for narrative scenes.
• Frame interiors three quarter rather than dead on, so the video model has depth to work with.

Seedance 2.5 takes clips up to 30 seconds and up to 50 reference images in a single generation. The current ceiling is 720p, which is a real limitation for a finished deliverable and a non-issue for everything up to that point. Upload every sheet and plate you approved, then write one prompt that covers the whole film.

The prompt has two halves and both matter. First the premise, in plain prose: what kind of scene this is, what happens start to finish, what the feeling is. These models are smart enough to use a premise to make sensible choices in the gaps, and a prompt without one produces fifteen technically correct shots that do not add up to a story. Second the beats, written as explicit timestamps. Zero to two seconds, two to four, four to six, all the way out. Timestamps are how you turn one clip into a cut sequence rather than one continuous drifting take.

Fifteen beats across 30 seconds was the shape in the worked example, roughly two seconds a shot. Each beat names what happens, where it happens, who is in it, what they are wearing, which props are present, any dialogue, and the look. Every character and prop is referenced by its uploaded asset rather than re-described in words, so the model matches the image instead of re-inventing from a sentence. When the prompt is pasted in, the asset tags resolve against the images already loaded in the session.

1. 1Upload every approved sheet and plate to the generation before you paste anything.
2. 2Open with the premise paragraph so the model knows the arc it is serving.
3. 3State the global look once: camera, lens, grade, and how often to cut.
4. 4Write the beats in order with explicit second ranges that tile the full runtime with no gaps.
5. 5In each beat, point at assets rather than describing them again.
6. 6Read the beats back as a list and check that the story still lands with the visuals stripped out.

You can cut a 30 second clip into as many beats as you like, down to one per second. That freedom is the main decision in the whole process and it is easy to get wrong in both directions.

Long beats of four to six seconds are right when a shot has to breathe: a performance, a slow reveal, a held emotional moment, anything where the audience needs time to read a face. The tell that your beats are too long is that motion inside a shot starts to wander and the subject slowly stops looking like itself, because the model is improvising past the end of your instruction.

Short beats of one to two seconds are right for montage, for time passing, for energy. The tell that your beats are too short is that nothing registers. The audience never spends long enough in a frame to understand what they are looking at, and the sequence reads as a trailer for a film rather than a film. For a first attempt, two seconds a beat is the safe default, with the two or three beats that carry the emotional payoff stretched to four.

Almost every bad result in this workflow traces back to one of five causes, and each has a distinct signature. Learning to read the signature is what turns a frustrating afternoon into a ten minute fix.

• A face or outfit that changes between shots means the sheet was too vague or the variant was written from scratch instead of from the approved sheet. Fix the sheet, not the scene.
• A character that looks pasted onto the background means the plate is weak or the sheet had its own scene baked into it. Regenerate the sheet on neutral gray.
• A scene that ignores half your instruction usually means the beats overlap or leave gaps. Timestamps that do not tile cleanly get resolved by the model, not by you.
• A prop that mutates across the film means it never got a product sheet. Anything recurring is a character.
• A sequence that looks good and feels like nothing means the premise was missing. Beats without an arc are a slideshow.

There is one more failure that is worth naming separately because it costs the most money. Generating the full sequence at final settings before you have looked at a cheap version of it. Every structural problem in the list above is visible in a low resolution pass. Find them there.

Everything above is a manual process and it works entirely by hand. It is also the same seven moves every single time, which is exactly the shape of thing worth turning into a skill. Inside the club there are two that do it.

The first one builds assets. You type a slash command and it interviews you: are you making a character, a product or an environment, what is it called, describe it, anything else it should know. Then it calls the image model directly through the connected integration and hands back a finished sheet. The second one handles the scene: it takes your approved assets and your premise, and it produces the full timestamped beat prompt ready to paste.

What the skills remove is not the thinking, it is the retyping. The premise, the aging arc, the choice of which prop deserves a sheet and how long each beat runs are all still yours, and they are the part this guide is about. The skills and the exact prompt text live in the club along with everything else we build.

The shortest honest path from finishing this page to having a real 30 second film. Budget about two hours the first time and under one hour after that.

1. 1Ask a chat model for five one-line ideas for a 30 second short in your genre, and pick the one whose premise forces a visible change.
2. 2List your cast and your recurring props. Most short films need two characters and one object.
3. 3Build one character sheet per character on a neutral gray background at 2K or higher, 16 by 9, and download each one.
4. 4Build any variants by feeding the approved sheet back in as a reference, so wardrobe and face carry across.
5. 5Build a product sheet for every object that appears more than once.
6. 6Paste your story in and ask for the environment plates, then generate the main interior at a higher batch size and pick your favorite.
7. 7Open the video model, upload every downloaded asset, and confirm they all loaded.
8. 8Write the premise paragraph, the global look line, and 12 to 15 timestamped beats that tile the full 30 seconds.
9. 9Generate one cheap low resolution pass and watch it for structural problems only, ignoring texture and detail.
10. 10Fix whatever the cheap pass exposed at the asset or beat level, then run the final generation once.

The last thing worth understanding is that you did not just make a film, you made a set. The sheets and plates are files. They load into the next generation exactly as they loaded into this one. Your second film in the same world costs you a premise and a beat list, and nothing else.

That is what makes this workflow worth the front-loaded effort. The first sequence feels slow because you are paying for the whole world up front. Every sequence after that is nearly free, which is the point at which making short films stops being a novelty and starts being something you can actually do on a schedule.