How To Make a Game Trailer with AI (Full Workflow + Prompts)
How to make a game trailer with AI: locked character sheets, seven locations, a four-act shot list, and two 30-second Seedance 2.5 generations, all prompts included.

To make a game trailer with AI, you need three things: locked character and location references, a shot-by-shot script, and a video model that can animate those references consistently. This guide walks through the full workflow, from character sheets to final cut, using Cinema Studio for generation and Marketing Studio to turn the same assets into ad creatives afterward.
Key takeaways
A trailer this size runs on surprisingly few locked references: this example uses 3 designed characters and 7 locations, reused across every shot instead of regenerated per scene.
Not every look needs its own asset. A character in a different outfit, or a location's own built-in equipment, can be described directly inside a scene's prompt when that look only needs to hold within one storyline thread.
Plan the full shot list before generating anything. This trailer breaks into 4 acts across two 30-second generations, about a minute total, decided on paper first.
The whole trailer cost about 391 credits, roughly $20 at a typical rate of 20 credits per dollar, and almost all of it went to the two video passes.
Once the trailer is done, the same locked files go straight into Marketing Studio for a batch of ad variants, no rebuilding required.
What is a game trailer, and what type does this guide build?
A game trailer is a short promotional video built to sell a game. There are a few recurring types: teaser and announcement trailers that introduce a game without revealing much, cinematic or story trailers that dramatize character and world the way a short film would, gameplay trailers that show actual mechanics and moment-to-moment feel, and launch trailers timed to release to convert interest into wishlists and purchases.
This guide builds a cinematic trailer: a fictional example built for a video game named LETA that leans entirely on character, world and atmosphere rather than gameplay footage, since that's the format an asset-first AI workflow is best suited for. If your own game needs a gameplay trailer instead, the same locked-reference workflow still applies, only the shot list changes from dramatized scenes to actual mechanics.
What do you need before generating a game trailer with AI?
Before generating anything, you need:
A concept and a full shot-by-shot script. A list of scenes, dialogue, sound cues and cuts included.
A written description for each character, detailed enough that someone else could picture them from the text alone. That's what keeps a face consistent later.
A locked reference sheet per character, generated once so the same face holds across every shot.
Separate references only for identity changes, like the same character at a different age. A change of clothes doesn't need its own reference, an outfit can be written directly into the scene prompt on top of the same locked face.
Once all of that exists, generate the script scene by scene, in order.
How do you storyboard a game trailer before generating anything?
Write the full trailer as a shot list before generating a single frame, cuts, dialogue and sound included, then generate against that list instead of improvising order later. Most trailers, whatever the story, break into a similar four-part arc:
The hook: something unresolved in the first few seconds, before the trailer explains anything.
The rise: world and characters established, tone still light or neutral.
The turn: tone shifts, tension escalates, stakes become clear.
The climax: ends unresolved rather than tidy, then the title card.
How do you make a game trailer on TAVOA AI?
Cinema Studio also gives you deterministic camera control, and for a game trailer that matters more than almost anything else. You can pick a virtual camera body, lens and focal length, choose specific movements like a dolly, a crane rise or a crash zoom, and stack several camera moves inside one shot, instead of hoping the model improvises something usable. Every lens lock and push-in in the trailer prompts below is written against that control, which is why the shot descriptions can be as specific as they are.
The whole method in one line: plan the story, lock character references, lock locations, write the full shot list, generate the trailer in passes, then finish the cut. Everything below walks that sequence on a real example.
What's the story behind LETA?
LETA is a fictional near-future detective game built around a memory-editing procedure that's licensed for therapy and quietly used for cover-ups. Kai Weil runs the procedure, clinical and unreadable by training. His current client is Leta Moren, an heiress whose husband Damen Voss ordered the extraction to erase her memory of his affair and buy time before she finds out he's after her inheritance. Partway into the extraction, Kai finds his own face inside a memory he shouldn't recognize, and the deeper he goes, the more it looks like Leta's family history hides something Damen needs buried for good.
Mapped onto the four-part arc above, LETA's trailer runs: a dark, unresolved teaser, a fast rise through Leta's warm years, a turn where the relationship and the family history both start cracking, and a climax in the clinic that ends on a standoff, not an answer.
Who are the characters in LETA?
Kai Weil, the memory editor (@kai)
What it does: generates Kai's reference sheet, the version of his face used in every clinic and memory shot.
Damen Voss, the husband (@damen)
What it does: generates Damen's adult reference sheet, used in every present-day beat, with his suit and frown lines locked so they never drift between cuts.
Leta Moren, the heiress (@leta)
What it does: generates Leta's base reference sheet in her everyday clothes, used in every memory scene outside the clinic.
What locations does the LETA trailer need?
Seven locations carry the story, all generated the same way. A few smaller settings appear in single beats only, the bedroom, the family dining table and the doorway, and those were written straight into their scene prompts instead of being generated as separate assets.
The clinic (@clinic), laid out the same way in every shot so nothing drifts between the wide plan and the close inserts.
The playground (@playground), generates for the earliest memories.
The beach (@beach), generates for the memory where Leta and her friends run toward the water at sunset.
The bar (@bar), generates for the warm night out with friends during Leta's younger years.
The school (@school), generates for the classroom where Leta and Damen first meet, used in both halves of the trailer.
The parents' office (@parents-office), generates for scenes where Leta breaks down.
The backyard (@backyard), the memory the whole trailer is built around.
How do you keep the same face and location across every generation?
Once every asset above exists, keep it in one place so the same file goes into every shot instead of a fresh face per scene. The handles above are your own naming, useful for finding assets fast. The prompts themselves read references by upload position: the first image you attach becomes @image_1, the second @image_2, and so on. That makes upload order part of the prompt, and the two halves of this trailer use different orders, so each one lists its own sequence. Locations work differently again. In this trailer they were not attached as references at all, they are written out in words inside each scene prompt, with the locked location images used as the visual target while writing those descriptions.
One rule matters more than any other with Seedance: it doesn't process negative prompts. Writing “she doesn't smile” tends to produce the opposite. Describe the state you actually want instead.
Type
Quantity in this trailer
Model
Credits
≈ USD (at ~20 credits/$1)
Characters (3 mains plus age variants and parents)
~$0.05
Locations
0.875
~$0.04
Video scenes
2 (30s each)
Seedance 2.5, 720p
~390
~$19.50
Total
~391.875
~$19.59
At roughly 20 credits per dollar, a trailer this size costs approximately $20 in credits, and almost all of it goes to video. Still images are effectively free by comparison: all fifteen character and location references together cost under 2 credits, well under one percent of the total. Video pricing scales with length, 195 credits per 30-second generation at 720p, so your own cost depends almost entirely on how much finished runtime you generate. The $1 ≈ 20 credits rate is a rough approximation, not tied to any specific subscription tier, check the pricing page for the exact rate on your plan.
The trailer cut
Two generations carry the full four-act trailer, about a minute end to end.
Both prompts ask for a deliberately rendered look rather than a photographic one: near-photoreal fidelity with the tells of a real-time engine, subsurface sheen on skin, strand-clumped hair, computed bokeh, an even grain. That is a choice. A cinematic game trailer has to read as belonging to a game, and an image that looks photographed reads as a film instead. The character and location references stay photographic, since they only carry identity and layout, and the video prompt sets the final look.
Part 1: the setup (0:00 to 0:30)
What it does: generates the entire first half in one pass, the teaser flashes, the clinic reveal, the interface prompt, Kai's glance, the warm montage and the exchange with Damen. Because references attach by upload order, load them in exactly this sequence: 1 Kai, 2 Leta in her mid-twenties, 3 Leta at 29, 4 Damen, 5 Leta at six, 6 Damen at sixteen.
Part 2: the reveal (0:30 to 1:00)
What it does: generates the second half in one pass, first love, the fight, the family beats, the accusation, the backyard reveal and the standoff, ending on the title card. The upload order changes here: 1 Kai, 2 Leta at 29, 3 Damen, 4 Leta at sixteen, 5 Damen at sixteen, 6 Leta's parents.
How do you finish the cut?
Generation isn't the last step. The two passes come back as separate 30-second clips, so the finish is ordinary editing: watch both passes, regenerate any beat that missed (retries on a pass cost the same 195 credits, so a failed beat is cheaper to fix at the prompt level than to accept), join the halves on the hard cut between them, check that the clinic grade matches across the seam, confirm the two dialogue lines and the monitor beep survived generation, then export. The title card and call to action are already part of the second pass, so nothing needs to be added on top.
What else can these same assets do?
The reference library was already built once for the trailer, and the same characters and locations move directly into Marketing Studio for a batch of ad creatives, no rebuilding required. Same reference set, so every version still reads as the same game.
That's what an AI-native creative suite buys here: the trailer and the ad batch come out of one pipeline, not two separate projects built from scratch.
How do you enter the TAVOA AI Global Film Festival?
Pro tip: if you want to go deeper into AI filmmaking, start with TAVOA AI Academy. The AI Filmmaking Pipeline course walks the same method as this guide, scripting, asset creation and scene-by-scene prompt engineering, taught on the football drama from How To Direct an AI Short Film Like a Filmmaker. Learn the pipeline on a trailer, scale it to a short film, then take that film somewhere.
The same workflow scales past a trailer. TAVOA AI's Global Film Festival runs August 7 to 31, 2026, with a $1,000,000 prize pool for films of 3 minutes or longer, generated entirely on tavoa.ai. Everything in this guide, locked references, a full shot list, generation in passes, is exactly how a festival entry gets built, a one-minute trailer just needs to keep going. Full rules, prizes and eligibility are on the festival page.
How To Make a Game Trailer with AI (Full Workflow + Prompts)
Got any questions left?
You lock the face as a reference sheet before generating a single scene, then attach that same file to every shot the character appears in. In TAVOA AI the prompt reads references by upload position, so keep the attachment order identical every time, and write the same physical details into the prompt as a backup.
This one-minute trailer cost about 391 credits, roughly $20 at a typical rate of 20 credits per dollar, and over 99% of that went to the two 30-second video generations. Still references, characters and locations together, cost less than 2 credits. Exact rates depend on your plan, check the pricing page for yours.
A single 30-second video pass generates in minutes, and the references are faster still, so the real time sits in iteration: writing the shot list, reviewing each pass and regenerating the beats that missed. Count one pass plus retries per video half, and expect the writing and reviewing to take longer than the generating.
Yes. In TAVOA AI, once a character or location is locked, it drops directly into Marketing Studio to batch short ad variants, no separate shoot required.
Some video models, including Seedance, don't process negative instructions the way text models do. Writing "she doesn't smile" can produce the opposite result, so it's safer to describe the state you actually want.
A cinematic trailer dramatizes character and world, while gameplay footage shows how the game actually plays. Some studios capture cinematic trailers directly in their game engine, while an asset-first AI workflow builds them from locked character and location references animated shot by shot instead.
by TAVOA AI