Filmathon · Nov 5–6
Creator’sguide
Something to hold on to for the evening, not a rulebook.
How an idea gets a beginning, a middle and an end, what the camera setups mean, and what to write in the prompt to get the picture you have in mind. Use as much of it as you like: the film is yours, and you make it your way.
The evening runs in Hungarian, and we’re happy to help in English. Your film can be in English too.
Where are you starting from?Director’s assistant
Let your Claude walk you through it.
This guide also comes as a conversation. The director’s assistant introduces itself, asks you about the story, the visual plan and the shots, writes the image, video and sound prompts, and works out how many credits you’ll need. At the end it gives you a prompt sheet where you can copy everything with one click. The film stays yours: it asks and suggests.
The assistant was written in Hungarian. Ask it to work in English and it will.
Claude doesn’t generate images: you run the prompts in ElevenLabs Studio (here’s how).
Claude on the web or in the desktop app
- Download the skill.
- In Claude: Customize › Skills, click “+”, then Upload a skill, and pick the file you downloaded.
- In a new chat, write: “Hi, can you help me with the Filmathon? Please work with me in English.”
If you don’t see the Skills menu, turn on code execution under Settings › Capabilities (“Code execution and file creation”). According to Claude’s help center, it works with a free account too.
Claude Code
Paste this sentence into Claude Code and it installs the skill itself:
00Where are you starting from?
Everyone comes from somewhere else.
Pick one and the guide adapts to you: whatever you don’t need right now gets folded away. You can switch any time, and every chapter can be opened.
Jury and mentors
Who’ll be there with you.
On Thursday evening the mentors help you at Puzl. On Friday morning the jury decides in the cinema. If you keep the jury’s five criteria in mind while you plan, your film will hold together better too.
The jury
The jury sees the films for the first time on Friday, together with the audience. They hear the mentors too, but the prizes are the jury’s call. There is no audience vote.
- Patrik PálvölgyiWestendHead of the Innovation Center at Gránit Pólus Group.LinkedIn
- Zoltán BaranyiTelekomFilm
- Norbert GaálAI BudapestOne of the founders of AI Budapest and an organizer of Filmathon.LinkedIn
- A delegate from ElevenLabsElevenLabsFrom the ElevenLabs team. If they can’t make it, the jury present also decides the ElevenLabs Prize.
What does the jury look at?
Every film is judged on the same five criteria. The films are made in one evening, in about five hours, so the jury looks at what you got out of that time, not at the standards of a professional production.
- Story and message
Can you say in one sentence what the film is about? Does it have a beginning, a middle and an end, or a strong emotional arc? How fresh is its take on the theme?
Story → - Visual world and control
Are the character, the style and the light consistent? Do we see deliberate choices rather than the machine’s accidents?
Visual plan → - Film language and editing
Shot size, camera movement, rhythm, a strong first and last shot: what turns generated shots into a film.
Film language → - Sound
Narration, music, noises and sound design working together with the picture.
Sound → - Purposeful use of AI
Does the film show something no camera could shoot, or is the AI just there for show? When you brainstorm, ask yourself: could this be shot with a camera?
Idea →
The five criteria count for all three prizes, and each prize has one criterion of its own. A film can win one prize at most.
- ElevenLabs PrizeFor the film that makes the most of ElevenLabs: image, video, voice, music and text created in one place, in one creative process. 6 months of ElevenLabs Scale for each team member.
- Westend Next Mindset PrizeFor the most modern, open and forward-looking thinking and visual world: a new-generation outlook, openness, fresh thinking, innovation. 3 months of ElevenLabs Scale for each team member.
- TelekomFilm PrizeFor innovative storytelling and visual narrative: not a copy of reality, but a film we’d never have thought of without AI. 3 months of ElevenLabs Pro for each team member.
If you hand in your film out of competition (another generation tool, a bigger plan or extra credits), it can’t win a prize. If the jury finds it outstanding, it can still get a special mention.
Who can you ask on site?
The mentors go from table to table and help with the story, the visual plan, the prompts and the sound. You don’t have to wait for them: if you get stuck, just ask. The best time to talk your plan through with them is between 18:30 and 19:30, before you start generating.
- Bernadett Dian (Betti)Art director, Moonrock Animation StudioCame to 3D animation from painting and production design for film. Has worked as an art director and character designer, and taught 3D courses at the University of Theatre and Film Arts in Budapest. At Moonrock, Betti brings a fine-art approach to AI-based workflows, with a close eye on visual quality and creative control. Co-founder of Magyar Gépmesék, a community for AI animation and film.
Ask aboutvisual world, character, visual plan
LinkedIn - Gábor MarinovFounder, creative producer, Moonrock Animation StudioHas worked in computer animation and visual effects since 1991. Was the first VFX supervisor and producer at Digic Pictures when the studio started, worked on the effects of Terminator 3, and taught production design for ten years at the University of Theatre and Film Arts. Co-founder of Magyar Gépmesék.
Ask aboutfilm language, effects, how shots become a film
LinkedIn - Réka Harsányi DLAMedia designer, university lecturerGraduated from MOME and earned a DLA at the Hungarian University of Fine Arts. Has taught interactive technologies, visual programming and concept design at university since 2010, and AI and visual storytelling at the Ainomalia school. Experiments with AI film, generative animation and hybrid visual techniques every day in the wowsomevisual project.
Ask aboutconcept, AI film workflow, generative animation
LinkedIn - The ElevenLabs mentorElevenLabsFrom the ElevenLabs team. At Filmathon, ElevenLabs also provides the creative tools and the credits: image, video, voice, music and editing all in one place.
Ask aboutthe platform, which model for which scene, sound
01How the build goes
From an idea to a film.
This arc is our suggestion, not a required order. One piece of advice is worth taking, though: plan first and try things out with still images, and only then bring them to life, because motion eats most of the time and most of the credits.
-
118:30
One sentence
The theme is out. What comes to mind? Write one sentence: who the film is about, and what happens to them.
you can tell the whole film in one breath.
Story → -
218:45
Beginning, middle, end
Split it into three parts. Decide on the first and the last image, then build the middle between them.
you know what the last image is.
Story → -
319:00
Shot list
One shot per line: how many seconds, what shot size, what the camera does, what happens, what we hear.
you know what we see in every line.
Director’s sheet → -
419:30
The look
A character sheet (portrait and full body), one empty image of each location, and a style line. You refer back to these in every image that follows.
your character looks the same in two different images.
Visual plan → -
519:45
Starting frames
One key frame for each shot: the character at the location, in that shot’s framing. This is where you pick and fix things, because images are faster and cheaper than video.
you have a good image for every line.
Film language → -
620:15
Motion
Only animate the images you’ve approved. In the prompt, write what changes: the action, the camera, the sound.
you have every shot, plus a spare.
Seedance → -
721:00
Sound
Music, sound effects and, if you need it, narration. In a team, this can run alongside the motion step.
you have the music and every important sound.
Sound → -
821:30
Edit and hand in
Put the shots in order, cut to the music, lay the sound effects underneath. From 22:15, export and email the film to the organizers. The edit suite closes at 22:30.
the film is sent, and you played it once from start to finish before sending.
Submission →
In a team
You’ll spend time waiting on the machine, so work in parallel: while one of you generates images, another works on sound and music, and the third is already editing. Put the plan together as a team, though, so you’re all making the same film.
Where you work: ElevenLabs Studio
A film doesn’t come out of one prompt. One generation is one shot of 4–15 seconds, and you build the film from those. In an ElevenLabs Studio video project everything is in one place: you generate images, video, voice, music and sound effects there, and cut them together on the timeline.
- 1. Your accountLog in to ElevenLabs. On the night of the Filmathon, every creator got a one-month Pro plan with 600,000 credits.
- 2. The projectIn Studio: “+ New blank project”, “Video project (New)”, Create. Set the aspect ratio at the top to 16:9.
- 3. Images firstLibrary, “Create +”, Image: character, location, then one key frame per shot. Images are cheap, so fix things here.
- 4. Then motion“Create +”, Video, Seedance 2.0. Use the approved key frame as the “Start frame”. Ask for one 5-second version at a time.
- 5. Timeline and soundDrag the clips onto the timeline in order. Then “Create +”: Speech for narration, Music, SFX, each on its own track.
- 6. Test export before 21:00Export two clips and check the file: 1920 × 1080, 24 fps, sound included. If Studio doesn’t give you that, edit in CapCut or DaVinci Resolve.
Studio Agent drafts a first cut from your idea. It’s good for putting finished clips in order and adding sound effects. Don’t let it generate your shots: it picks the model itself, and it’s hard to keep track of the credits. In a team, everyone generates in their own account and you edit in one person’s project. Clips from the separate Image & Video page go over with the “Export to Studio” button.
Credit plan: how far will it go?
ElevenLabs gives every creator a one-month Pro plan with 600,000 credits (if you registered after noon on November 4, you buy it yourself). This is the competition budget, and everyone works from it. Video eats most of it, so do the math up front: set the length of your film, how many times you generate each shot, which model you use, and how many of you there are.
Estimated video cost
- Fewer rejectsThree versions instead of five may be enough: time is tight anyway, and it’s fine if not every shot is perfect. Approve the key frame as an image before you animate it.
- In a teamYou can’t pool credits, but you can split the work: someone else can generate some of the shots, the sound and the effects, and you put it all together in one person’s Studio.
- If you run shortLess important shots can go on a cheaper model. You can also buy more credits (top-up), but you have to tell us when you hand in, and then your film screens out of competition: we show it, but it can’t win a prize. The same applies if you work on a bigger plan, or generate images, video, sound or music with any other tool, even a free one.
Prices are for one 15-second, 16:9, 1080p generation with sound, measured in the ElevenLabs interface on October 5, 2026: Seedance 2.5: 93,982, Seedance 2.0: 61,849, Kling 3 Pro: 15,272, Kling O3: 12,726 credits. The calculator scales this per second and only counts video, so leave some room for images and sound. The interface shows the exact price before every generation.
02Story
Beginning, middle, end.
A short film has no room for subplots. You need a character who wants something, something that gets in the way, and by the end, something has changed. That’s enough for the audience to follow.
Say it in one sentence
If it fits in one sentence, the film will come together too. Four blocks, in this order:
With a story
A grandmother on a video call teaches her grandson to make pancakes, but the connection keeps freezing. In the end, the grandson flips his first one in his own kitchen.
Mood piece, no plot
The sentence only needs a place and a feeling: the last night tram runs empty through the city, and the driver opens the doors at every stop.
Split it into three
The proportions are similar at any length. Pick a length, and the blue bar plays out how long that really is.
Beginning
Who is this, where are we, what are they missing? Open with something odd, a beautiful image or a question: the first few seconds decide whether the audience pays attention.
Middle
They try, and it keeps getting harder. Every shot should move things a step forward. If it doesn’t, cut it.
End
A turn, then a closing image that stays with people. Once you have your last image, the film is half done.
An AI shot in the finished film usually runs 4–6 seconds, so for this length you need about 5–8 shots. Plan smaller than you’d like: a finished half minute beats an unfinished minute.
Decide this before you generate
These are your parts. Once you have them, the prompts almost write themselves.
- 1×
Character
Three visible traits and an object, for example a gray bun, a striped scarf, a worn leather bag. Describe them in exactly the same words in every prompt.
- 1–2×
Location
Few locations, many shot sizes. That keeps the world consistent, and you regenerate less.
- 1×
Style
One visual world all the way through: one reference (a film, a photographer, a painter) and three colors. Don’t switch styles halfway through.
- 1×
Sound
Does anyone speak? Narration, dialogue, or just music and sound effects? This shapes the images too.
- 1×
Closing image
Know the last image in advance. Every shot leads up to it.
Short, clean movements work best with AI. Long dialogue and crowded scenes where characters grab and touch each other take more attempts, so tell as much as you can with images, music and sound.
No idea? Five forms
Crutches, not required templates. If you have a form of your own, bring it.
Want and obstacle
Someone wants something, and it isn’t easy to get. The classic.
Twist
The last image rewrites everything we’ve seen.
Mood piece
No plot. One feeling builds: place, light, sound.
Fake trailer
The trailer for a film that doesn’t exist. It’s made of fragments, so you don’t need continuity, which is hard with AI.
One moment
A single second, stretched out, seen from many angles.
03Visual plan
Draw it before you move it.
Design the character, the location and the style as still images first, with an image generator. That way you can see what you want, and the machine has something to look back at in every shot.
Four sheets, in this order
The Grok Imagine team works this way too: character sheet, locations, one visual world, and only then the key frames they animate. If an object carries the story (a pocket watch, a letter, a key), make a sheet for it the same way.
Character sheet 1×
One neutral portrait and one full-body image against a gray background. If we see the character from behind, you need a back view too, and a costume change needs a new sheet. Give Seedance 2.0 these as separate images, not as one multi-view sheet.
Location sheet 1–2×
One empty image of each location, with no character in it, at the time of day and in the light it will have in the film. In the prompt, include the shot size, the time of day, the mood, the camera, the lens and the film stock.
Style line 1×
Start from a moodboard or a style image and write one sentence from it: camera, lens, film stock, colors. Add it word for word to the end of every image prompt, and to the video prompt too if you don’t start from an image.
Key frame 5–8×
You place the character from their sheet into the location: give an image model both images as references. This is the image you animate later. If the still isn’t good, the video won’t be either.
You can start from an image you like: your own photo, a painting, a film still. Several of the ElevenLabs image models accept references (Nano Banana 2.1, for example, takes up to fourteen). Ask it to take the light, the colors and the mood, not the content: use image 3 only as a style reference: its light, colors and grain. Don’t upload a real person’s face, and don’t copy a specific work or a well-known character. If you like a director’s or a photographer’s style, leave their name out of the prompt and describe what the style means instead (composition, camera position, palette, light, era).
How does the world stay consistent?
In three layers. The detailed description isn’t for the video model: the images hold the look together, and only the fixed tags go into the prompt.
- for you, the team and the chatbot
World description
It can be long and detailed: the era, the rules of the world, materials, colors, who’s who. The director’s assistant works from it when it writes prompts. Don’t paste the whole thing into the video model: according to ByteDance’s guide, a full script only confuses it.
- for the machine
Reference images
The character sheet, the location sheet and the key frames. These hold the look together, not long text. Seedance 2.0 accepts up to nine images, but ByteDance recommends four or five references in total. 2.5 takes up to thirty.
- in every image prompt
Fixed tags
Two or three identifying traits for the character, plus the style line, in exactly the same words every time. You need them in the video prompt if you don’t start from a key frame, or if the clip has a cut, something blocking the view, or a change of angle. Otherwise the image already carries them.
How much fits in a prompt?
A detailed prompt helps when the details are about the image: ByteDance’s guide for 2.5 explicitly asks for a director-style prompt organized in blocks. What hurts is the excess: backstory, explanations, piles of adjectives, contradictory requests, or the whole script pasted in. ByteDance doesn’t publish an official character limit. Third-party API docs put it at roughly a thousand English words, which is plenty for one shot.
Goes in the prompt
- the shot size and the camera
- two or three of the character’s traits, or the number of their image
- who stands where, and what they do
- the location and the light
- if several things happen: when (0–4s, 4–8s)
- the style line and the sound
Stays out
- the character’s life story
- the rules and history of the world
- the full script
- descriptions of feelings: show a physical sign instead
- what happens in the other shots
04Film language
What the camera sees, and what goes in the prompt.
The models understand film terms best. Every card shows the same scene with a different shot size, angle or movement. Under each one is the English phrase: click it to copy.
Shot size: how big is the character in the frame?
The same moment in six shot sizes. The closer we get, the more the shot is about a feeling and the less about the place.
Extreme wide EWS
Where are we? The place is the lead, the person is small in it. Good for an opening shot.
Wide WS
The whole figure and the space around it. We see what she’s doing, and where.
Medium MS
Waist up. You see the face and the hands, so most dialogue is shot this way.
Close-up CU
Face and shoulders. This is where we see what she feels.
Extreme close-up ECU
Only the eyes, the mouth or a hand. It’s a strong image, so use it rarely.
Insert Detail
An object that will matter. A grateful shot with AI: there’s no face in it to fall apart.
Alternate your shot sizes: a close-up after a wide, an insert after a close-up. Two similar shot sizes in a row look like a jump in the edit.
Angle: where are we looking from?
The height and tilt of the camera tell us how to see the character. The dashed line is the camera’s field of view.
Eye level
Neutral, as if we were standing next to her.
Low angle
Seen from below, she looks strong, big, even threatening.
High angle
Seen from above, she looks small, exposed, alone.
Bird’s-eye view
Straight down from above: a map, a pattern, order.
Dutch angle
A tilted horizon: something isn’t right.
Over the shoulder
We look at one person from behind the other one’s shoulder. For dialogue.
Point of view
We see through the character’s eyes. The viewer is inside the scene.
Two together
A cut is the relationship between two images: a close-up after an extreme wide says more than either one alone.
Movement: what is the camera doing?
Seen from above, the way the crew would see it. The camera moves, the character stays. Ask for one movement per shot.
Static camera
The safest choice with AI: the model pays attention to the character, not the camera.
Push-in
We slowly move closer: attention, tension.
Pull-back
We back away and reveal what’s around her. Loneliness, a reveal, a closing shot.
Pan
The camera turns sideways on the spot and shows us something.
Tilt
Up or down on the spot: from her shoes to her face.
Tracking
We move alongside her as she walks.
Orbit
We circle around her: it singles her out and stops time.
Crane
We rise and move away. A classic closing shot.
Handheld
A slight shake: documentary, urgency, nerves.
One shot, one movement. Ask for a pan, a push-in and an orbit at once and the model mixes them up. A zoom and a push-in aren’t the same thing: a zoom magnifies from where the camera stands, while in a push-in the camera moves closer and the space moves with it.
Lens and focus
The focal length (mm) decides how much fits in the frame and how close the background seems. The aperture (f) decides how blurry it gets. Drag the sliders.
Light: where does it come from, and what is it like?
Here the orange circle is the light source. Light sets the mood, often more than the story itself.
Golden hour
A low, warm sun and long shadows. Nostalgia, warmth.
Backlight
The light is behind the character: a silhouette with a glowing outline. Mystery, farewell.
Low key
A dark image with a single hard light. Tension, noir.
High key
Bright and soft, few shadows. Light-hearted, like an ad or a comedy.
From above
Hard light from overhead, shadows under the eyes. Cold, like an interrogation.
Practical light
The light source is in the frame: a lamp, a candle, neon, a phone.
Soft light (an overcast sky, a big window) feels friendly. Hard light (midday sun, a bare bulb) casts sharp shadows and feels dramatic. Write down which one you want: soft light or hard light.
Aspect ratio and resolution
Your film plays in a cinema, on a big screen. That means two things.
Landscape, 16:9
The screen is landscape. Generate every image and video in this format.
Portrait, 9:16: don’t
In a cinema it becomes a narrow strip squeezed between two big black bars. Keep the phone format for your posts.
Resolution
You can test at a lower resolution if the model allows it: it’s faster and cheaper. Whatever goes on the screen should be 1080p (1920 × 1080), 24 fps. You can hand in 720p, but it won’t be sharp enough on the big screen.
Add the widescreen look (2.39:1) in the edit, with two black bars. If you ask for it in the prompt, the model sometimes paints the bars into the image, and then you can’t remove them.
05Prompt
A prompt is built from bricks.
A good video prompt can be short, but it should be complete: it tells the model what the camera sees, what moves and what the light is like. Build it brick by brick and copy it at the end. Write it in English, that’s what the models understand best.
Starting image: the image prompt
It describes what’s in the image: shot size, character, place, light, style. This is the faster and cheaper step, so experiment here until it’s right.
Close-up of an old woman with a grey bun and a striped scarf at an empty tram stop at night, lit by a flickering neon sign, shot on 35mm film, muted colors.
Motion: the video prompt
If you start from an image, don’t describe the image again. Write only what changes: the action, the camera and the sound.
She slowly looks up at the sky. Slow push-in. Sound: wind and a distant tram bell.
- One shot, one action, one camera movement.
- Write verbs, not feelings: “puts down the mug with a trembling hand” says more than “she’s sad”.
- Describe the character with the same words in every prompt, or give the model an image of them as a reference.
- Describe what you want to see. A ban (“no people”) often brings in exactly what you wanted to keep out.
- Drop the empty words (beautiful, stunning, 8K, masterpiece). A specific lens, light and material are worth more.
- If it’s not working, change one thing in the prompt at a time, or you won’t know what helped.
06Seedance in ElevenLabs
How to write for Seedance.
ElevenLabs gives you several video models to choose from (Kling, Veo and others). We recommend Seedance. Two versions matter, 2.0 and 2.5: you write for them in a similar way, but they can do different things.
| In ElevenLabs | Seedance 2.0 | Seedance 2.5 |
|---|---|---|
| Length | 4–15 seconds | 4–30 seconds |
| Resolution | 480p, 720p, 1080p | 480p and 720p according to the documentation, though a test in early October also showed 1080p. Check in the interface. 720p can be screened, but it won’t be sharp enough on the big screen, so it’s worth upscaling. |
| Several cuts in one | with numbered shots (Shot 1:, Shot 2:), no timestamps | with seconds (0–4s:, 4–8s:), no gaps |
| For testing | 2.0 Mini or Fast: quicker, and Mini costs about half the credits | try 480p first |
| Best for | shots planned one by one, final 1080p | a longer scene with several cuts in one go |
The order
Subject, action, location, style, camera, sound. Leave out what you don’t need. This is ByteDance’s own recommendation for 2.5.
A tired tram driver slowly takes off his cap inside an empty tram at night. Muted 35mm film look. Slow push-in. Sound: the low hum of the engine.Several cuts in one generation
For 2.0, write numbered shots; it’s unreliable with timestamps. 2.5 understands seconds, as long as the segments follow each other with no gaps.
Shot 1: wide shot, the tram stops at an empty stop.
Shot 2: close-up, the driver looks at the empty seats.0–4s: wide shot, the tram stops at an empty stop.
4–8s: close-up, the driver looks at the empty seats.References
You refer to uploaded images by number (@Image1, @Image2), in the order you uploaded them. The first time you mention one, tie it to the character and say what to take from it and what not to. Check exactly what the interface calls them when you practice at home.
The woman from @Image1 (keep her face and striped scarf, not the background) walks into the kitchen from @Image2.The character image
Give it a neutral portrait and a full-body image of the character against a plain background. 2.0 easily mistakes a multi-view character sheet (front, side and back in one image) for several different people. Don’t upload a photo of a real person: the model may refuse it, and we wouldn’t put it on the screen either.
Starting frame or reference
2.0 Fast, 2.0 Mini and 2.5 accept either a start and end frame or reference images, not both together. If you need an exact composition, use a starting frame. If you need to carry the character through many scenes, use references.
Sound in the prompt
ByteDance’s notation: music ( ), sound effects < >, dialogue { }. If you don’t want music, say so at the end; if it still plays, say it at the start too, in more words. 2.5 speaks eleven languages, English among them, so English dialogue can go straight into the prompt. Hungarian isn’t on the list: for Hungarian lines, use ElevenLabs voices.
She whispers {"one more time"}. <a distant tram bell>
No music, no score, no soundtrack.The interface shows the price before you generate, and four variations cost four times as much as one. Test with 2.0 Mini or at 480p, and only animate starting images you’ve already approved. A 15-second 1080p clip with sound costs roughly a quarter of the credits with Kling that it does with Seedance 2.0 (a test from early October). If your budget is tight, the key scenes can go through Seedance and the rest through Kling. Do the math with the credit planner.
If you generate at 720p, it’s worth upscaling for the big screen: ElevenLabs has Topaz Video Upscale for this. Export at 24 fps. If you don’t want to bother with that, generate your final shots with 2.0, at 1080p.
When a shot is hard: seven blocks
The Grok Imagine team breaks longer video prompts into seven blocks, and ByteDance’s 2.5 guide gives a similar structure. Their prompts are long and detailed, just well organized. Their rule is for when something isn’t working: then you shorten, because the problem is rarely a missing detail.
CAMERA: static, eye level, 50mm lens, shot on 35mm film. ACTION: the old woman from @Image1 slowly lifts the cinema ticket to the light and smiles. ENVIRONMENT: empty tram stop at night, wet asphalt, light fog, a flickering neon sign. PHYSICS: her scarf moves gently in the wind; the paper ticket bends slightly. AUDIO: no music. Wind and distant traffic. SFX: paper rustle as she lifts the ticket. CONSTRAINTS: one continuous shot, no cuts, no text on screen, only one person in frame.
07Sound
Half the film is sound.
In a cinema, a weaker image gets by. Bad sound doesn’t. In ElevenLabs, voice, music and sound effects sit in the same place as the image, so you can build the sound there too.
Narration
ElevenLabs voices (Eleven v3 and v4) speak English, and Hungarian too if you want a line in Hungarian. Pick a ready-made voice, or describe a new one (Voice Design). Don’t ask for effects like echo or a phone filter in the description. Add those in the edit.
Native English. Female, in her seventies. Warm, slightly raspy. Persona: a grandmother telling a bedtime story. Emotion: tender, amused. Speaks slowly, with small pauses.ElevenLabs, Voice DesignDialogue
Seedance 2.5 speaks English, so English dialogue can come straight from the video model. It doesn’t speak Hungarian: for a Hungarian line, make the voice in ElevenLabs and add it with lip sync (under Image & Video: Sync, OmniHuman). In any language, it’s easier if the speaker doesn’t face the camera: seen from behind, on the phone, through a door, or as narration.
Micro sounds
Small sounds bring the image to life: a rustling sleeve, a breath, crinkling paper, a spoon in a cup, a click. A silent close-up feels dead. Generate them separately as short single sounds (one-shot), close up, without reverb. Ask for a few variations, and place each one on the exact frame where the movement happens. Even one footstep has layers: shoe, ground, clothing.
Close foley: the slow rustle of a wool coat sleeve as an arm lifts, single one-shot, dry, no reverb, about 1 second.A paper cinema ticket bent slowly between two fingers, very close, dry, one-shot.ElevenLabs, Sound effectsNoise and room tone
For bigger sounds (a door slam, a tram, rain), describe the source, the material, the distance and the rhythm. Ask for the sound of the place (the city, wind, the hum of a room) as a loop, and lay it under the whole scene. One sound can be 30 seconds at most. Generate a complex sequence in pieces and put it together in the edit.
Footsteps on wet cobblestones, close, slow, night street ambience.Distant night city ambience with a faint tram bell, seamless loop.Music
Answer five questions, or the result will be generic: genre, mood, instruments, tempo (BPM) and era. Describe the arc in time order. If you don’t want vocals, write instrumental only.
Melancholic minimalist piano with soft strings, 72 BPM, intimate 1970s European film score, close-mic'd piano. Starts with solo piano, strings enter at 15 seconds, ends on a single held note. Instrumental only.ElevenLabs, Music best practicesMixing in Studio
On the ElevenLabs Studio timeline, narration, music, noise and micro sounds each run on their own track, so you can line them up. The voice should be loudest, with the music underneath. Micro sounds work when nobody notices them, but everyone would miss them. Studio Agent can also put together a first cut, which you then adjust.
ElevenLabs, Studio- Voice
Narration or dialogue. What the image doesn’t say.
- Micro sound
The tiny sound of a movement: clothing, breath, paper, a cup.
- Noise
The bigger events: a door, a step, a tram.
- Room tone
What surrounds you: rain, the city, the hum of a room. Without it the image feels empty.
- Music
The feeling and the rhythm. You cut to it.
Cut to the music: let each cut land on a beat. A moment of silence before the end is worth more than any effect.
08When you’re stuck
Something falls apart for everyone.
The most common problems, and what to try first. If that doesn’t work, ask the mentors: they’ve seen it before.
- The character looks different in every shot.
Describe them with the same words in every prompt, and give a reference image. Start from an image, not from text.
- The face falls apart when it moves.
Ask for a slower, smaller movement and a shorter shot. Film the face moments closer. Save big movement for wide shots, where the face is small.
- Too much is moving at once.
One movement, one camera move. Write
static camera, and take out the second character. - It floats, looks rubbery, the physics is off.
Simpler action, heavy objects, a shorter clip. Things that fall, break or pour often work better in a detail shot.
- Captions or text in the image come out unreadable.
Don’t generate text. Add the title and captions in the edit.
- The hands look strange.
Move the hands further away or hide them. If they matter, give them a separate detail shot.
- A plastic, too-perfect “AI look”.
Name a specific lens and light source, add film grain and a little imperfection (
handheld,overcast,slightly overexposed). Drop the empty superlatives. - The same person appears twice.
For every character, say which image they’re based on, and end with
only one person in frame. Don’t give 2.0 a multi-view character sheet. - The camera doesn’t do what you asked.
Ask for one movement only, in one clear word, and don’t mix it with anything else. Models confuse a push-in with a zoom, so add
no lens zoom. - Captions or music show up when you didn’t ask for them.
Keep text off your reference images, and generate in landscape. Rule out music with several words at the start and the end of the prompt (
no music, no score, no soundtrack). - Adding more and more doesn’t help.
Detail helps when it’s about the image: who stands where, what they do, where the light comes from. If a prompt doesn’t work, first remove whatever isn’t visible in the image (backstory, explanation, piles of adjectives, requests that contradict each other), and only then add. The Grok Imagine team found the same: cutting the prompt down helps more often than adding detail.
- Time is running out.
At 21:30, start editing with what you have. A detail shot or a longer closing shot can stand in for a missing shot.
09Director’s sheet
The plan on one page.
Fill it in on Thursday evening, once you know the theme. Your browser remembers it, and you can print it. If you work with the director’s assistant, it walks you through the same questions and gives you a prompt sheet at the end.
| # | sec | Shot size | Camera | What happens | Sound |
|---|
10Before the evening and submission
What’s worth doing at home.
You only learn the theme on site, but you can try the tools before. If you’ve made one image and one video clip at home, you won’t be hunting for buttons on Thursday.
Before tick them off
At submission Thursday, by 22:30
- File
- MP4 (H.264 video, AAC audio)
- Image
- landscape 16:9, recommended 1920×1080, at least 1280×720, at most 2560×1440
- Frame rate
- 24 fps
- Sound
- rendered into the video file
- Length
- 30–90 seconds
- File name
- teamname_filmtitle.mp4
- Send to
- by email, to the organizers
- A 720p film can be screened, but it won’t be sharp enough on the cinema screen, so upscale it to 1080p before you edit (Topaz Video Upscale in ElevenLabs does this). We don’t screen anything below 720p. Set the export to 24 fps.
- We don’t touch the file you hand in: we screen exactly what you sent. Films without sound don’t get screened.
- Email your exported film to the organizers. Put your team name and the film title in the email. If the file is over 25 MB, send a Google Drive or WeTransfer link, and make sure the link opens for anyone. After 22:30 there’s no submission and no extension, not even for technical problems.
- If you used another generation tool, a bigger plan or extra credits, say so in the email. Your film will be screened in a separate block, out of competition.
- Export to MP4 at the best quality and don’t re-compress it. Don’t upload it to social media and download it from there: that crushes the image.
- The cinema projector is dimmer than your laptop. What looks dark on your screen turns muddy on the big screen, so don’t push the shadows too deep.
- After 21:30, don’t start a generation your film can’t be finished without.
11If you want to go deeper
Who we learned from.
This guide builds on these public materials. If you have time, take a look: most of them are short and full of examples.
- Inside Episode 1: The OathGrok Imagine (@imagine), Dev Walker and Wonder Studios · X · Oct 27, 2026
How a seven-minute AI episode was made, step by step, with the real working files: script, shot list, character sheet, key frames, a seven-block video prompt.
- Seedance 2.5 prompt guideByteDance · in Chinese · Aug 7, 2026, updated Oct 13
The official guide: order, time segments, references, sound tags, common mistakes.
- Seedance 2.0 prompt guideByteDance, BytePlus · in English
The rules for version 2.0, from character references to multi-shot generation.
- How to Prompt Seedance 2.5Emily Peng, Kapwing · Aug 21, 2026
A director’s brief for the machine, and a table of what goes wrong and what to rewrite.
- The Complete AI Camera Prompt GuideEmily Peng, Kapwing · Sep 11, 2026
77 shot sizes, angles and movements, each with a video and a prompt to copy.
- Video camera terms: examples & cheatsheetBlack Forest Labs
119 example videos, from shot size to light, each with its prompt.
- Ultimate Guide to Camera ShotsStudioBinder
The classic film basics with pictures: shot sizes, angles, movements.
- How to Make a Dynamic 60-Second Short FilmJason Hellerman, No Film School · Sep 22, 2024
What fits into a minute, and when the stakes, the obstacle and the twist should come.
- Image & VideoElevenLabs · documentation
The list of models, start and end frames, references, upscaling and lip sync.
- The ultimate prompting guide for Veo 3.1Khulan Davaajav, Hussain Chinoy, Google Cloud · Oct 16, 2025
If you want to try Veo too: dialogue, sound tags, timed shots.
Found something good? Send it to kamill@ai-budapest.com and we’ll add it to the next version.