Make a scene for Dub'Up

The Dub'Up catalogue is opening up to creators. You make a scene, you submit it, we watch it; if it is selected, it joins the game with your name on it. This page says what a Dub'Up scene is, what gets one selected, how to make it step by step, and how to submit it.

1. How it works
2. Before you start
3. What a Dub'Up scene is
4. The five rules that decide
5. The method, step by step
6. What you submit
7. Submitting
8. With Claude Code: our skills
9. What happens next

How it works

Dub'Up is a dubbing game: you pick a scene, everyone records their voice over one character, and the app puts the video together. The game is worth what its scenes are worth — and until now, they all came out of our own workshop. They are generated: characters and sets with an image generator, the video with a video generator that produces picture and voice together. That method is what we are handing over here.

What we are offering:

• You make a scene with the tools and the method described below.
• You submit it from the app, with an account.
• We review it: every submitted scene is watched by a person. Nothing goes live on its own.
• If it is selected, we finish it (music, mix settings) and it joins the catalogue: everyone can play it, and your creator name is shown on the scene and in the showcase.
• The best creators — those whose scenes are selected and played — will get powerful advantages in the game, on top of the visibility the showcase gives.

This page exists in every language of the app — English, French, German, Spanish, Italian and Portuguese (links at the top). It is written first in French: dubup.app/createurs. The skills further down are in French — they are our workshop's own working notes.

Before you start

• You need an account. You play Dub'Up without one, but you do not submit without one: a scene belongs to its author, and that is what lets us credit you and reach you. Create it in the app, Account tab (an email address and a password; 15 and over).
• The scene is yours. You submit it because you made it. No character, brand, universe or music belonging to someone else — a blue genie with gold cuffs, a moustached plumber in red overalls, a clip from a film: declined, even “inspired by”. The generators themselves block that kind of figure; so do we.
• By submitting, you allow us to show the scene in the game and in the showcase, to mix it with players' voices, and to show the videos that come out of it, with your name on them. You remain the author; you can ask us to take it down.
• The game is 12+. Humour, tension, drama, absurdity: yes. Sex, hate, gratuitous violence, politics: no. Players will say whatever they want over it; the scene itself has to be watchable by anyone.
• We select. A scene can be declined over a single one of the rules below. We say why, and you can submit again.

What a Dub'Up scene is

A short generated film that players will dub, one character each. What we expect, in numbers:

Our workshop uses GPT Image 2 for the stills and Seedance 2.5 (in Higgsfield's web interface) for the video. You may use other tools, as long as the result respects everything above — in particular lips that really articulate the lines, which only generators that produce picture and voice together know how to do.

The five rules that decide

They come from failed scenes. Any one of them, on its own, is enough to get a video declined.

The method, step by step

1. The idea, then the breakdown — before any prompt

A scene fits in one sentence: “a birthday toast that goes off the rails”, “a police check where it is the officer who is scared”. Then lay out a table, in seconds, shot by shot:

Check on this table, before spending a single credit: by 0:02 something has happened; silences long enough to carry the sound; a silent tail at the end; no noisy gesture under a line; every line between 2 and 4.5 s, and how many words it can hold — 5 words in 2 s, 7 in 3 s, 9 in 4 s, 12 in 5 s.

2. Sheets: one per character, one per place, generated once

Never count on a generator to hold a face from one shot to the next. Produce frozen sheets, which the video then takes as references:

• Character sheet (landscape, 2048 × 1152): three views on a single board — full front, full back, close-up of the face — neutral studio light, neutral expression, arms at the sides. It is a model sheet, not a film frame: emotion and lighting belong to the scene, not to the sheet.
• Location sheet (portrait, 1536 × 2592): three stacked views of the same set — an overall view, then the angles your shots will use — with nobody in it. It is the geometric anchor: without the reverse angle as a reference, the generator invents what is behind the camera, and the set shifts between two shots.
• A sheet is generated once, read, approved, and never regenerated.

3. The multi-shot prompt

The whole scene in one generation — never shot by shot, you would lose the continuity. The structure that works:

[Visual profile: light, contrast, focal length, grain — the same words for every shot.]

CHARACTERS: @image_1 is Tom, <visual markers>. @image_2 is Anna, <visual markers>.
  No exaggerated, theatrical or cartoonish expressions, no wide-eyed stares,
  no raised or knitted eyebrows. All expressions change gradually and
  continuously, never between one frame and the next.
SET: @image_3 is the dining room, wide view. @image_4 is the same room, reverse angle.
  The whole scene plays in this room; its geometry never changes.
ACTING: naturalistic, understated performance, as in a French realist family
  drama; the actors underplay; gestures small and economical; the voice does
  most of the work at ordinary conversational volume.
SOUND: real domestic sound, recorded in the room, at its natural loudness: no
  cinematic sound design, no added impact, no reverb. Every line is spoken at a
  slow, natural, unhurried pace, about two words per second, with real breaths;
  lines are short on purpose and must never be sped up to fit the shot.

Shot 1 (0-4 s) — wide, static. From the first frame Tom is already tapping his
  knife on a glass; three small tings at 0.2 s. From about 0.8 to 3.5 s Tom says
  "I signed my sister up for 'Roommate From Hell'." No noisy gesture during the line.
Shot 2 (4-8 s) — close on Anna. She sets her fork down at 4.3 s. From about 5 to
  7.5 s she says "You did what." No noisy gesture during the line.
…
Shot N (27-30 s) — nobody speaks, nobody makes a sound. Held reaction to black.

Spoken in English, slowly and naturally, never sped up.

• An emotion is written as a sequence of gestures, never as a state. his face reddens gives you a red that drops in one frame; his jaw sets; his breathing goes deep and slow gives you a man taking it. One change at a time, each with its duration. No skin colour, no direction on eyebrows or eyes, no bare emotion adjective.
• No stage verbs. shouts → speaks, loud, his voice tight; smashes → sweeps off the table with the back of his hand. A hesitation before every strong gesture: that is what makes it human.
• The speech window of every line, in seconds, inside the shot — not just the shot's length.
• The shot size and the camera move, for every shot. Without them the generator reframes at random.
• References are written @image_1, @image_2… (lower case, underscore); their number is the order in which you uploaded them.

4. Generating

1. Upload the sheets in the order of the prompt: characters, then places, then any key frames.
2. Set: the length from your table (30 s at most per generation), ratio 16:9, resolution 720p. The audio is generated with the picture — that is what we want.
3. Run a short test first — 4 to 6 seconds, at 480p — on the one shot you are unsure about (a line, a gesture). A 30 s take is expensive, and a short test tells you straight away whether the register is right.
4. Generate, download the mp4. Keep it as it is: that is the video with the voices, the one you will submit.

There is no seed: regenerating gives you another take, not a variant. Hence the importance of settling the breakdown and the prompt first.

Beyond 30 seconds — up to 45 — the scene is made in two generations, joined end to end. With no seed, the second take always drifts a little (faces, light); the cut only holds if it is justified by the story: a lamp going out, a door opening onto another room, a hard change of framing. Generate a key frame of that exact instant, approve the first take before launching the second, and give the second that key frame as its starting point, with the same sheets — removing any reference that contradicts the starting state (the lamp still lit). Submit both files as they are, in order: we do the joining.

5. Checking what came back

6. The scene sheet

A plain text that comes with the video. It lets us find who speaks when, and write the French version players will read:

Title: The Toast
Synopsis (280 characters at most): At the birthday dinner, Tom raises his glass
  and tells his sister he has signed her up for a reality show.
Characters:
  - Tom, man — the brother, thirties, creased shirt
  - Anna, woman — the sister, calm, never lets go of her fork
Lines (in order, with the approximate window):
  1. Tom, 0.8 → 3.5 s — "I signed my sister up for 'Roommate From Hell'."
     French adaptation: « J'ai inscrit ma sœur à Colocataire de l'enfer. »
     Intent: light, like good news
  2. Anna, 5 → 7.5 s — "You did what."
     French adaptation: « Tu as fait quoi. »
     Intent: without moving, the voice alone
  …
Length: 29 s. Last line finished at 26.5 s.

If you do not speak French, leave the adaptation out — write the English lines and their intent, and we will do the French.

What you submit

No music. Each scene's music is composed on our side and mixed in the edit: do not put music in the video — it would get in the way of separating the voices. We also take care of stripping the voices, measuring the lines, encoding to the catalogue profile and setting the mix. If you already master all that (a voiceless clip, a scene manifest), send it as well, never instead of the video with the voices.

Submitting

1. In the app — on a phone or in the browser, dubup.app/jouer, where your files are — open the Account tab: create an account if you do not have one, and verify your email address.
2. In the Creator card, tap Submit a scene: the title, the sheet, the video (the second generation is added next to it), the images if you have them.
3. Tick “this scene is my own creation and I allow Dub'Up to show it”, then Submit. Files go one at a time; a 45 s video takes a few dozen seconds.

The submission then appears under My submissions, with its state — pending, selected, declined — and the review note. You can withdraw it while it is pending. Any question: [email protected].

With Claude Code: our skills

Everything above is the summary of our workshop's method. The full method — with the traps met scene after scene — is written as skills for Claude Code: instruction files the assistant reads so it can write the image prompts and the video prompt for you, go over your breakdown, and draft the sheet. If you work with Claude Code, take them. They are written in French; Claude reads them in French and will answer you in your own language.

Installing: in your project folder, create .claude/skills/dubup-image/SKILL.md, .claude/skills/dubup-video/SKILL.md and .claude/skills/dubup-battle/SKILL.md from these files (renamed SKILL.md), and .claude/skills/dubup-video/references/manifest-checklist.md for the checklist. Then, in Claude Code: “Prepare a Dub'Up scene: a birthday toast that goes off the rails, three characters” — the assistant will walk you through the steps above, in order.

These skills are written for our workshop: some steps (stripping the voices, measuring the lines, encoding, the catalogue) talk about our tools and our folders. They are not your problem — you submit the video with the voices, and we do the rest. What is yours is everything before the generation, and the checking after.

What happens next

• The review. We watch every submitted scene against the rules on this page — the freeze frame at 1 s, the lips, the silence under the lines, the acting register. Allow a few days. A decline comes with a reason; you can fix it and submit again.
• The finishing. A selected scene goes through our workshop: voices stripped, lines measured, music composed and mixed, avatars, French text. We may suggest trimming the end or adapting the lines; the scene stays yours.
• Going live. The scene joins the catalogue, in its universe, with your creator name; the videos players make from it and publish in the showcase carry it too.
• The creators who count. The most played and best rated scenes lift their author: that is where the advantages reserved for the best creators are decided, and their place in the showcase.

‹ Back to the home page · Cette page en français · Privacy policy (in French)