Learn
Formats & Exhibition
AI Film Scene Remake

Formats & Exhibition · AI filmmaking

AI Film Scene Remake

External-tool reference examples from the source experiments, not a claim of current FlashBoards model availability or installed automation. Use your own supported tools and authorized media; settings, timings and comparisons describe the tested runs.

Also called: AI scene remake, recreate a movie scene with AI, AI recast, film scene with AI characters, AI shot-for-shot remake

Remaking a scene is copying everything except the faces and the voices. The reference method starts with a scene you have permission to use and your own characters: cast them in the film’s roles, take the edit, the angles and the lines from the corpus, make one picture per angle and keep the actors’ timing in your characters’ voices. The example run below is the coffee-shop scene of Heat (1995) with our demo characters Dex and Mara — every prompt, every measurement and the five tries before it.

Watch first: the whole guide in 1:47, hosted by Yaro

Video script

Hi, I'm Yaro. Today we remake a real movie scene with our own AI characters. The rule: copy everything except the faces and the voices.

Step one: cast first. Dex plays Neil, Mara plays Hanna, in the coffee-shop scene of Heat. Unclear who plays whom? Ask. I guessed once and got cast as the ketchup.

Don't invent the edit. The corpus has every cut, angle and line: twenty-seven point nine seconds, five shots, two angles. The director already did the homework.

Step two: one picture per camera angle, so a dialogue needs two. GPT Image 2.5 swaps only the face looking at the camera. Swap both, and the shot falls apart.

Step three: voices. Text-to-speech made Dex sound dead, like a sat-nav reading poetry. So we keep the actors' performance and change only the timbre: Audio Isolation, then Speech-to-Speech.

Watch the defaults: "or I'll die in my sleep" came back as "while tying my shoelace." Stability zero point seven fixed it.

Under it all goes the film's own room tone, looped from its quiet gaps. Not fake diner noise.

Step four: shoot. MiniMax H3 takes parts up to fifteen seconds, each ending on a cut. Mara speaks off-screen? Tell Dex to keep his mouth shut. He did. Unlike me.

The source tested H3 Max at fifteen seconds and Turbo at thirteen seconds. Turbo goes angle to angle, Max when an angle comes back.

Copy the scene, not the faces. Every prompt and every try, flops included, is right below. Scroll down. I'll bring the coffee.

What it does: Your AI characters take over the roles of a real film scene, cut where the film cuts and speaking where the actors spoke.
Use: authorized scene notes for the edit, GPT Image 2.5 for one picture per angle, ElevenLabs isolation + speech-to-speech, MiniMax H3 in parts ≤ 15 s.
Keep: the film’s edit and the actors’ timing; change only faces and timbre.

✓ Source experiment pick

Copy the scene, not the faces

Your characters take over the film’s roles and play the scene cut where the film cuts, every word where the actors said it, in your characters’ voices. Example run: our demo characters Dex and Mara in Heat — he tells her about the dream where he is drowning; she asks what it is about. 27.9 seconds, five shots, two angles. Under every remake on this page runs the original video, in sync.

Cast: your characters replace the film’s roles — {character A} for role 1, {character B} for role 2; ask when it is unclear
Scene: the film’s own edit — units, angles, cuts and lines from your authorized scene notes; the example run is the coffee-shop scene of Heat (1995)
Pictures: one per camera angle — two for a two-angle dialogue, every face replaced
Voices: the actors’ performance → ElevenLabs Audio Isolation → Speech-to-Speech in the voices from your characters’ cards, the film’s room tone
Video: MiniMax H3 · parts ≤ 15 s ending on a cut · 480P

See every prompt →

Reference guide

filmscene — a reference guide linking the steps

For footage you have permission to use, this external reference guide explains how to cast your own characters first, read the cuts, angles and lines, make one picture per angle, re-voice the actors in your characters’ voices, write every part shot by shot and check each generation against the original — with the exact settings and prompts from this page.

⬇Download reference guidefilmsceneA reference guide connecting the creative steps — corpus · pictures · sound · shots

1filmscene-corpusCast your characters, then units, setups, cuts and lines from the corpus
2filmscene-setupsOne picture per camera angle
3filmscene-soundThe actors’ timing in your characters’ voices
4filmscene-shotsParts ≤ 15 s on H3, one at a time

⬇ filmscene-corpus ⬇ filmscene-setups ⬇ filmscene-sound ⬇ filmscene-shots

The short answer

A scene remake works when you copy everything but the faces and the voices. Cast your own characters in the film’s roles, take the edit, the angles, the lines and their timing from the film itself; make one picture per camera angle with your characters in it; keep the actors’ performance and only change the timbre of their voices; then let MiniMax H3 shoot it in parts that end where the film cuts.

Cast✕ demo characters, or a guess at who plays whom✓ your characters, one per film role — ask when unclear
Edit✕ a montage we invented✓ the film’s units, cuts and angles from the corpus
Pictures✕ a new frame for every shot✓ one picture per angle — two for a dialogue
Voices✕ text-to-speech lines, dead and off the lips✓ the actors’ delivery in your characters’ timbre
Sound✕ two audio tracks, a generated “ambience”✓ one track as long as the video, the film’s own room tone
Process✕ the whole scene in one go✓ one part, compared with the original, approved, then the next

Draft or quality: Seedance 2.5

The external method on this page uses MiniMax H3 for drafts and many takes. Seedance 2.5 reference-to-video accepts the same pictures and sound as references (@Image1, @Audio1), up to 30 seconds a call, with its own synced sound at 480p, 720p or 1080p. The source preferred Seedance for quality. Side by side on our product commercial, its reported cuts were approximately one second early and half a second late; it drew finer detail and added its own sound effects in sync. This side-by-side result is from a product commercial, not a measurement on every format.

1 · The scene as data

any scene of the FlashBoards corpus · example: Heat (1995) · scene 95

Cast first. Before anything else, fix who plays whom: {character A} replaces <film role 1>, {character B} replaces <film role 2> — your own characters, from their character cards: the sheet, the face board, the voice. If it is unclear who plays which role, ask; never fall back to the demo characters. Every later step — the setup pictures, the speakers in the prompt, the voices — follows this cast.

Nothing in the remake is invented. The source experiment used a prepared scene corpus, not supplied here. For your own authorized scene, record its units with their HD times, the camera setup of every unit (the same letter is the same angle), the camera moves, the cut times and the subtitle lines with the film role that says them. Each role maps to the character you cast for it.

H3 takes up to 15 seconds, so the scene is cut into parts that end on a cut. A part that runs a little over loses silence at the head before the first line, never the tail; a part shorter than whole seconds is generated at the next whole second and trimmed. A subtitle row can hold two speakers — the diarized transcript of the original splits it.

Example run: Heat (1995), scene 95. Setup A is the angle on Neil over Hanna’s shoulder, C the reverse; our demo character Dex plays Neil and Mara plays Hanna. Part 1 ran 15.14 s — we trimmed 0.14 s of silence at the head. Part 2 is 12.888 s, generated as 13 s and trimmed. The two-speaker row: “Enough time to do what you wanna do? That’s right.”

Part	Unit	Setup	In the part	Length	Camera	Lines
1	996_0	A · on Dex, over her shoulder	0.000–8.212 s	8.212 s (0.14 s trimmed at the head)	Static slow	Dex: “I have one where I'm drowning.” Dex: “And I gotta wake myself up and start breathing, or I'll die in my sleep.”
1	997_0	C · on Mara, over his shoulder	8.212–11.747 s	3.535 s	Static slow	Mara: “You know what that's about?”
1	997_1	A · on Dex, over her shoulder	11.747–15.000 s	3.253 s	Static zero	Dex: “Yeah. Having enough time.”
2	998_0	C · on Mara, over his shoulder	0.000–4.849 s	4.849 s	Static slow	Mara: “Enough time to do what you wanna do?”
2	998_1	A · on Dex, over her shoulder	4.849–12.888 s	8.039 s	Static slow	Dex: “That's right.” Mara, off-screen: “You doing it now?” Dex: “No, not yet.”

The acting comes from the film too: a one-frame-per-second contact sheet of each part, and every shot’s action and acting written from it — where the eyes go, when the head drops, when he nods.

Reference guides for this step · Markdown instructions for Claude Code or any agent

⬇Download reference guidefilmscene-corpusThe scene as data — units, setups, cuts, lines, parts ≤ 15 s

2 · One picture per angle

GPT Image 2.5 Flare edit

A scene in one place shot from two angles needs two pictures: one per camera setup, with your cast in them — the film frame of each setup, your characters’ sheets as the other pictures. Everything else — which angle when, what they do — is the prompt. Each picture is a film frame with the letterbox cropped (1920 × 800) through GPT Image 2.5 Flare edit. The prompts are blind: they name the pictures’ roles and never describe the frame or anyone’s look.

Example run: setup A and setup C of the Heat scene with Dex and Mara.

Dex — Setup A · only the man facing the camera replaced

film — Setup A · only the man facing the camera replaced

Mara and Dex — Setup C · both people placed, every extra invented

Replace only the face that faces the camera when the other person is a blurred shoulder in front. Replacing both broke the over-the-shoulder geometry.
Every extra gets an invented face, and the prompt says there is no text, logo or watermark — the source frames carried a channel watermark.

The method · prompt for a setup with one face to the camera

Picture 1 is only the look of {character A}: their face, hair and their whole outfit. Picture 2 is a frame from a film. Shoot picture 2 again exactly the same: the same camera, framing, focus, set, light and composition. The one person who faces the camera is {character A} from picture 1, in their outfit from picture 1, in the same pose, head turn, gaze and expression — take them from picture 2. Everything out of focus in the foreground stays exactly as it is. Every other person keeps their place, pose and clothes but gets a completely different, invented face. There is no text, logo or watermark anywhere in the picture. Change nothing else.

The method · prompt for a setup with both people placed

Picture 1 is only the look of {character A}: take only their face and hair from it, never their clothes. Picture 2 is only the look of {character B}: take only their face and hair from it, never their clothes. Picture 3 is a frame from a film. Shoot picture 3 again exactly the same: the same camera, framing, focus, set, light, poses, clothes, head turns, gazes and expressions; but the two main people in it, from left to right, are {character A} (picture 1) and {character B} (picture 2), even where one of them is seen from behind or out of focus. Every other person in the frame keeps their place, pose and clothes but gets a completely different, invented face that belongs to nobody from any film or real life. There is no text, logo or watermark anywhere in the picture. Change nothing else.

Example run · the two prompts as sent, with Dex and Mara

Setup A (one person)

Picture 1 is only the look of Dex: his face, hair, moustache and his whole outfit. Picture 2 is a frame from a film. Shoot picture 2 again exactly the same: the same camera, framing, focus, set, light and composition. The one person who faces the camera is Dex from picture 1, in his outfit from picture 1, in the same pose, head turn, gaze and expression — take them from picture 2. Everything out of focus in the foreground stays exactly as it is. Every other person keeps their place, pose and clothes but gets a completely different, invented face. There is no text, logo or watermark anywhere in the picture. Change nothing else.

Setup C (two people)

Picture 1 is only the look of Mara: take only her face and hair from it, never her clothes. Picture 2 is only the look of Dex: take only his face, hair and moustache from it, never his clothes. Picture 3 is a frame from a film. Shoot picture 3 again exactly the same: the same camera, framing, focus, set, light, poses, clothes, head turns, gazes and expressions; but the two main people in it, from left to right, are Mara (picture 1) and Dex (picture 2), even where one of them is seen from behind or out of focus. Every other person in the frame keeps their place, pose and clothes but gets a completely different, invented face that belongs to nobody from any film or real life. There is no text, logo or watermark anywhere in the picture. Change nothing else.

Reference guides for this step · Markdown instructions for Claude Code or any agent

⬇Download reference guidefilmscene-setupsOne picture per angle — your characters in, every extra out

3 · The actors’ performance, your characters’ voices

ElevenLabs Audio Isolation · Speech-to-Speech

Text-to-speech voices read the lines well and sound nothing like a scene: no hesitation, no breath, and the timing is theirs, not the film’s. So the lines are not re-read but re-voiced:

The part’s sound is cut from the film and run through ElevenLabs Audio Isolation — only the voices remain.
It is split by speaker, and each piece goes through Speech-to-Speech into the voice of the character you cast for that role — the ElevenLabs voice on their character card (eleven_multilingual_sts_v2, stability 0.7, similarity 0.95). A character with no voice yet: pick or clone one first; never a demo voice. The actors’ isolated voices are a working file and are never published.
The pieces go back at their offsets on a silent track exactly as long as the generation — that is Audio 1.

Example run: Dex in a soft-spoken male voice (Spence), Mara in a low female one (Hope) — demo voices, never for your characters. The length and the word times stayed the actor’s (8.34 s in, 8.36 s out). The default settings garbled a mumbled ending — “or I’ll die in my sleep” came back as “while tying my shoelace”; stability 0.7 fixed it. Every track is checked with Scribe v2 against the original — each line starts where the actor started it:

Part	Who	Line	Original	Audio 1	Shift
1	Dex	“I have one where I'm drowning.”	2.14 s	2.16 s	+0.02 s
1	Dex	“And I gotta wake myself up…”	4.86 s	4.86 s	+0.00 s
1	Mara	“You know what that's about?”	10.58 s	10.60 s	+0.02 s
1	Dex	“Yeah. Having enough time.”	12.44 s	12.44 s	+0.00 s
2	Mara	“Enough time to do what you wanna do?”	1.60 s	1.60 s	+0.00 s
2	Dex	“That's right.”	5.18 s	5.18 s	+0.00 s
2	Mara	“You doing it now?”	8.90 s	8.94 s	+0.04 s
2	Dex	“No, not yet.”	10.62 s	10.62 s	+0.00 s

The room tone is the film’s own. There is no clean background track, so cut the scene’s gaps where nobody speaks, crossfade them into one loop for the whole scene and lay it under the voices at the film’s level. In the example run: 11.2 s before “There’s a flip side to that coin” and 4.6 s after “You doing it now?” — crossfaded into one loop and laid about 13 dB under the voices, as in the film. A generated “diner ambience” was louder and read as noise.

Reference guides for this step · Markdown instructions for Claude Code or any agent

⬇Download reference guidefilmscene-soundThe actors’ performance, your voices — isolation, speech-to-speech, room tone

4 · One part, one generation

MiniMax H3 · 480P · 21:9

Each part is one prompt with every shot written out: the cut time from the corpus, which setup picture it is, the camera (static, as in the film), the action as beginning → then → end, the acting, and the dialogue with the measured word times of Audio 1. A line said off-screen is marked on_screen: false, and the rules say the face on screen keeps its mouth closed. Every generation is checked — where the cuts landed, how far each line moved — and sent next to the original for approval before the next part. Each video below is that comparison: the remake on top, the original video underneath, the sound of the remake.

The prompt structure, filled from your cast — looks come from the setup pictures only, never written into look:

The method · prompt structure for one part

speakers: {character A}: S1, {character B}: S2
{ output: one continuous film scene, D seconds, 21:9, N shots with hard cuts exactly at the given times; Audio 1 is exactly D seconds…,
  references: { Image 1: setup A — every {character A} shot looks exactly like it…, Image 2: setup B — every {character B} shot…, Audio 1: the whole dialogue… },
  characters: { {character A}: {look: the person facing the camera in Image 1…, voice: {character A}'s voice of Audio 1 (name it by what tells the two voices apart, e.g. the man's / the woman's / the lower voice)…}, {character B}: {…} },
  environment: { location…, light…, sound: only Audio 1; no music },
  style: live-action 35 mm film, natural and restrained…,
  shots: [ { shot, cut_at, from, to, image, camera (setup + "static" from the corpus move), action: BEGINNING… THEN… END…, acting,
             dialogue: [ { speaker: S1, on_screen: true, when: 00:02.140–00:03.200, line: <d>[English] "…"</d> } ] } ],
  rules: [only the person facing the camera speaks…; an off-screen line is heard from the person seen from behind and the face on screen keeps its mouth closed; no cut to the speaker inside that shot; one full-screen shot at a time; no phones] }

Example run · every generation · video, inputs, prompt

Part 1 · H3 Max

Dex on the dream, Mara’s question, his answer — setups A → C → A, cuts at 8.212 and 11.747 s.

Model: MiniMax H3 Max reference-to-video
Settings: 15 s · 480P · 21:9 · enhance disabled
Inputs: Image 1 setup A, Image 2 setup C, Audio 1
Time: 8.7 s of generation

Cuts at 8.79 and 11.50 s; Dex’s lines exact (0.00 s), Mara’s line re-voiced by the model 0.95 s early, still inside her shot.

Audio 1 · the actors’ lines in Dex’s and Mara’s voices

Inputs and prompt

The inputs H3 saw · click for each one’s prompt

speakers: dex: S1, mara: S2
{
 output: one continuous film scene, 15 seconds, 21:9, three shots with hard cuts exactly at the given times; Audio 1 is exactly 15 seconds and plays from 00:00.000 to the end,
 references: {
  Image 1: setup A of this scene — the over-the-shoulder close-up on Dex; every Dex shot looks exactly like it: the same framing, faces, focus, place and light,
  Image 2: setup B of this scene — the over-the-shoulder close-up on Mara; every Mara shot looks exactly like it: the same framing, faces, focus, place and light,
  Audio 1: the whole dialogue, both voices in turn; every line is lip-synced by its speaker exactly at the given times
 },
 characters: {
  dex: {name: Dex, look: the man in the white suit facing the camera in Image 1, exactly that face, hair and moustache, voice: the man's voice of Audio 1, a tired, quiet, even voice},
  mara: {name: Mara, look: the young woman facing the camera in Image 2, exactly that face, slicked-back black hair and dark jacket, voice: the woman's voice of Audio 1, low and calm}
 },
 environment: {
  location: a crowded diner at night, two people face to face across a small table, people and lights soft and blurred behind them; exactly as in the images,
  light: cool dim overhead light with a blue-green cast, soft and low-contrast, deep shadows on the faces, warm small practical lights far behind,
  sound: only Audio 1; no music
 },
 style: live-action 35 mm film, natural and restrained, tiny real movements, no slow motion, no text,
 shots: [
  { shot: 1, from: 00:00.000, to: 00:08.212, image: Image 1,
    camera: an over-the-shoulder close-up on Dex at eye level, Mara's dark hair and shoulder out of focus in the left foreground with her back to the camera; static, a barely noticeable slow push-in,
    action: BEGINNING Dex looks off past Mara to the left, lips pressed together in a sour little grimace, as if a bad memory came back; THEN he lowers his eyes to the table and says the first line quietly to himself; THEN he lifts his eyes to Mara and keeps them on her, his head slightly tilted, and says the second line evenly, almost matter-of-fact; END he holds her gaze, still,
    acting: weary, honest, guarded; small movements of the head only, no gestures,
    dialogue: [
     { speaker: S1, on_screen: true, when: 00:02.140–00:03.200, line: <d>[English] "I have one where I'm drowning."</d> },
     { speaker: S1, on_screen: true, when: 00:04.860–00:07.750, line: <d>[English] "And I gotta wake myself up and start breathing, or I'll die in my sleep."</d> } ] },
  { shot: 2, cut_at: 00:08.212, from: 00:08.212, to: 00:11.747, image: Image 2,
    camera: the reverse angle — an over-the-shoulder close-up on Mara at eye level, Dex's dark hair and white jacket out of focus in the right foreground with his back to the camera; static,
    action: BEGINNING Mara sits leaning slightly forward, listening, her eyes fixed on Dex, head a little tilted; THEN she lifts her chin and her eyebrows slightly and asks the question softly, with real curiosity; END she waits for his answer, looking straight at him,
    acting: attentive, sharp, almost gentle; blinks rarely,
    dialogue: [
     { speaker: S2, on_screen: true, when: 00:10.600–00:11.360, line: <d>[English] "You know what that's about?"</d> } ] },
  { shot: 3, cut_at: 00:11.747, from: 00:11.747, to: 00:15.000, image: Image 1,
    camera: back to the over-the-shoulder close-up on Dex exactly as in shot 1; static, no camera movement,
    action: BEGINNING Dex looks at Mara and gives a small nod; THEN he says "Yeah" and pauses for a breath, his eyes dropping for a moment; THEN he looks back at her and says "Having enough time" quietly, with a faint resigned half-smile; END he stays still, looking at her,
    acting: calm, fatalistic, a touch of irony,
    dialogue: [
     { speaker: S1, on_screen: true, when: 00:12.440–00:12.660, line: <d>[English] "Yeah."</d> },
     { speaker: S1, on_screen: true, when: 00:13.420–00:14.220, line: <d>[English] "Having enough time."</d> } ] }
 ],
 rules: [only the person facing the camera speaks, and the lips match every word at the given times; the person seen from behind never speaks; nobody else talks; it is always one single full-screen shot; no split screen; no phones]
}

Part 2 · H3 Max

“Enough time to do what you wanna do?” — setups C → A, cut at 4.849 s, one line off-screen.

Model: MiniMax H3 Max reference-to-video
Settings: 13 s · 480P · 21:9 · enhance disabled
Inputs: Image 1 setup A, Image 2 setup C, Audio 1
Time: 8.7 s of generation

Cut at 4.75 s; all four lines at 0.00 s, Dex keeps his mouth shut while Mara asks off-screen.

Audio 1 · the actors’ lines in Dex’s and Mara’s voices

Inputs and prompt

The inputs H3 saw · click for each one’s prompt

speakers: dex: S1, mara: S2
{
 output: one continuous film scene, 13 seconds, 21:9, two shots with one hard cut exactly at 00:04.849; Audio 1 is exactly 13 seconds and plays from 00:00.000 to the end,
 references: {
  Image 1: setup A of this scene — the over-the-shoulder close-up on Dex; every Dex shot looks exactly like it: the same framing, faces, focus, place and light,
  Image 2: setup B of this scene — the over-the-shoulder close-up on Mara; every Mara shot looks exactly like it: the same framing, faces, focus, place and light,
  Audio 1: the whole dialogue, both voices in turn; every line is heard exactly at the given times
 },
 characters: {
  dex: {name: Dex, look: the man in the white suit facing the camera in Image 1, exactly that face, hair and moustache, voice: the man's voice of Audio 1, a tired, quiet, even voice},
  mara: {name: Mara, look: the young woman facing the camera in Image 2, exactly that face, slicked-back black hair and dark jacket, voice: the woman's voice of Audio 1, low and calm}
 },
 environment: {
  location: a crowded diner at night, two people face to face across a small table, people and lights soft and blurred behind them; exactly as in the images,
  light: cool dim overhead light with a blue-green cast, soft and low-contrast, deep shadows on the faces, warm small practical lights far behind,
  sound: only Audio 1; no music
 },
 style: live-action 35 mm film, natural and restrained, tiny real movements, no slow motion, no text,
 shots: [
  { shot: 1, from: 00:00.000, to: 00:04.849, image: Image 2,
    camera: an over-the-shoulder close-up on Mara at eye level, Dex's dark hair and white jacket out of focus in the right foreground with his back to the camera; static, a barely noticeable slow push-in,
    action: BEGINNING Mara looks at Dex, her head slightly tilted, eyebrows a little raised, as if she just understood him; THEN she says the line slowly, half asking and half confirming, with a small nod on "time"; END she keeps her eyes on him and waits, lips closed,
    acting: thoughtful, perceptive, almost sympathetic; only small movements of the head,
    dialogue: [
     { speaker: S2, on_screen: true, when: 00:01.600–00:04.340, line: <d>[English] "Enough time to do what you wanna do?"</d> } ] },
  { shot: 2, cut_at: 00:04.849, from: 00:04.849, to: 00:13.000, image: Image 1,
    camera: the reverse angle — an over-the-shoulder close-up on Dex at eye level, Mara's dark hair and shoulder out of focus in the left foreground with her back to the camera; static, no camera movement,
    action: BEGINNING Dex looks at Mara and answers with a small nod; THEN he lowers his eyes to the table and presses his lips together, thinking, his mouth pulled down a little; THEN, while Mara asks her question off-screen, he stays still with his eyes down and his lips closed — he does not speak, only the back of Mara's head moves slightly; THEN, still looking down, he answers quietly, almost to himself, with a tiny shake of the head; END he stays still, eyes lowered, then lifts them slightly toward her,
    acting: private, honest, a little sad; no gestures,
    dialogue: [
     { speaker: S1, on_screen: true, when: 00:05.180–00:05.640, line: <d>[English] "That's right."</d> },
     { speaker: S2, on_screen: false, when: 00:08.940–00:09.740, line: <d>[English] "You doing it now?"</d> },
     { speaker: S1, on_screen: true, when: 00:10.620–00:11.320, line: <d>[English] "No, not yet."</d> } ] }
 ],
 rules: [the lips of the person facing the camera move only on his or her own lines, at the given times; Mara's question in shot 2 is heard off-screen from the woman seen from behind, and Dex's mouth stays closed while she speaks; there is no cut to Mara in shot 2; nobody else talks; it is always one single full-screen shot; no split screen; no phones]
}

Part 2 · H3 Max Turbo

The same part on Turbo: setup C as the first frame, setup A as the last frame, our voices pinned.

Model: MiniMax H3 Max Turbo image-to-video
Settings: 13 s · 480P · enhance disabled
Inputs: first frame setup C, last frame setup A, target audio
Time: 3.7 s of generation

Cut at 4.67 s; our audio exactly as given (correlation 1.00); 3.7 s of generation against 8.7 s on Max.

Audio 1 · the actors’ lines in Dex’s and Mara’s voices

Inputs and prompt

The inputs H3 saw · click for each one’s prompt

speakers: dex: S1, mara: S2
{
 output: one continuous film scene, 13 seconds, 21:9, two shots with one hard cut exactly at 00:04.849; the given audio is exactly 13 seconds and plays from 00:00.000 to the end,
 references: {
  first frame: shot 1 starts exactly on the given first frame — the over-the-shoulder close-up on Mara,
  last frame: shot 2 ends exactly on the given last frame — the over-the-shoulder close-up on Dex; the whole of shot 2 has exactly this framing, faces, focus, place and light,
  audio: the whole dialogue, both voices in turn; every line is heard exactly at the given times
 },
 characters: {
  dex: {name: Dex, look: the man in the white suit facing the camera on the given last frame, exactly that face, hair and moustache, voice: the man's voice of the given audio, a tired, quiet, even voice},
  mara: {name: Mara, look: the young woman facing the camera on the given first frame, exactly that face, slicked-back black hair and dark jacket, voice: the woman's voice of the given audio, low and calm}
 },
 environment: {
  location: a crowded diner at night, two people face to face across a small table, people and lights soft and blurred behind them; exactly as on the given frames,
  light: cool dim overhead light with a blue-green cast, soft and low-contrast, deep shadows on the faces, warm small practical lights far behind,
  sound: only the given audio; no music
 },
 style: live-action 35 mm film, natural and restrained, tiny real movements, no slow motion, no text,
 shots: [
  { shot: 1, from: 00:00.000, to: 00:04.849, image: the given first frame,
    camera: an over-the-shoulder close-up on Mara at eye level, Dex's dark hair and white jacket out of focus in the right foreground with his back to the camera; static, a barely noticeable slow push-in,
    action: BEGINNING Mara looks at Dex, her head slightly tilted, eyebrows a little raised, as if she just understood him; THEN she says the line slowly, half asking and half confirming, with a small nod on "time"; END she keeps her eyes on him and waits, lips closed,
    acting: thoughtful, perceptive, almost sympathetic; only small movements of the head,
    dialogue: [
     { speaker: S2, on_screen: true, when: 00:01.600–00:04.340, line: <d>[English] "Enough time to do what you wanna do?"</d> } ] },
  { shot: 2, cut_at: 00:04.849, from: 00:04.849, to: 00:13.000, image: the given last frame,
    camera: the reverse angle — an over-the-shoulder close-up on Dex at eye level, Mara's dark hair and shoulder out of focus in the left foreground with her back to the camera; static, no camera movement,
    action: BEGINNING Dex looks at Mara and answers with a small nod; THEN he lowers his eyes to the table and presses his lips together, thinking, his mouth pulled down a little; THEN, while Mara asks her question off-screen, he stays still with his eyes down and his lips closed — he does not speak, only the back of Mara's head moves slightly; THEN, still looking down, he answers quietly, almost to himself, with a tiny shake of the head; END he stays still, eyes lowered, then lifts them slightly toward her,
    acting: private, honest, a little sad; no gestures,
    dialogue: [
     { speaker: S1, on_screen: true, when: 00:05.180–00:05.640, line: <d>[English] "That's right."</d> },
     { speaker: S2, on_screen: false, when: 00:08.940–00:09.740, line: <d>[English] "You doing it now?"</d> },
     { speaker: S1, on_screen: true, when: 00:10.620–00:11.320, line: <d>[English] "No, not yet."</d> } ] }
 ],
 rules: [the lips of the person facing the camera move only on his or her own lines, at the given times; Mara's question in shot 2 is heard off-screen from the woman seen from behind, and Dex's mouth stays closed while she speaks; there is no cut to Mara in shot 2; nobody else talks; it is always one single full-screen shot; no split screen; no phones]
}

Reference guides for this step · Markdown instructions for Claude Code or any agent

⬇Download reference guidefilmscene-shotsOne part, one generation — the prompt, Max or Turbo, the checks

5 · H3 Max or H3 Max Turbo

the same part, measured

	H3 Max reference-to-video	H3 Max Turbo image-to-video
Pictures	up to 9 references — both angles anywhere in the part	a first and a last frame — the part must start on one angle and end on another
Our audio	a reference: the first speaker kept exactly, the second may be re-voiced and moved (−0.95 s in part 1)	pinned as is (correlation 1.00)
Part 2 · cut at 4.849 s	4.75 s	4.67 s
Part 2 · time	8.7 s	3.7 s
Use for	a part where an angle comes back (A → C → A)	a part from one angle to another (C → A), or one angle only

Turbo needs the faces of both angles at the ends of the part. In the example run part 1 goes A → C → A, so Mara’s face (setup C) is in neither end frame — on Turbo it becomes two generations split at a cut.

How we got here

example run · same scene, six tries

✕How not totext-to-speech and our own montage

ElevenLabs voices on one track, seven parts, shots we picked ourselves from the scene.
Mara stayed in sync, Dex did not, and his voice sounded dead.

✕How not toone line per generation

Two separate voice tracks confused the model; one track with the timings in the prompt still had the second speaker moved by about a second — so every line got its own generation, cut by its words.
H3 added speech in the silence and stuttered on “…”; the edit was still ours, not the film’s.

✕How not tothe film’s window, text-to-speech

The first unit exactly as in the corpus, first and last frames on Turbo, lines placed at the subtitle times.
In sync — but the delivery is not the actor’s.

✓How tothe film’s window, the actor re-voiced

The same window with the actor’s own delivery in Dex’s voice: the pauses and the weight are back.

✓How totwo pictures, the rest in the prompt · H3 Max

Part 1: three shots, two angles, cuts from the corpus, the film’s room tone.

✓How tothe same on Turbo

Part 2 from angle to angle on Turbo, our audio pinned exactly.

Settings and measurements

fal · ElevenLabs

Step	Tool	Time
A setup picture	GPT Image 2.5 Flare edit, low, 1920 × 800	20–30 s
Voices only	ElevenLabs Audio Isolation	seconds
The actors in your voices	ElevenLabs Speech-to-Speech, multilingual v2	seconds
Check the words	ElevenLabs Scribe v2 on fal	≈ 5 s
A part on H3 Max	reference-to-video, 480P, 2 images + audio	8.7 s
A part on H3 Max Turbo	image-to-video, 480P, first + last frame + audio	3.7 s
27.9 s of the scene	2 parts on H3 Max

What works best

our conclusions

✓ Source experiment pick

Copy the scene, not the faces

Cast: your characters replace the film’s roles — {character A} for role 1, {character B} for role 2; ask when it is unclear
Scene: the film’s own edit — units, angles, cuts and lines from your authorized scene notes; the example run is the coffee-shop scene of Heat (1995)
Pictures: one per camera angle — two for a two-angle dialogue, every face replaced
Voices: the actors’ performance → ElevenLabs Audio Isolation → Speech-to-Speech in the voices from your characters’ cards, the film’s room tone
Video: MiniMax H3 · parts ≤ 15 s ending on a cut · 480P

See every prompt →

Cast first. Your characters replace the film’s roles, one each; ask when it is unclear, never use the demo cast.
Copy the edit, don’t invent it. Units, angles, cuts and lines from the corpus; acting read off the original frames.
One picture per camera angle. A dialogue in one place is two pictures; the prompt does the rest.
Keep the actors’ performance, change the voice to the one on your character’s card. Isolation and speech-to-speech keep every pause where the film has it; the actors’ own voices are never published.
One audio track, exactly as long as the video, with every line’s time in the prompt and no “…” in the text.
Mark off-screen lines and say the face on screen keeps its mouth closed.
The film’s own room tone from its silent gaps, under the voices at the film’s level.
Turbo from angle to angle, Max when an angle comes back.
One part at a time, next to the original, approved before the next.

FlashBoardsRemake a film scene with your characters

The scene, its angles and every part on one board.

Open FlashBoards

FAQ

6 questions

How do I recreate a film scene with AI characters?

Take the scene’s structure from the film — its shots, camera angles, cut times and lines — make one picture per camera angle with your characters in it, re-voice the actors’ lines in your characters’ voices, and generate the scene in parts of up to 15 seconds that end on a cut, each prompt listing the shots with their cut times, acting and dialogue times.

How many reference images does a dialogue scene need?

One per camera angle. A two-person dialogue in one place shot from two angles needs two pictures; which angle plays when, and what the actors do, goes into the prompt.

How do I keep the original timing of the lines?

Don’t re-read the lines with text-to-speech. Isolate the voices of the original, convert each speaker with speech-to-speech into your character’s voice, and lay the result on a track exactly as long as the generation. Word times stay where the actors put them.

How do I make a character speak off-screen in MiniMax H3?

Mark the line as off-screen in the prompt and add a rule that the person facing the camera keeps the mouth closed while the other one speaks. In our test H3 Max kept the listener’s mouth shut and the line in place.

H3 Max or H3 Max Turbo for a dialogue scene?

Turbo image-to-video when the part goes from one angle to another — the first and last frames carry both faces and the audio is pinned exactly. H3 Max reference-to-video when an angle comes back inside the part; it may move the second speaker’s line.

What were the tested scene settings?

H3 Max was tested at 480P for 15 s and H3 Max Turbo at 480P for 13 s, with two setup pictures.

Examples use external tools; available FlashBoards workflows may differ. Downloads are reference guides, not installed automation.

AI Film Scene Remake

Copy the scene, not the faces

filmscene — a reference guide linking the steps

The short answer

Draft or quality: Seedance 2.5

1 · The scene as data

2 · One picture per angle

3 · The actors’ performance, your characters’ voices

4 · One part, one generation

Example run · every generation · video, inputs, prompt

Part 1 · H3 Max

Part 2 · H3 Max

Part 2 · H3 Max Turbo

5 · H3 Max or H3 Max Turbo

How we got here

Settings and measurements

What works best

Copy the scene, not the faces

FAQ

Related