---
name: filmscene
description: "Reference guide for remaking a film scene with your own AI characters — the source reference approach (measured 9.10.2026 on a two-person dialogue scene): the scene's exact units, cuts, setups and lines recorded from your authorized footage; one replaced picture per camera setup; the actors' own performance re-voiced in your characters' timbre (Audio Isolation → Speech-to-Speech) with the film's room tone; MiniMax H3 in ≤ 15 s parts that end on a cut, the edit and the acting written from your scene notes, 480P. Points to filmscene-* sub-skills. Use when the user says remake a scene, recreate a film scene, воссоздать сцену, ремейк сцены, сцена из фильма, диалог из фильма, кино и сериалы. Two external routes: the H3 draft method of this skill, and the quality route — Seedance 2.5 reference-to-video; the source preferred Seedance for quality, but the side-by-side measurement was a product commercial, not every format."
---

These creative references describe source experiments with external tools. Use your own supported account and tools; this download installs no automation. Settings and timings are experimental observations, not current FlashBoards availability. Use only media and voices you have permission to use.

# filmscene · reference workflow

> **Source author's sound preference (9.10.2026), not FlashBoards policy:** avoid ElevenLabs Sound Effects (`/v1/sound-generation`, `fal-ai/elevenlabs/sound-effects`) for effects, room tone, ambience and jingles; the author found them strange and louder than the voices and preferred ElevenLabs for voices (TTS v4, Speech-to-Speech, Audio Isolation). Historical demonstrations below and in related guides still include ElevenLabs Music beds and an earlier Sound Effects ambience test; those are separate experiments, not a universal voices-only rule. For your own mix, use permissioned film sound (line-free gaps, Demucs without vocals), the video model's native sound or supported video-to-audio tools; keep effects under the voice.

Work for the person's own character(s): their photo / sheet / voice from their character card. The example runs below belong to demo characters.

Measured on source experiments (9.10.2026, a two-person dialogue scene re-shot with two characters; the demo film and characters are in the last section). Article with every prompt and picture:
https://flashboards.yaroflasher.com/learn/formats-and-exhibition/ai-film-scene/

## 0 · Cast first

Before anything else, fix who plays whom: **{character A} replaces <film role 1>, {character B} replaces <film role 2>** — the person's own characters from their character cards (sheet, {face board}, {voice A} / {voice B}). Ask the person who plays each role if unclear — never cast demo characters. Every later step (setup pictures, `speakers`, voices) follows this cast.

## 🏆 Source experiment pick — the reference approach

1. **Nothing invented: record the scene from your authorized footage** — units with HD times, setups (A/B/C), cut times, camera moves, subtitle lines with times and speakers → `filmscene-corpus`. The source used a private corpus; neither it nor an extractor is delivered or required.
2. **One picture per camera setup**, not per shot: a dialogue in one place with two angles needs **two** pictures. Every face replaced (the person's characters in, every extra invented, no watermark) → `filmscene-setups`.
3. **The actors' performance, the characters' voices**: the original dialogue → ElevenLabs Audio Isolation → Speech-to-Speech per speaker span into each character's own voice (from their character card). Timing, pauses and intonation stay the actor's, so the lip-sync and the edit land where the film has them. **Room tone from the film itself** (gaps with no lines) → `filmscene-sound`.
4. **Parts ≤ 15 s that end on a measured scene cut**; one prompt per part with every shot: `cut_at`, setup image, camera, action and acting read off the original frames, dialogue with exact times, off-screen lines marked → `filmscene-shots`.
5. **One generation at a time, then self-review** next to the original (your comparison file, e.g. `cmp.mp4`, original on top). 480P. Check cuts, faces and sync before continuing; no approval from the source author is required.



## Two external routes: draft or quality

| Route | Model | When |
|---|---|---|
| **Draft** — the method of this skill | MiniMax H3 Max / H3 Max Turbo, 480P | drafts, many takes, a tight budget |
| 🏆 **Quality** | **Seedance 2.5** reference-to-video (`bytedance/seedance-2.5/reference-to-video`, `task: reference`): the same pictures as `image_urls` (@Image1…), the same audio as `audio_urls` (@Audio1…), 4–30 s a call, native synced sound, 480p / 720p / 1080p | the final cut, when the look matters |

**Source observation (9.10.2026): Seedance 2.5 was preferred for quality; H3 remains the draft route.** Measured side by side on a 15 s product commercial (`ad-spot`): Seedance 2.5 at 720p — cut deviations up to about 1 s in the listed measurements (versus up to 2 s on H3), finer liquids and detail, its own sound effects in sync — took 7 min against 9 s on H3 Max 480P. Not yet measured on every format; write the same shot-by-shot prompt in plain text and refer to the references as @Image1, @Audio1. Compare the routes before choosing a generation.

## Which engine

| The part | Engine | Why |
|---|---|---|
| Starts on one setup, ends on another (A → B), or one setup only | **H3 Max Turbo image-to-video** — `image_url` = first setup, `end_image_url` = last setup, `target_audio_url` = the voices | pins our audio exactly (corr 1.00), cut within 0.2 s, faster in the measured comparison |
| A setup returns in the middle (A → B → A) | **H3 Max reference-to-video** — Image 1/2 = the setups, Audio 1 = the voices | sees both faces; keeps the first speaker exact but may move the second one (−0.95 s measured) — or split the part at the cut and use Turbo twice |

## Classify

| The job | Go to |
|---|---|
| Units, cuts, setups, lines, speakers, part boundaries | `filmscene-corpus` |
| The setup pictures with the person's characters | `filmscene-setups` |
| Voices in the characters' timbre, room tone, checks | `filmscene-sound` |
| The prompt, the request, checks, the comparison | `filmscene-shots` |

## What we tried and dropped (same scene)

| Approach | Why dropped |
|---|---|
| ElevenLabs TTS on one track, montage invented, 7 parts | the second speaker out of sync, a dead voice, not the film’s edit |
| Two voice tracks; then one track with timing in the prompt | two tracks confused the model; with one, H3 moved the second speaker by +0.94…+0.98 s |
| One line per generation, cut by words | H3 adds speech in the silence, stutters on "…", the edit is still invented |
| TTS at the corpus times on Turbo (window w01) | in sync, but the delivery is not the actor's |

## Global rules

1. **No invention**: units, cuts, lines, times, speakers — all measured from your authorized footage and kept in your scene notes; acting from the original frames.
2. **No actor's face on screen**, extras included; the actors' own voice is never published — only the timbre-swapped one.
3. **Only your own cast**: your characters' sheets and reader-selected, permissioned voices; never substitute demo pictures, prompts or clones. Library voice identifiers in the historical example are external samples, not credentials or an automatic cast selection (`filmscene-sound`).
4. **The audio is exactly as long as the generation** (pad with silence to whole seconds); every line has its time in the prompt.
5. **Never "…" in a line** — H3 stutters on it.
6. **One generation, then your own review** — no batches; decide whether it meets your brief before the next part.
7. The source used cloud tools (fal, ElevenLabs) for heavy compute and local ffmpeg for light editing (`-threads 2`). Use your own supported environment; no Mac or author-specific runner is required.

## Example run (reference only — never reuse its files, prompts, faces, voices or frames for a person)

*Heat* (1995), the coffee-shop scene (scene 95; shot units 996–998), re-shot with the demo characters **Dex** (replacing the role facing the camera in setup A) and **Mara** (setup C).

The source's scene files and corpus are not delivered or required. Use your own authorized footage, scene notes, character references and comparison outputs.