---
name: avatar
description: "Reference guide for an AI avatar that speaks a voice track of any length — one picture of the speaker and the voice, MiniMax H3 Max Turbo: a silent 2-second stub from the picture, then extend-video step by step (output continuation) with the voice in ≤ 14 s pieces, the stub cut off at the end. Standing talking head, a solo video podcast / YouTube host in 16:9 with two camera angles, or walk-and-talk (room tour, show-around). Classifies the job and hands off to the avatar-* sub-skill with exact settings, prompts and times. Use when the user says avatar, AI presenter, speaker, talking avatar, long monologue, room tour, walk and talk, make my picture speak this audio, аватар, спикер, ведущий, говорящий аватар, рум тур, озвучить картинку, длинная речь."
---

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.

# avatar · reference workflow

**Work for the person's own character: their picture / look sheet / voice from their character card. The example run below belongs to a demo character.**

Measured on source experiments (9.10.2026, see the example run).

Slots used in all avatar-* skills: {character} · {picture} (the speaker, the first frame) · {look sheet} · {voice} · {speech} (the whole voice track) · {place} · {route} (where the character walks) · {mood}.

## 🏆 Source experiment pick — a speaker of any length (MiniMax H3 Max Turbo)

1. **{speech}** → the person's audio as is, or ElevenLabs **v4** with {voice} and emotion tags (`reels-voice`, `reels-emotion-tags`).
2. **{picture}** → the person's frame as is, or a start frame from {look sheet} on GPT Image 2.5 Flare edit, quality low, 1152×2048 (30 s; prompt structure in `reels-lipsync-turbo`).
3. **Pieces** → cut at pauses into even pieces ≤ 14 s, each attached through your supported tool and transcribed with Scribe; keep the words and durations.
4. **Video** → through your supported external tool: a silent 2 s stub from {picture} on `minimax/h3-max-turbo/image-to-video`, then one `minimax/h3-max-turbo/extend-video` per piece (the piece in `reference_audio_urls`, its words and acting beats in the prompt), **output `continuation`** — only the new footage comes back and the clip is assembled locally (sharper than `extended`, measured 9.10); the result **starts with the speech** — the stub is cut off → `avatar-long-speech`.

At 768P: 26–42 s of generation per piece, a minute of footage in ≈ 4 min. The source also listed 480P, 1080P and 2K for extension; these were not measured in this comparison.

## Best result: Seedance 2.5 (30 s a call)

`bytedance/seedance-2.5/reference-to-video` makes **4–30 s per call** with the voice as a reference (`audio_urls`, ≤ 30.2 s a file, `@Audio1` in the prompt) and the frame as a reference image (`image_urls`, `@Image1`): **the same prompt** as an H3 piece, a whole ≤ 30 s segment in one generation — no chain, no stub. The source listed 480p, 720p and 1080p for this route. Compare a whole segment on Seedance with the H3 continuation chain. The source preferred Seedance for quality; this does not establish a measured comparison on every avatar format.

## Classify

| The job | Go to |
|---|---|
| One line up to 15 s, the exact frame matters | `reels-lipsync-turbo` (pinned audio, one pass) |
| One line up to 15 s, the most natural "creator" acting | `reels-talking-head` (H3 Max) |
| A speech longer than 15 s — standing or sitting, talking to camera | `avatar-long-speech` |
| A solo video podcast or a YouTube host, 16:9, a minute and longer — two camera angles, a chain per segment | `avatar-podcast` (+ `avatar-long-speech` for each chain) |
| The character walks while talking: room tour, show-around, walk and talk | `avatar-walk-talk` (+ `avatar-long-speech` for the chain) |
| Pick or make the voice, clone it, add emotion | `reels-voice`, `reels-voice-clone`, `reels-emotion-tags` |

## Global rules

1. **Output `continuation`, not `extended`** (choose the mode explicitly): `extended` re-generates and re-encodes the whole clip at every step and the picture softens faster.
2. **Extend has no pinned audio.** The voice piece goes in `reference_audio_urls` and the prompt says "Audio 1 is <their> voice: <they> say(s) exactly these words…". The words are kept (waveform correlation with the input 0.96–0.98 on a standing speaker), but **pauses and laughs are re-timed** — the clip drifts up to +3 s a minute. Lips follow the clip's own sound: **never lay the original track over the result**.
3. **Pieces ≤ 14 s, at pauses**, and ≥ 2 s (the API: 2–15 s per audio, 15 s combined). The source of each step is the whole clip so far: **≤ 60 s**, so a single chain holds about a minute of speech.
4. **The stub is silent and short (2 s)** and always cut off; the model may put its own noise on the stub frames — cut at the stub's end, not at the first silence (the source assembly removed it).
5. Prompts are blind as in `reels`: never describe {picture} in words — "everything stays exactly as in the video". Expression may be named.
6. Every piece's prompt has its **own words and its own acting beats tied to words** (the `reels-talking-head` way). A shared prompt for all pieces is not measured.
7. `enable_prompt_expansion: false`, no seed.
8. Never reuse the example run's files, prompts, voices or images for a person.

## Not measured yet

- **Dubbing** (the same video in another language with the same voice) — the next test; until then: voice the translated speech with the person's {voice} and run `avatar-long-speech` on the same {picture}.
- A speech longer than a minute in one shot — instead cut to the other angle with a fresh stub (`avatar-podcast`, measured: 96 s from two chains).

## External tools

Use the named image-to-video and extend-video endpoints through your own supported tool. Keep an assembly plan with the picture or stub, stub duration, resolution and each piece’s audio, seconds and prompt. Trim the repeated tail before appending each continuation, then remove the silent stub. No chain runner is included.

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

These output labels and the demo clone identifier are historical context, not delivered or required files or a supplied voice. Prepare your own picture and speech, select a permissioned voice and keep your own assembly plan; no local chain runner is included.

- Demo character **Mara Vale**, 9.10.2026, voice `mara` (`iV8DaEoP6jYNZUgTCz0e`),
- `t1_*`: stub 2.3 s + one extend 7.5 s — the speech is the input file (correlation 0.985 on the first 2.5 s; control: the same file through AAC 1.000).
- `t2_*`: a 27 s line, 2 pieces 13.8 + 13.3 s, standing in a laundromat — 30.7 s; words 0.96–0.98, laughs 0.50–0.58, drift 1.3–3.6 s.
- `room*`: a 56.6 s room tour, 5 pieces 10.4–12.2 s — 60.3 s; speech correlation median 0.67 (≥ 0.9 in 5 of 19 three-second windows), drift up to +2.9 s.
- Podcast «Slow Talk» 16:9 , see `avatar-podcast`: two angles, 2 × 4 pieces, 96 s; `continuation` vs `extended` on the same stub.