---
name: avatar-podcast
description: "A solo video podcast or a YouTube host in 16:9 on MiniMax H3 Max Turbo — a reference episode found with vidIQ (breakout videos, its transcript for the structure), the person's own script in segments of ≤ 50 s, two camera angles of one studio (medium + close-up start frames on GPT Image 2.5 Flare), one stub + extend chain (output continuation) per segment, the angles alternating at the cuts, glued with ffmpeg; thumbnail. Part of the avatar reference set."
---

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-podcast · solo podcast / YouTube host, 16:9

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

Slots: {character} · {look sheet} · {voice} · {topic} · {studio} (the place and the light) · {reference} (a breakout episode on the topic, optional).

## Why segments and two angles

- One chain holds ≤ 60 s of source (`avatar-long-speech`), and the picture still softens along a chain (the face ≈ 30 % softer by the last quarter, measured 9.10).
- A cut to the other angle starts from a fresh stub, so every segment starts sharp — and a podcast cutting between a medium shot and a close-up looks normal.
- Segments of 40–50 s of speech: 3–4 pieces each, the source before the last piece stays well under 60 s.

## Steps

1. **Reference (optional)** — use your own supported vidIQ account to look for breakout episodes on {topic}: a keyword, the last six months, at least 50,000 views and up to 8 results were the source filters. A breakout has many more views than the channel usually gets. Read the chosen episode’s transcript and take its **structure**, never its text:
   the host's welcome → the hook (a question the viewer feels) → the turn (a new way to see it) → one or two images →
   one small thing to do tonight → a quiet goodbye.
2. **Script** — the person's own words on {topic} in that structure, in segments of ≈ 90–110 words (≈ 45 s at a calm pace).
   Short sentences, room for pauses; ElevenLabs v4 tags only where they help (`[softly]`, `[warmly]`, `[exhales]`).
3. **Voice** — make one file per segment with your chosen voice on ElevenLabs v4.
4. **Two start frames** of the same studio — `openai/gpt-image-2.5/flare/edit`, quality low, `{"width": 2048, "height": 1152}`,
   {look sheet} as picture 1 (25 s each in the source test):
   ```text
   Picture 1 is a character sheet of one <man/woman/person>. A horizontal 16:9 live-action film still of the same <man/woman/person> hosting <their> <mood> video podcast in {studio}: <they> sit(s) at a desk behind a black podcast microphone on a boom arm, <the lights and the things behind them>. A medium shot from the waist up, the camera at eye level on a tripod, <they> sit(s) a little left of center, the microphone in front of <them> below <their> chin so <their> mouth is clearly visible, <they> look(s) straight into the lens with <a soft calm smile>, about to speak. <Their> face, hair and clothes exactly as in picture 1. Natural film look, shallow depth of field, no text.
   ```
   The close-up: the same prompt with "A close-up from the chest up from a second camera at eye level a little to <their> left, the microphone at the lower right edge of the frame below <their> chin so <their> mouth is clearly visible, <the lights> softly blurred behind <them>". **The mouth must be clear of the microphone** in both — check the frames before video generation.
5. **Pieces** — split each segment at pauses into transcribed pieces ≤ 14 s.
6. **A chain per segment** — odd segments on the medium frame, even on the close-up; job as in `avatar-long-speech` with the
   stub and piece prompts below; select output `continuation` explicitly in your supported external tool.
   Chains of different segments are independent — run them in parallel (job names must differ: `<show>/seg_1`, `<show>/seg_2`).
7. **Glue** — one encode, the clips' own sound (never the original voice files: the lips follow the clip):
   ```bash
   ffmpeg -i seg_1.mp4 -i seg_2.mp4 -filter_complex "[0:v]scale=1366:768,setsar=1,fps=24[v0];[1:v]scale=1366:768,setsar=1,fps=24[v1];[0:a]aresample=48000[a0];[1:a]aresample=48000[a1];[v0][a0][v1][a1]concat=n=2:v=1:a=1[v][a]" -map "[v]" -map "[a]" -c:v libx264 -crf 17 -c:a aac -b:a 192k -movflags +faststart episode.mp4
   ```
8. **Thumbnail** — draw it from {look sheet} with the image tools (face big, one strong expression, ≤ 5 words, high contrast);
   use vidIQ for the words and the idea (`generate_titles`, `generate_thumbnail` with the reference thumbnail in
   `referenceImages` as a style sample). vidIQ draws its own people: `refine_thumbnail` with `subjectImage` failed on its side
   on 9.10 — not a way to put {character} on it yet.

## 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.

## Stub prompt (tripod)

```text
One continuous take from the given first image, a horizontal 16:9 video podcast shot on a still tripod camera. The <man/woman/person> sits still and silent behind the microphone, looking straight into the lens, lips closed, one soft natural blink. Everything stays exactly as in the first image. Quiet room tone only, no voice, no music.
```

## Piece prompt (seated host)

```text
The same <man/woman/person> <starts/keeps> talking straight into the camera, calm and warm, like hosting <their> quiet podcast. Audio 1 is <their> voice: <they> say(s) exactly these words, <their> lips exactly in sync with Audio 1: "<the piece's words>" <acting beats tied to words>. <They> stay(s) seated at the desk behind the microphone. The camera holds still on a tripod. <Their> face, hair and clothes, the studio and the warm lamp light stay exactly as in the video. One continuous shot, no cuts, no music, no text on screen.
```

- Beats are small and seated: a breath before the first word, an open hand toward the lens, a hand on the chest, a nod, a lean toward the microphone on a quiet line, a smile on the last words. Big head shakes and hand moves near the face blur (measured: a dip in sharpness on "taps her temple … shakes her head").
- "starts" for the first piece of a segment, "keeps" for the rest.

## Measured (9.10.2026, the example run)

| | Result |
|---|---|
| Episode | 96 s, 2 segments × 4 pieces, 768P |
| Time | ≈ 35 s per piece; two segments in parallel ≈ 3 min |
| `continuation` vs `extended` on the medium segment | sharper in every quarter: the frame 198 / 185 / 157 / 161 vs 179 / 147 / 114 / 132 |
| Close-up vs medium | the close-up softens less along its chain (≈ −9 % vs ≈ −26 % for the whole frame) |

## Not measured yet

- 1080P against 768P; more than two segments; a cut to a fresh stub inside one segment of the same angle.
- A thumbnail with {character} through vidIQ.

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

The script, start-frame, job and clip filenames below identify historical outputs only; none is delivered or required. Use your own script, permissioned voice, studio frames and assembly plan. In the ffmpeg command, `seg_1.mp4` and `seg_2.mp4` are your own segments and `episode.mp4` is your chosen output; no local chain runner is included.

- Demo character **Mara Vale**, «Slow Talk» episode 1, 9.10.2026.
- Reference: vidIQ `outliers` «solo podcast talking to camera» → *Talk to Yourself Like This and Manifest Anything* (Gentle English Podcast, 16.2k subscribers, 235k views, breakout 62×); its transcript gave the structure (welcome → the voice we talk to ourselves with → words as instructions → images → a small evening practice → goodbye).
- Script `script_a.txt` / `script_b.txt` (47.0 + 43.7 s, voice `mara`, ElevenLabs v4), start frames `try_podcast_start_a.webp` (medium) / `try_podcast_start_b.webp` (close-up), jobs `chain_a_cont.json` / `chain_b_cont.json`, episode `slow_talk_ep2.mp4`; the first cut `slow_talk_ep1.mp4` was made with `extended`.
- vidIQ thumbnail in the reference's style: `thumbnail_1.png` (its own drawn woman, not Mara).