narrated-multi-scene

Prerequisites

Install and load these skills before generating (skip if already in context via @pruna):

Or install the full suite once: npx skills add PrunaAI/pruna-skills@pruna -y

Follow each skill's Before generating / craft sections — do not restate guide content here.

Workflow habit

In every reply, name `narrated-multi-scene` in backticks. State phase gates using exact phrases approve plan, approve stills, approve clips (user types these to proceed). Do not same-turn plan + paid video. Skip-review / burn-credits → follow generation-diversity Red flags.

Feedback gates (required)

Execute phases with parallel curl batches — never batch p-video before still and TTS review.

Intake: ask before generating

Open intake → generation-diversity clarification intake.

Do not start scene 1 until the whole scene plan exists in writing (manifest or table):

Ask follow-ups until every scene row has enough to build input without guessing.

Scene table (template — fill during intake)

Mode: T2V · I2V · I2V+last · triple (image + last_frame_image + audio — omit duration)

How the agent runs this

1. Write the scene table (or plan JSON) → approve plan.
2. Hero → parallel p-image-edit start/end stills (pruna-api parallel batches) → approve stills.
3. Parallel Gemini TTS → duration gate on every MP3 → upload → listen → proceed.
4. Parallel p-video triples once all anchors ready → approve clips.
5. ffmpeg concat (± crossfade) → optional bed.

Workflow (after intake)

Phase 0 — Stills (parallel when independent)

1. Hero anchor — one approved p-image or upload.
2. p-image-edit per scene — start still (edit_prompt) from hero; end still (last_frame_edit_prompt) from start still. Parallel after hero exists.
3. Frame chain (selective): set chain_from_previous: true only when scene i continues directly from i−1. Use composed start still + hard cut for new beats.

Phase 1 — Audio (parallel)

gemini-3.1-flash-tts per scene → upload each to /v1/files.

Duration gate (required):

ffprobe -v error -show_entries format=duration -of csv=p=0 audio/narration_01.mp3

If any scene exceeds ~19s, fix before p-video — output truncates at the 20s API max even when input.audio is set.

If a line is too long (pick one or combine):

Phase 2 — Video (parallel when all anchors ready)

Scene anchor triple — one p-video job per row:

{
  "prompt": "...",
  "image": "START_URL",
  "last_frame_image": "END_URL",
  "audio": "NARRATION_URL",
  "resolution": "720p",
  "fps": 24,
  "save_audio": true
}

Omit duration. Always include uploaded audio in input. Poll all get_url until done; retry failed scenes only. Parallel pattern: pruna-api.

Phase 3 — Review

Adjust prompt, stills, or narration; re-run that scene only.

Phase 4 — Assembly

Hard-cut concat (narration already embedded):

# clips.txt: file 'clips/01.mp4'\nfile 'clips/02.mp4' …
ffmpeg -y -f concat -safe 0 -i clips.txt -c copy film.mp4

Optional short crossfade between chained scenes (~0.15s) — use xfade / acrossfade when joins need softness; hard-cut elsewhere.

Optional bed — stable-audio-2.5 under VO:

ffmpeg -y -i film.mp4 -i bed.mp3 \
  -filter_complex "[1:a]volume=0.12[bed];[0:a][bed]amix=inputs=2:duration=first[a]" \
  -map 0:v -map "[a]" -c:v copy -c:a aac film_with_bed.mp4

Phase 5 — Manifest

Scene table + all six URLs per scene (start, end, audio in/out) + prediction ids.

Frame-chain + narration example (dog story)

Scene 1: composed start,  last=play_end,   audio=vo_1   chain→2
Scene 2: extract(clip_1), last=loss_end,   audio=vo_2   hard cut→3
Scene 3: composed start,  last=search_end, audio=vo_3   chain→4
Scene 4: extract(clip_3), last=tree_end,   audio=vo_4   chain→5
Scene 5: extract(clip_4), last=reunion,   audio=vo_5

See video-prompting for when to chain vs hard cut, and OPEN/MID/CLOSE prompt structure.

Related

Related skills: