narrated-multi-scene

Use when the user needs a multi-scene narrated film, episodic B-roll story, chaptered promo, or several linked video beats with voiceover and no in-scene character dialogue.

Multi-scene AI video (Pruna p-video only)

Each scene = one p-video job (same model, separate predictions). Assembly is outside Pruna (ffmpeg or your editor). No p-video-avatar in this workflow.

See p-video (first/last frame chaining), scene-anchor-triple.md, scene-anchor-pair.md (visual-only alternative), audio-post-production.md, and pruna-api.md.

For visual transitions without narration, use visual-transition-reel instead.

For educational explainers (history, science, nature, how-it-works) with narrator + in-story character dialogue, use interactive-explainer instead of narrator-only tables.

Staged generation: staged-generation-gate.md · workflow-feedback-gates.md

Feedback gates (required)

Phase	What to show	Proceed when
0 — Plan	Scene table, narration lines, style_bible	approve plan
A — Stills	Hero + start/end stills per scene	approve stills
A2 — TTS	audio/narration_*.mp3 per scene — listen	Lines OK
B — Video	p-video clips with embedded VO	approve clips
D — Bed	Optional Stable Audio under concat	User accepts

Execute phases manually or with phased curl — no bundled runner. Never batch p-video before still and TTS review.

Intake: ask before generating

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

Topic	Questions
Story	Order of scenes (1…N)? What changes between scenes (location, time, emotion)?
Per scene i	Primary prompt? First frame (image), last frame (last_frame_image), narration (audio URL)? resolution / fps / draft?
Continuity	Per scene: chain_from_previous only when motion continues (same moment/location). Otherwise composed OPENING still + hard cut. End stills via p-image-edit; extract last frame when chaining.
Audio	Scene anchor triple (preferred): TTS → upload → p-video with image + last_frame_image + audio (omit duration; save_audio: true). Each scene line ≤ ~19s — P-API caps audio-led clips at 20s. Optional Stable Audio bed in post only.
Visual style	Locked style_bible? One specific subject/location per still (painterly illustration, period film still, etc.)? Avoid unrelated branding (e.g. science-show nebula) unless the brief asks for it
Global	Default aspect_ratio for text-only scenes? Global seed policy?
Runtime	Target total duration after assembly?
Assembly	Concat order; narration mux; bed mix volume (~0.08–0.15 under VO)?

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

Scene table (template — fill during intake)

#	Prompt	First frame (image)	Last frame (last_frame_image)	Narration (audio)	Mode
1	motion prompt	start still	end still → scene 2	TTS line → upload	triple
2		= scene 1 end	end still → scene 3	TTS line → upload	triple

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

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 TTS per scene → upload each to /v1/files.

Duration gate (required): after TTS, ffprobe each MP3. If any scene exceeds ~19s, fix before p-video — output truncates at the 20s API max even when input.audio is set. Use validate_narration_duration in runners.

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

Remedy	When	Action
Shorten copy	One beat has too many facts	Cut clauses; keep dates/names; target ≤ ~45 words (~17–18s) per scene
Faster pace	Line is right length but slow delivery	Tighten Gemini style_prompt (e.g. ~2.3 words/sec, brisk, no filler); regenerate TTS only
Split scene	Two story beats in one row	Add scene row + edit_prompt / last_frame_edit_prompt / scene_lines entry; one narration file per row

Phase 2 — Video (parallel when all anchors ready)

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

json

{
  "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 — silent duration-only renders truncate narration on concat. Poll all get_url until done; retry failed scenes only.

Phase 3 — Review

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

Phase 4 — Assembly

1. Concat clips in scene order (narration already embedded).
2. Optional bed — stable-audio-2.5 under VO (launch_background_music.py).

Phase 5 — Manifest

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

Frame-chain + narration example (dog story)

text

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 scene-anchor-triple.md for when to chain vs hard cut, and OPEN/MID/CLOSE prompt structure.

Related

• Single clip: single-scene-ai-video
• Talking avatars: multi-scene-avatar-video
• Audio layering: audio-post-production.md
• Parallel vs phased: parallel-execution.md
• Generic chain: pruna-generative-pipeline