avatar-multi-scene

Use when the user needs multiple talking-head segments, motion-transfer comparison reels, mixed host and animate clips, or multi-scene UGC with character continuity.

Multi-scene avatar & motion-transfer video (Pruna only)

Purpose

Produce a coherent multi-scene piece stitched later in your own editor or pipeline (Pruna does not ship a concat endpoint in this skill set). Each beat is one of:

Beat type	Model	Deliverable
avatar	p-video-avatar	Talking-head clip from approved still + voice_script
animate	p-video-animate + slider render	Motion-transfer clip, usually wrapped in a left → right slider comparison MP4 (motion template vs animated subject)

Mix types in one announcement reel—e.g. avatar hook → animate slider demo → avatar CTA (same pattern as the Pruna × Tellers cut, with optional animate beats between speaking scenes).

Visual continuity comes from Pruna p-image / p-image-edit on uploaded references—not from other vendors’ image APIs.

Follow this skill in plain language when talking to the person requesting the video: explain cast, voices, motion templates, and scene order the way you would in a production meeting. Use natural, speakable copy in every voice_script.

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

Quick reference

Resource	Path
Photoreal dynamic personas	realistic-persona-showcase.md
Cast ledger, character sheet, voice/video prompts	prompt-templates.md
Animate rows, sliders, alignment	animate-beats.md
Examples	examples.md
Feedback discipline	requesting-generation-feedback
Slider script	generate_video_comparison.py

Feedback gates (required)

Phase	What to show	Proceed when
0 — Plan	Scene table, read-through, cast ledger	approve plan
A — Stills	Hero + per-scene plates	approve stills
B — Video	Avatar / animate clips + sliders	approve clips
C — Assembly	Concat reel + optional bed	User accepts

Intake: ask before generating

Do not call POST /v1/predictions until the user has answered and you have recorded the answers (use defaults only if the user explicitly opts in):

Topic	Questions
Goal	What is the piece for (pitch, tutorial, trailer, episode)? Primary audience?
Scope	How many speaking scenes or beats? Approximate total runtime after assembly?
Cast	Who speaks, in what order? One character throughout or multiple?
Look	Aspect for stills and feel (9:16 / 16:9)? Avatar output 720p or 1080p?
Voice	For each named character, pick one Pruna voice and voice_language and reuse it in every scene that character speaks. Any words that must be pronounced exactly (names, acronyms)?
Style	Agreed style bible line for all image prompts?
Character sheet	Per speaker: age range, wardrobe baseline, hair, skin/realism level, personality adjectives—record before hero generation (see Character sheet below).
Scene variety	Each scene must differ in camera angle, background/setting, and/or energy—no two consecutive scenes with the same framing and location unless the user asks. Plan visual_style_tag, setting_tag, camera_tag, lighting_tag per row; cast diversity (gender, age, ethnicity) on launch reels — visual-variety-bible.md.
Seeds	Random seed ritual at hero — pick and state a random integer; record as project_seed; reuse for hero regen and every p-video-avatar call unless A/B testing motion.
References	Which files to upload; rights cleared?
Beat mix	Which scenes are avatar vs animate? All avatar, all animate, or mixed announcement?
Narrated B-roll cutaways	Optional p-video beats using scene anchor triple alongside avatar rows
Motion templates (animate beats)	Source .mp4 per animate row—owned/licensed? Match pose/framing to reference still?
Slider delivery (animate beats)	Comparison MP4 only, animated-only strip, or both? Canvas default 1920×1080.
Assembly	How clips will be joined and leveled (tool-agnostic plan)?

If anything material is unknown, ask before the first upload or prediction.

Cast ledger & character sheet

Maintain a cast table in the manifest: one Pruna voice + voice_language per recurring character — never swap presets mid-story unless the user requests a recast.

Before hero generation, fill a character sheet per speaker (age, face, realism, wardrobe baseline, personality, locked project_seed). Templates and manifest JSON: prompt-templates.md.

Rule: New locations and styles = p-image-edit off the approved hero URL — not unrelated fresh p-image identity pulls.

Scene plan (dynamic beats)

Every piece needs a scene table — each row avatar or animate. Example columns and manifest JSON: prompt-templates.md · animate-beats.md.

Motion-transfer alignment (animate beats)

P-Video-Animate animates a reference image using motion, timing, and camera movement from a source video. The better the subject's features, pose, framing, and proportions align with the motion template, the better the result.

Alignment	Typical outcome
Same shot type, similar pose, similar scale	Clean motion transfer; slider demo reads instantly
Same character type, slightly different angle	Good with optional p-image-edit repose toward a template keyframe
Meme / cartoon / mascot on human full-body motion	Limbs, gait, and contact points may warp or slide
Tiny head / extreme proportions on dance or arm-heavy motion	Hands, legs, and depth cues often break
Reference facing camera, source subject in profile	Shoulder/head turn and occlusion artifacts

Rule: Treat severe pose or proportion mismatch as a pre-flight risk. Repose with p-image-edit or pick a closer motion template before burning p-video-animate credits.

Alignment prep (per animate row):

1. Match shot size and facing direction between still and template.
2. Match limb visibility—if the template waves arms, the still must show arms.
3. Repose when close but not exact — p-image-edit from the hero anchor: "Change only: match pose and camera to reference video frame; keep identity and outfit."
4. Run p-video-animate-quality-checklist.md on the pair before animate.

Anti-patterns (all types): two identical office avatar scenes back-to-back; corporate brochure voice_script; human dance template + chibi meme still without repose; serial API jobs when scenes are independent; motion templates that prompt smile/wave only (avatar stays silent — see Motion templates for animate beats below).

Motion templates for animate beats

When p-video-avatar generates a motion template (source video for p-video-animate), treat it as a speaking beat — not a portrait pose.

Field	Requirement
Motion-source still_edit	mouth clearly visible ready to speak — not passive smile only
video_prompt	speaks directly to camera, clear lip movement, explain gestures, head nods — before any wave/smile close
voice_prompt	Delivery throughout the line — not “wave energy at the end” only
Camera	Prefix: Camera moves continuously for the full clip — … never locked-off

Silent motion templates break slider demos and animate transfers. Prompt templates: prompt-templates.md. Full animate pipeline: animate-beats.md.

Mixed reels with animate rows

Combine avatar talking-head beats and animate slider demos in one scene table. Common patterns:

Pattern	Structure
Interleaved	avatar hook → animate demo → avatar proof → animate demo → avatar CTA
Slider-heavy	N animate slider rows → final avatar CTA on hero

End product launches with a speakable avatar CTA unless the user opts out. See animate-beats.md for model roles, alignment, and slider assembly.

Seed policy

1. Hero p-image: set seed once; store in manifest as project_seed.
2. Regenerate hero only: reuse project_seed + same prompt unless the user resets identity.
3. p-image-edit: seed support varies—continuity comes from the hero file URL, not re-rolling identity.
4. p-video-avatar: pass seed: project_seed on every clip for reproducible motion/delivery when the API accepts it; if a scene needs a motion retry, bump seed only for that scene and note it in the manifest.

Natural voice (mandatory for avatar social / founder content)

voice_script = speakable dialogue (contractions, short breaths). voice_prompt = performance direction only — never marketing copy or script text.

Good/bad pairs, per-scene video_prompt patterns, and shared cast voice line: prompt-templates.md.

Source portrait / hero (same character across styles and scenes)

For each recurring character:

1. Land one approved source still via p-image (photoreal prompt + locked seed) or upload. Run the slop gate on the hero before sign-off. Treat the approved file URL as the identity anchor.
2. Every later look—including a new background, emotion, prop, or style variation (e.g. “same cat, but ink-line instead of pencil”)—should be produced with p-image-edit from that same source URL, plus the shared style bible and a short delta (“change only: …”). Do not mint a fresh identity with unrelated p-image text prompts mid-run unless the user resets the character.
3. Each new scene still starts from the same character source as step 2 so faces stay one continuous role across the arc.

This keeps later scene opens and style experiments aligned with the same underlying portrait the user signed off on.

Confirmation gate (mandatory)

After intake is complete and you have drafted work in natural human language:

1. Present a read-through package: scene order and type per row; full voice_script for avatar rows; motion templates + reference stills + alignment risks for animate rows; cast ledger; hero URL(s); chosen resolution; legal/CTA lines verbatim if supplied.
2. Ask clearly for approval (e.g. “Reply approve or go when this script and cast are final.”).
3. Do not upload binaries for generation, call POST /v1/predictions, or run automation until the user explicitly confirms.

Optional edits after feedback; repeat confirmation if the script or cast changes materially.

Script and run package (after confirmation)

Once the user confirms:

1. Write a concrete generation package: phased curl steps or a small script that performs uploads, p-image / p-image-edit, parallel p-video-avatar (avatar rows), parallel p-video-animate (animate rows), generate_video_comparison.py slider renders (animate rows), and downloads—matching the approved scene table exactly. Parallelize independent lanes within each phase (parallel-execution.md).
2. Execute that package when execution is possible (PRUNA_API_KEY present, network available). Prefer one subagent per scene lane (still pipeline: edit → gate; or avatar: create → poll → download) launched in parallel after the hero anchor exists. Parent agent owns confirmation, manifest merge, and assembly. If the environment cannot call the API, hand the user the same script and exact commands so they can run it locally without guesswork.

The script is the contract: what runs must match what was approved.

Core rules

1. p-video-avatar input.image — use an approved still URL from /v1/files (upload, p-image, or p-image-edit output) that passed generation-quality-checklists.md.
2. Run the slop gate on every hero and scene still before any avatar job.

text

Hero:     p-image (or upload) → slop gate → approve anchor
Scene N:  p-image-edit(anchor) → slop gate → p-video-avatar

Use the approved hero as the reference for p-image-edit, not a rejected intermediate.

API surface (this workflow)

Step	Model	Skill
Upload binaries	POST /v1/files	pruna-api.md
Style-locked stills	p-image, p-image-edit	p-image, p-image-edit
Talking clips	p-video-avatar	p-video-avatar
Motion transfer	p-video-animate	p-video-animate
Slider comparison (animate rows)	generate_video_comparison.py	local; install via npx skills add PrunaAI/pruna-ai-content-generation-skills/skills --skill avatar-multi-scene --agent cursor -y

Use PRUNA_API_KEY and the apikey header on every call. Async + parallel by default: batch all avatar jobs once approved stills pass slop; batch all animate jobs once motion + still URLs are ready; poll all get_url together. See parallel-execution.md.

Parallel execution & subagents

After the confirmation gate and hero anchor are locked:

Phase	Parallel?	Subagent split
Hero p-image → gate	Sequential	No — identity anchor
Per-scene p-image-edit	Yes — all scenes	One subagent per scene still lane
Slop gate	Yes — review in parallel	Parent or per-lane subagent
p-video-avatar	Yes — all avatar rows	One subagent per clip (create + poll + download)
p-video-animate	Yes — all animate rows	One subagent per clip (create + poll + download)
Slider render	Yes — all animate rows	One subagent per comparison MP4
Assembly	Sequential order only	Parent agent

Rule: Never dispatch subagents before user confirmation. Parent merges all lane outputs into one manifest.

Workflow

Step	Action
1–3	Intake → speakable script → confirmation gate (no API until approve)
4–5	Upload refs → p-image hero per character (locked seed) → slop gate
6–7	Parallel p-image-edit scene stills → slop gate each
8	Parallel p-video-avatar (cast ledger voices, unique video_prompt per scene)
9	Parallel p-video-animate + slider renders — animate-beats.md
10	ffmpeg concat ± optional bed — stable-audio-2.5
11	Manifest: paths, prediction ids, slop notes, cast snapshot

Field names and curl shapes: prompt-templates.md · p-video-avatar.

References

• generation-quality-checklists.md
• prompt-templates.md — cast ledger, character sheet, voice/video templates
• animate-beats.md — p-video-animate, motion templates, sliders
• requesting-generation-feedback
• examples.md

Related

• Pruna-only pipeline overview: pruna-generative-pipeline
• One-scene avatar: single-scene-avatar-video
• Cinematic B-roll (non-avatar): single-scene-ai-video, multi-scene-ai-video
• Still upscale slider demos: p-image-upscale-comparison
• Motion transfer tool: p-video-animate