visual-transition-reel

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 `visual-transition-reel` in backticks. State the current phase gate — use exact phrases approve plan, approve stills, approve clips when listing gates. Do not same-turn plan + paid video. Skip-review / burn-credits → follow generation-diversity Red flags.

Skill boundary

Montage with transitions between composed video clips — not a picture-book slideshow.

Redirect before intake:

• Picture-book / illustrated slideshow / Ken Burns story with narration → `illustrated-story-reel`
• Cinematic multi-scene B-roll chapters (full p-video-2 audio-led scenes) → `narrated-multi-scene`

When NOT to use

Use a different skill instead:

Feedback gates (required)

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.

Anchor pairs + video-prompting

Start/end stills and transition motion use video-prompting scene-anchor pairs — physically reachable end states; same subject in both plates (identity preserved). Craft OPEN/MID/CLOSE in transition prompts; still OPENING/CLOSING prefixes on p-image-edit plates per image-prompting.

Scene table (template)

How the agent runs this

1. Copy templates/transition-plan.template.json → fill from intake → approve plan.
2. Hero → parallel start stills → parallel end stills → approve stills.
3. Parallel (or sequential for extract-chain) p-video-2-pro standard pair jobs by default; use p-video-2-pro-reference for rows where imported image or video references should condition identity/style/motion/timing/performance, with optional audio guidance or a kept track (keep_audio), p-video-2 for 1080p/draft/scene-anchor triple, or p-video for simpler clips → approve clips.
4. ffmpeg concat ± per-join crossfade → optional bed.

Generation phases

Workflow (after intake)

Phase 0 — Hero (p-image-ideogram)

One approved anchor photo when generating a new image:

1. p-image-ideogram with hero_prompt + style_bible + ritual seed from generation-diversity (p-image for a cheap draft)
2. Slop gate — approve before branching edits

Skip when every scene uses uploaded start/end images.

Phase 1 — Start stills (p-image-edit, parallel)

For each scene without an uploaded start image:

1. Upload hero URL to /v1/files
2. p-image-edit with edit_prompt + hero in images[]
3. Download → {scene_id}.png

Run all start stills in parallel after hero exists (pruna-api).

Phase 2 — End stills (p-image-edit, parallel)

For each scene with last_frame_edit_prompt:

1. Upload start still URL
2. p-image-edit with last_frame_edit_prompt + start still in images[]
3. Download → {scene_id}_last.png

Run all end stills in parallel once start stills exist.

Phase 3 — Video (p-video-2-pro default pair)

Scene anchor pair — one job per row (duration 5–15s, no audio; generated audio comes from the prompt). Iterate mode: cost / speed; final quality:

{
  "prompt": "OPEN: hold. MID: dolly in, subject turns. CLOSE: settle on end pose.",
  "image": "START_URL",
  "last_frame_image": "END_URL",
  "duration": 5,
  "resolution": "768p",
  "mode": "speed",
  "prompt_upsampler": "turbo"
}

1080p, imported audio, or draft previews → p-video-2 instead (720p/1080p, fps, optional audio).

For a scene where imported image or video references should condition the result, use p-video-2-pro-reference: map one role to each image/video asset and any optional audio guidance, remove or prioritize competing references, and treat optional first/last frames as soft keyframes. Audio alone does not qualify the request. The clip carries generated audio; reference audio guides timing or performance unless keep_audio: true keeps exactly one uploaded track as the soundtrack. Use 480p/768p and cost/speed only; validate current limits and reference billing in the tool skill.

Extract last frame when chaining:

ffmpeg -y -sseof -0.05 -i clips/01.mp4 -frames:v 1 stills/02_from_prev.png

Poll all get_url until done; retry failed scenes only.

Phase 4 — Review

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

Phase 5 — Assembly

1. Normalize clip audio (48 kHz stereo) if concat fails on mixed formats
2. Concat — hard cuts:

ffmpeg -y -f concat -safe 0 -i clips.txt -c copy reel.mp4

Per-join crossfade (~0.12–0.15s) on chain joins — use xfade (video) + acrossfade (audio); hard-cut joins stay at 0.

3. Optional bed — stable-audio-2.5 under native SFX (amix, bed ~0.10–0.15)

Phase 6 — Manifest

Scene table + start/end URLs + prediction ids + chain_from_previous flags.

Transition prompt shape

Write video_prompt as motion between the two plates — not a repeat of the still descriptions:

OPEN: [what holds at start frame]
MID: [camera + subject motion developing]
CLOSE: [how motion settles into end frame]

Limits: prefer 4–5s beats; avoid extreme camera whips; start/end plates should differ clearly but share identity and lighting.

When to chain vs hard cut

Related

Related skills: