visual-transition-reel

Use when the user needs a visual montage, transitions between stills, action-sequence reel, or multi-scene piece where narration is optional or absent.

Scene transition video (Pruna p-video + p-image / p-image-edit)

Each scene = one p-video job steered by two stills and a transition prompt. Assembly is outside Pruna (ffmpeg concat). Narration is optional — default mode is visual-only (scene anchor pair).

Canonical spec: scene-anchor-pair.md

See p-video, p-image, p-image-edit, parallel-execution.md, and pruna-api.md.

For narrated films (pair + TTS audio), use narrated-multi-scene instead.

Staged generation: staged-generation-gate.md · workflow-feedback-gates.md. Default runner --phase stills.

Feedback gates (required)

Phase	What to show	Proceed when
0 — Plan	Scene table, transition prompts, style_bible	approve plan
A — Stills	Hero + start/end PNGs	approve stills
B — Video	clips/*.mp4	approve clips
D — Bed	Final after concat + optional bed	User accepts

Intake: ask before generating

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

Topic	Questions
Story	Scene order (1…N)? What changes between beats (location, time, emotion)?
Per scene i	Start still (edit_prompt or upload)? End still (last_frame_edit_prompt)? Transition video_prompt (OPEN/MID/CLOSE motion)? duration_seconds?
Continuity	Per scene: chain_from_previous only when motion continues. Otherwise composed OPENING still + hard cut.
Stills source	Generate via p-image hero + p-image-edit, or user-supplied photo pairs?
Global	style_bible? aspect_ratio? ritual_seed? frame_chain_mode (extract_last_frame vs parallel_vignettes)?
Audio	Native SFX only (default), optional stable-audio-2.5 bed in post, or upgrade to triple + TTS?
Assembly	Concat order; chain crossfade (~0.12–0.15s) vs hard cut (0)? Target total duration?

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

Scene table (template)

#	Start (image)	End (last_frame_image)	Transition prompt	Duration	Chain?
1	edit_prompt → still	last_frame_edit_prompt → still	OPEN/MID/CLOSE motion	5s	no
2	= scene 1 end or extract(clip 1)	end still	motion prompt	4s	yes / no

Workflow (after intake)

Phase 0 — Hero (p-image)

One approved anchor still when generating from text:

1. p-image with hero_prompt + style_bible + random seed ritual (SSoT)
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.

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)

Scene anchor pair — one job per row:

json

{
  "prompt": "OPEN: hold. MID: dolly in, subject turns. CLOSE: settle on end pose.",
  "image": "START_URL",
  "last_frame_image": "END_URL",
  "duration": 5,
  "resolution": "720p",
  "fps": 24
}

frame_chain_mode	Start frame when chain_from_previous: true	Render order
extract_last_frame	ffmpeg last frame from prior clip	Sequential for chained scenes
parallel_vignettes	always composed start still	Parallel (montage / hard cuts)
planned_stills	prior scene end still URL	Parallel once all stills exist

Use p_video_payload.py — duration set, no audio.

Poll all get_url until done; retry failed scenes only.

Phase 4 — Review

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

Phase 5 — Assembly

1. Normalize clip audio (48 kHz stereo) if concat fails on mixed formats
2. Concat via concat_clips.py with per-join crossfades
3. Optional bed — stable-audio-2.5 under native SFX

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:

text

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.

Portable runner

Default --phase stills. Phased flow:

bash

mkdir -p output/core/visual-transition-reel/my-transitions/{stills,clips}
cp workflows/core/visual-transition-reel/templates/transition-plan.template.json \
   output/core/visual-transition-reel/my-transitions/plan.json

python3 workflows/core/visual-transition-reel/scripts/run_from_plan.py \
  --plan output/core/visual-transition-reel/my-transitions/plan.json \
  --out-dir output/core/visual-transition-reel/my-transitions

python3 .../run_from_plan.py --plan ... --out-dir ... --approve-stills --phase video
python3 .../run_from_plan.py --plan ... --out-dir ... --approve-clips --phase assemble

Flags: --phase hero|stills|video|assemble|all · --approve-stills · --approve-clips · --assemble-only · --only SCENE_ID … · --regen-stills · --regen-clips · --skip-assembly

Requires PRUNA_API_KEY. Optional bed needs REPLICATE_API_TOKEN.

When to chain vs hard cut

Use chain_from_previous: true	Use chain_from_previous: false
Same location, motion continues	New location or story beat
Subject mid-action into next beat	Emotional pause or time jump
You will extract_last_frame from prior clip	Montage vignettes (parallel_vignettes)

Related

• One beat: image-to-video
• Narrated triple: narrated-multi-scene
• Scenario hub: pruna-generative-pipeline (Recipe Q)
• Example prompt: ./example-prompt.md