B-roll Generator

Generate short, hyper-realistic complementary B-roll clips with Pruna's p-video model on Replicate. The clips show only the people / objects / environments you describe — the main presenter/avatar never appears — so they can play under an avatar voice-over in a final composite (like the inserts used in the original lolo reels).

This is the visual counterpart to the avatar-talking-video skill: that one renders the talking head, this one renders the cutaways that reinforce what's being said.

Setup

Uses the same Replicate token as the other Replicate-based skills (avatar-talking-video, voice-clone, gpt-image-2, ...). If any of those is configured, this skill finds the token automatically.

pip3 install -r ~/.cursor/skills/broll-generator/scripts/requirements.txt
# Only if no sibling skill has a token yet:
python3 ~/.cursor/skills/broll-generator/scripts/setup_key.py YOUR_REPLICATE_API_TOKEN

ffmpeg/ffprobe are used to normalize to the exact length and strip audio (recommended, but the skill degrades gracefully without them).

Quick start

python3 ~/.cursor/skills/broll-generator/scripts/generate_broll.py \
  "manos sosteniendo un smartphone en penumbra, revisando el Instagram de un ex de noche, primer plano, foco selectivo" \
  --duration 6 --camera push_in --avatar-dir /Users/.../virtual-avatar/lolo

Output: <avatar>/broll/<NNN>_<slug>.mp4 plus a manifest.json entry. A JSON summary (saved path + metadata) is printed to stdout for use by an orchestrating skill.

How it works

1. Prompt wrapping — your scene description is wrapped in a photoreal/cinematic style block, a camera move (--camera) plus a "whole frame stays in motion" clause, and an explicit "the presenter does NOT appear / no on-screen text" instruction (skip with --raw-prompt). This fights the static/frozen-background look that makes AI B-roll feel fake. p-video has no negative-prompt field, so these constraints live in the positive prompt.
2. Generate — prunaai/p-video renders text-to-video at the exact requested duration (1–20s), 720p/1080p. save_audio is off for clean B-roll.
3. Normalize — the clip is trimmed to the exact duration and audio stripped via ffmpeg (best-effort).
4. Manifest — every clip is recorded with its description, prompt, camera, model, dimensions and seed.

Making B-roll look believable

The biggest realism win is camera motion — without it, p-video tends to animate only the subject over a static background. Always pick a --camera:

• push_in / pull_out — classic reel inserts; the whole frame travels.
• pan_left / pan_right — good for landscapes, windows, walking subjects.
• orbit — wraps around a subject (objects, a person at a table).
• handheld (default) — organic drift + micro-shake for documentary feel.
• static — only when you genuinely want a locked tripod (scene still kept alive).

Describe subject + action + framing + lighting/mood, one concrete situation per clip (Spanish works well). Avoid naming the presenter or any real person.

Avoid the "mannequin" look (people who don't move/talk). Text-to-video freezes people into stiff poses unless you give them explicit, continuous performance. For any clip with people interacting, pass --action with vivid verbs — gestures, who points/turns/steps, and (for conversations) that their mouths move as they speak:

python3 .../generate_broll.py \
  "una mujer y un hombre discutiendo en la sala de una casa, luz natural fría, plano medio" \
  --duration 6 --camera handheld \
  --action "ambos discuten hablando con vehemencia, las bocas se mueven al hablar; ella gesticula con las manos, señala y sacude la cabeza; él levanta las manos y se gira; intercambian miradas de enojo, postura tensa en movimiento constante"

Key options

Typical workflow (reel composite)

1. Analyze a reference reel with video-scene-analysis to get the per-scene focus, emotion and B-roll descriptions.
2. For each B-roll scene, write a scene description + camera move and call this skill (match the scene's --duration).
3. Render the talking-head segments with avatar-talking-video.
4. A higher-level composite skill stitches talking-head + B-roll under the avatar voice-over.

Notes

• p-video supports any 1–20s duration, so 3s and 6s are produced natively (no padding/retiming).
• Audio is stripped by default so clips are clean for the avatar voice-over; pass --keep-audio to keep it.
• Use --draft to iterate cheaply on a prompt, then re-run without it for the final clip.
• Generation is a long-running operation (~30s–2min per clip).