Generate B-roll environment and product shots with Kling 3

B-Roll (Kling 3)

Kling 3 generates B-roll — environment shots, product close-ups, lifestyle moments. These are visual-only clips that play under the continuous voice track. Voice never cuts. B-roll swaps visuals only.

Prerequisites

• Script with [B-ROLL] segments tagged (from /persona)
• A-roll clips generated (from /animate) — needed for environment matching

Environment Matching

Extract a frame from the A-roll and use it as Kling's start image. This ensures B-roll looks like the same room — same floor, same lighting, same vibe. See references/orchestrator.md.

Generate context-specific first frames per scene. Don't use the podcast frame (headphones) for gym B-roll.

Cut Timing

Cut from A-roll to B-roll at a script BEAT point — not mid-sentence. The viewer should see the face deliver the setup, then cut to B-roll for the payoff visual.

Example: If Sarah says "she finishes the whole bowl" → cut to dog finishing the bowl. Match the visual to the words. Leave 0.5-1s buffer before the cut so the face isn't visible in a non-talking state.

Generation

USE scripts/generate-broll.sh — do NOT construct Kling API calls manually.

# B-roll with environment-matched start frame (RECOMMENDED)
bash scripts/generate-broll.sh \
  --provider fal \
  --image workspace/campaigns/<slug>/frames/environment-frame.png \
  --prompt-file workspace/campaigns/<slug>/broll-prompt.txt \
  --output workspace/campaigns/<slug>/clips/b-roll-01.mp4 \
  --log-file workspace/campaigns/<slug>/output-log.md \
  --label b-roll-01 \
  --seconds 5

# Faceless B-roll (no start frame needed)
bash scripts/generate-broll.sh \
  --provider fal \
  --prompt-file workspace/campaigns/<slug>/broll-prompt.txt \
  --output workspace/campaigns/<slug>/clips/b-roll-02.mp4 \
  --log-file workspace/campaigns/<slug>/output-log.md \
  --label b-roll-02 \
  --seconds 5

API Reference

Read references/kling-api.md for Kling 3 endpoint details, field names, and queue workflow on fal.ai.

Key Findings

• Kling generates in ~100s vs Sora's 5-10 min. Use Kling for ALL B-roll.
• Same character face image feeds both engines for identity consistency
• AI cannot generate realistic app screens or UI text. For product/app demos: use real screenshots or screen recordings.
• The orchestrator doc (references/orchestrator.md) has the full A/B-roll split methodology

Brand Memory Integration

Reads

File	Purpose
workspace/campaigns/<slug>/frames/environment-frame.png	Environment-matched start frame for Kling i2v
workspace/campaigns/<slug>/clips/a-roll-*.mp4	Source clips to extract environment frames from
workspace/campaigns/<slug>/scripts/<format>-script.md	B-roll segments and visual cues

Writes

File	Notes
workspace/campaigns/<slug>/clips/b-roll-01.mp4	One file per B-roll segment
workspace/campaigns/<slug>/output-log.md	Kling prompt, duration, provider (append-only)

Context loading

🎞️ B-Roll context loaded:
  ✓ Environment frame: workspace/campaigns/ridge-q1/frames/environment-frame.png
  ✓ A-roll clips: 3 found (workspace/campaigns/ridge-q1/clips/)
  ✓ Script: talking-head (B-roll segments: 2)
  ✓ Campaign: ridge-q1

Contract

Input

• Required: [B-ROLL] script segments plus either an environment frame or usable A-roll clips for extraction
• Optional: product screenshots, faceless B-roll prompts, campaign brief for scene grounding
• Format: workspace image/video files plus B-roll prompt text
• Source: /animate, /persona, and references/orchestrator.md

Output

• Produces: one visual-only B-roll clip per segment plus append-only generation logs
• Format: MP4 files in workspace/campaigns/<slug>/clips/ and rows in output-log.md
• Default behavior: use Kling with an environment-matched start frame; for app demos, use real screenshots or recordings instead of AI UI generation
• Downstream use: /assemble

Validation

• Pre-conditions: the cut point is clear in the script and there is enough scene context to match the A-roll environment
• Post-conditions: clip visually fits the voice beat, feels like the same world as the A-roll, and saves locally
• Failure checks: do not accept stock-looking B-roll, mismatched rooms, or fake app UI when the product needs a real screen

Output

• B-roll clips (MP4) in workspace/campaigns/<slug>/clips/
• Each clip named b-roll-<segment>.mp4 corresponding to a [B-ROLL] script segment
• Clips are visual-only — no dialogue
• Generation params logged to workspace/campaigns/<slug>/output-log.md

Next Step

B-roll done → run /assemble to stitch everything together with unified audio.