Character continuity — One AI-generated protagonist appears in all 4 shots (uses Vidu's reference2video so the same person shows up in every scene without per-shot drift).
Cinematic visuals — 4 storyboarded shots:
- Pain point (high cost in patient's home country)
- Modern destination clinic
- Wellness recovery in scenic location
- Triumphant outcome with brand CTA
AI narration — Replicate Kokoro TTS (af_bella voice) generates per-shot voiceover, time-aligned to each scene.
Background music — Soft synthesized ambient pad (C-major triad, low-pass filtered, fade in/out).
Brand overlays — Top descriptive captions (so viewers understand the story instantly) + bottom emerald-green brand text on each shot.
Output — Final MP4 uploaded to your Cloudflare R2 bucket, plus all intermediate clips for re-use.

How it works

Step 1: Wavespeed (Seedream 4.5) → 1 protagonist portrait → R2
Step 2: Vidu reference2video × 4 (parallel)  → 4 shot clips → R2
Step 3: Replicate Kokoro TTS × 4              → 4 narration clips
Step 4: ffmpeg concat                          → 30s silent video
Step 5: ffmpeg filter_complex                  → drawtext overlays + audio mix
Step 6: Upload final to R2

Cost & timing

Per run (one full 30s ad):

Item	Cost
Wavespeed Seedream 4.5 (1 portrait)	~$0.04
Vidu viduq2-pro reference2video × 4	~$2.50 (250 credits)
Replicate Kokoro TTS × 4	~$0.001
Total	~$2.55

End-to-end runtime: ~3 minutes (most time is Vidu video generation in parallel).

Required environment variables

VIDU_API_KEY — Vidu Platform API key (https://platform.vidu.com)
WAVESPEED_API_KEY — Wavespeed.ai API key (for the protagonist image)
REPLICATE_API_KEY — Replicate token (for Kokoro TTS)
R2_ACCOUNT_ID, R2_ACCESS_KEY_ID, R2_SECRET_ACCESS_KEY, R2_BUCKET_NAME, R2_PUBLIC_URL — Cloudflare R2 (S3-compatible) for storage

Required system binaries

node (≥ 18)
ffmpeg is bundled via the ffmpeg-static npm package — no system install needed.

Usage

# Customize the SHOTS array in make-xpilot-ad.ts with your storyboard,
# then run:
npx tsx make-xpilot-ad.ts

The script prints the final R2 URL at the end. To iterate on post-production (captions, narration, music) without re-spending Vidu credits, run:

npx tsx xpilot-ad-finalize.ts

This pulls the existing 4 video clips from R2, regenerates narration, and re-composites the final video. Free and fast (~45 seconds).

Example output

Final 30-second ad (8 MB MP4) — narration, ambient music, brand overlays: https://pub-22e3d3e3f43e400493bbd71306cae6bb.r2.dev/demo/medical-tourism-ad/v2/medtravel-final.mp4

Behind-the-scenes assets (all publicly hosted on R2):

Protagonist reference image (Wavespeed Seedream 4.5): https://pub-22e3d3e3f43e400493bbd71306cae6bb.r2.dev/demo/medical-tourism-ad/v2/reference-protagonist.png
Shot 1 — Sticker shock: https://pub-22e3d3e3f43e400493bbd71306cae6bb.r2.dev/demo/medical-tourism-ad/v2/shot-1-sticker-shock.mp4
Shot 2 — Nanning clinic: https://pub-22e3d3e3f43e400493bbd71306cae6bb.r2.dev/demo/medical-tourism-ad/v2/shot-2-nanning-clinic.mp4
Shot 3 — Bama wellness: https://pub-22e3d3e3f43e400493bbd71306cae6bb.r2.dev/demo/medical-tourism-ad/v2/shot-3-bama-wellness.mp4
Shot 4 — Detian triumph: https://pub-22e3d3e3f43e400493bbd71306cae6bb.r2.dev/demo/medical-tourism-ad/v2/shot-4-detian-triumph.mp4

Notice the same protagonist appears in all 4 shots — that's the power of Vidu's reference2video mode, which this skill encapsulates.

Customization

To make this skill work for a different brand/vertical (e.g., "Mexican dental tourism", "Thai cosmetic surgery", "Korean LASIK"), edit:

REFERENCE_PROMPT — describe your protagonist
SHOTS[*].prompt — describe each scene
SHOTS[*].narration — what the voiceover says
SHOTS[*].brandText — bottom brand caption
SHOTS[*].topCaption — top descriptive caption

The pipeline (parallel submission, polling, R2 mirroring, ffmpeg composition) stays the same.

Why Reference-to-Video?

Vidu has three video generation modes:

Mode	Pros	Cons
`text2video`	Simple	Each shot's character looks different
`img2video`	Visual continuity	Hard to change scenes (just continues motion)
`reference2video`	Same character across scenes	Slightly more setup

For multi-shot ads with a recurring protagonist, reference2video is the only mode that works. This skill encapsulates that workflow.

Known gotchas (saved you the debugging time)

Vidu CloudFront URLs contain unencoded ; — don't URL-encode it, that breaks the signature. Mirror to R2 immediately.
OpenAI / OpenRouter quotas run out fast — this skill uses Replicate Kokoro instead, which is dirt cheap.
Replicate rate-limits accounts under $5 credit to 6 req/min — script adds 11s delays between TTS calls.
ffmpeg drawtext apostrophe escaping is unreliable — use full words instead ("should not" instead of "shouldn't").
ffmpeg drawtext % is parsed as variable — escape or use words ("60 percent").
Multiple drawtext filters with commas in text break with , separator — use ; + intermediate labels instead.

xpilot-ad-maker

MedTravel Ad Maker

What it does