Turn still images into a 30-second cinematic reel

What you need

Video generation cost (the only real cost): Veo 3.1 fast ≈ $0.15/sec, standard ≈ $0.40/sec of video. A reel needs six 8-second clips: ≈ $7 fast or $19 standard. Standard is noticeably better with people and animals; fast is fine for pure scenery. Budget 1.5× for a few regenerations.

The handoff prompt

Paste this into Claude Code in an empty folder, with your images in ./stills. Adjust the brackets:

Read https://ai-video-playbook.pages.dev/
and follow its Recipe exactly.

Make a 30-second cinematic marketing reel from the images in ./stills.
Subject: [our seaside guesthouse in Crete / whatever it is].
Mood: [calm, warm, editorial luxury]. Brand name for title cards: [NAME].
Put [people / our dog / nothing alive] in the scenes.
My Google AI Studio key is in $GOOGLE_API_KEY.
Use Veo 3.1 [fast|standard]. QC every clip visually before using it.

Recipe · 1Prepare the stills

Center-crop every image to exactly 1920×1080 (PIL or ImageMagick). Veo uses the image as the literal first frame — off aspect ratios get cropped unpredictably.

Recipe · 2Generate motion clips (Veo 3.1)

One REST call per image — no SDK needed:

apt: POST https://generativelanguage.googleapis.com/v1beta/models/{MODEL}:predictLongRunning
MODEL: veo-3.1-fast-generate-preview  |  veo-3.1-generate-preview
header: x-goog-api-key: $GOOGLE_API_KEY

{
  "instances": [{
    "prompt": STYLE + " " + MOTION,
    "image": {"bytesBase64Encoded": "<b64 of jpg>", "mimeType": "image/jpeg"}
  }],
  "parameters": {
    "aspectRatio": "16:9",
    "resolution": "1080p",
    "durationSeconds": 8,
    "negativePrompt": "text, watermark, captions, distortion, warping,
                       morphing architecture, flicker, extra limbs,
                       deformed faces, cartoon",
    "personGeneration": "allow_all"
  }
}

→ returns {"name": "models/.../operations/xyz"}
→ poll GET /v1beta/{name} until done:true
→ video uri at response.generateVideoResponse.generatedSamples[0].video.uri
→ download that uri WITH the same x-goog-api-key header

Prompt formula

• STYLE (shared): “Photorealistic cinematic film of a real [subject], warm natural light, shallow depth of field, subtle elegant camera movement. Architecture and furniture stay exactly true to the original still image. No text, no captions.”
• MOTION, exteriors: “Slow dolly push-in across the lawn… palm fronds sway, water shimmers, two children run across the grass, a dog trots alongside.”
• MOTION, interiors: “Locked-off tripod camera, fixed framing, no camera movement — only the ceiling fan turns, light shifts, curtains breathe.” Without this, Veo drifts the camera through doorways.
• Have people and animals enter the frame from the side (“a dog trots down the path toward the camera”) — it reads natural.

Recipe · 3QC every clip with your eyes

For each clip, extract a 3-frame contact strip and actually look at it (Claude: use your vision):

ffmpeg -i clip.mp4 \
  -vf "select='eq(n,0)+eq(t,4)+gte(t,7.5)',tile=3x1,scale=1920:-1" \
  -frames:v 1 qc.jpg

Reject and regenerate (or trim) anything with warped architecture, morphing subjects, ugly camera drift, or ghosting. For clips where subjects enter over time, also strip t=1,2,3s to confirm the entry looks natural. This QC loop is the difference between “AI slop” and “how did you make this?”

Recipe · 4Music

Lyria RealTime with the same Google key (pip install google-genai) — stream ~36s to WAV. Prompt like “serene tropical luxury spa ambient, warm felt piano, soft kalimba, airy pads, sparse organic percussion”, with bpm≈88, scale=MAJOR, brightness≈0.55, density≈0.35, negative “vocals, heavy drums, EDM”. Fallback: any licensed 36s track, or local MusicGen.

Recipe · 5Edit with HyperFrames (HTML → deterministic video)

npx hyperframes init reel --example blank --non-interactive, then author index.html:

• Structure — <video> clips as direct children of the composition root, alternating data-track-index 0/1 (DOM order = stacking order); text overlays as <section class="clip"> on track 2; one <audio> music element on track 10. Six ~5.3s scenes → cuts at 5.3 / 10.1 / 14.9 / 19.7 / 25.0.
• Transitions (calm/luxury) — blur-crossfade on the GSAP timeline at each cut T: old clip blurs 22px + scales 1.06 over 0.7s then fades; new clip fades in blurred and sharpens by T+1.5. Add tl.set(old, {opacity:0}, T+1.0) after each fade — the linter requires the hard-kill.
• Ghost trim — a clip whose first seconds are bad plays from data-media-start="3".
• Type — serif display + clean sans, video-scale sizes (headline 66px+, wordmark 150px+), letterspaced uppercase labels, thin accent rules, bottom gradient scrims for legibility, a persistent top hairline with the brand name and a coordinates line. A subtle vignette overlay on top lifts the whole grade.
• Music fades — animate volume on the timeline (in 0.8s, out over the final 2.4s).
• Validate — npx hyperframes lint → validate → inspect, all clean, before rendering.

Recipe · 6Render, QC again, deliver

npx hyperframes render --quality high --output reel.mp4

# then look at a 6-frame contact sheet of the final render (esp. each transition), then:
ffmpeg -i reel.mp4 -c:v libx264 -crf 20 -preset slow -pix_fmt yuv420p \
       -c:a aac -b:a 192k -movflags +faststart Final.mp4
ffmpeg -i reel.mp4 -c:v libx264 -crf 26 -vf scale=1280:720 \
       -c:a aac -b:a 128k -movflags +faststart Final-social.mp4

A 30s structure that works

Total spend for the villa reel this playbook comes from: ~$26 of Veo across two full iterations, one afternoon, zero manual editing.