Personalized AI birthday video

A name, an age and one to twenty photos in; a 15–60 s video that is unmistakably theirs out: their name sung on the hook, their face (or their dog's), the details only their people know. Generic "Happy Birthday" templates are what everyone already gets; the reason to make one with an agent is that it can be about one person.

Generation runs on varg: REST https://api.varg.ai/v2 (the varg-ai skill) or the MCP server https://mcp.varg.ai/mcp (generate_image, generate_video, generate_music, call_tool, get_job, estimate_cost). Quote the price of every clip before it runs.

1. Ask five things, then go

1. Who and how to say it: the name as they spell it and how it sounds, the age (or "don't say it"), and who is sending it (partner, best friend, kids to a parent, a team).
2. Three true details: a habit, a place, a joke, a pet, a dream. They go into the song and the scenes.
3. Photos: one clear, front-facing photo of the birthday person for a face that must match (more angles help), or 6–20 photos across the years for a montage. Ask whether they agree to their face being animated; for a surprise, the sender vouches for it. Never animate a stranger's or a celebrity's face.
4. The format (section 2) and the tone: funny, heartfelt, epic or cute.
5. Where it plays: WhatsApp or iMessage (9:16, under 60 s, under 16 MB), an Instagram story (9:16, 15 s parts), or the TV at the party (16:9, 60–120 s). Ask the date: if it is today, pick a format that is done in one pass (A or B).

2. Pick the format

Formats, the prompts that made the viral examples and their styles: references/formats.md, references/styles.md. The song: references/song.md.

3. Stills first, then motion

1. Make the key still with gpt_image_2_5 (it spells short words right: put the exact text in quotes and spell it out, "the words "MAYA 30", spelled M-A-Y-A then 3-0"). With a face, pass the photo as the reference: generate_image { prompt, reference_image } or POST /v2/image with files. Look at it at full size before spending on video: the face, the hands, the text, no logos.
2. Animate it with the still as the first frame (POST /v2/video with files: [{url}]): wan_3 (best value, 2–30 s) or seedance_2_5 (native sound: the breath, the cheers, the crackle). Write the action in timed beats and say what must not change ("the letters MAYA 30 stay exactly the same").
3. Make the song (references/song.md) and transcribe it to find where the name is sung.
4. Join and export with scripts/assemble.sh (below).

# key still
curl -s -X POST https://api.varg.ai/v2/image -H "Authorization: Bearer $VARG_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"gpt_image_2_5","aspect_ratio":"9:16","prompt":"<still prompt>","files":[{"url":"FACE_PHOTO_URL"}]}'
# animate it
curl -s -X POST https://api.varg.ai/v2/video -H "Authorization: Bearer $VARG_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"wan_3","duration":10,"aspect_ratio":"9:16","files":[{"url":"STILL_URL"}],"prompt":"<timed action>"}'
# poll: GET https://api.varg.ai/v2/jobs/<id> until status is completed, then output.outputs[0].url

4. Join, lay the song, export

bash scripts/assemble.sh -s 9.5 -k birthday.mp4 song.mp3 clip1.mp4 clip2.mp4

It scales every clip to 1080×1920 (-W 1920 -H 1080 for a TV), joins them, starts the song at the given second (the line before the first sung name), keeps the clips' own cheers under it (-k), fades out, normalizes to -14 LUFS and writes H.264 + AAC with faststart: the file WhatsApp, iMessage and Instagram play without re-encoding. Cut a clip before it goes wrong with ffmpeg -i in.mp4 -t 4.6 -c:v libx264 -c:a aac cut.mp4.

5. Check before you send

• The face is theirs in every shot (eyes, hairline, skin, glasses). If it drifts, use fewer seconds per shot or varg-consistent-characters.
• The name is spelled right on every surface it appears (cake, banner, title) and in the song's transcript.
• Candles: once blown out they stay out. In our Wan 3 test they lit again at 4.8 s; cut before that or ask for "the candles stay out, only smoke rises".
• Five fingers, no melted cake, no extra guests appearing from nowhere, no brand logos on the cake or balloons.
• It plays with the sound off: the name and the age are on screen, not only in the song.

Tested

2026-10-30, a fictional 30-year-old "Maya" (no real person): the key still on gpt_image_2_5, 9:16 (22 credits, "MAYA 30" spelled right on the first try); seedance_2_5, 8 s, 720p, with sound, from the still (308 credits: one breath, the candles stay out, smoke, confetti, hands to the mouth, the face steady, the cheers in the native audio); wan_3, 10 s from the same still (100 credits: the same action, but the candles lit again at 4.8 s, so that cut ends at 4.6 s); the song on eleven_music_v2_5, 32 s (8 credits, every line and the name sung, vocal from 9.9 s); lyria_3_5 refused the name "Maya" and the traditional lyric but sang "Lena". Whisper transcription 5 credits. assemble.sh -s 17.8 -k put "make a wish, Maya, blow the candles out" on the blow-out (-14.3 LUFS): https://s3.varg.ai/files/72c55813-c0fa-4cf2-9f64-ec4f7ed0e7ae/file_7e6d05077855.mp4

Rules

• Consent: animate a real person's face or clone their voice only with their agreement (for a surprise, the sender's word that the person would want it). Never a celebrity, a public figure or a stranger.
• Children: make a child's birthday from drawings, toys, a pet or a cartoon character you invent, not by animating the child's real photo. Keep the files private.
• Music: write original lyrics; do not use commercial songs, and do not quote the traditional verse in Lyria.
• If it goes on public social media, say it was made with AI where the platform asks (TikTok, Instagram, YouTube label realistic AI content).

Related skills

varg-occasion-video (weddings, anniversaries, holidays, invitations), varg-music-video (a full music video from the song), varg-consistent-characters (the same face in every shot), hyperframes (lyric captions, title cards, photo montages rendered from HTML), varg-ai (models and prices).