How We Made an Animated Tamil Children's Song with AI, Shot by Shot

One sung line, one shot, and the shot shows literally what that line says. That sentence is the entire product. Everything else in this post is machinery for keeping it true.

It sounds obvious. It is not what the first two cuts did. They cut the picture on the beat grid — a new shot every bar or two, because that is how music videos feel — and then laid the lyrics over the top as captions. That guarantees musical cuts and guarantees nothing about meaning. When the line said "roll, roll, come", the character was sometimes rolling, sometimes hopping, sometimes standing in a field looking pleased. My own verdict on cut two, in the message that sent it back, was that very few scenes matched and the others did not match at all. A parent watching with a three-year-old feels that mismatch even if they cannot name it.

So the third cut started from the other end: the words first, then a shot per line, and a test for every shot — can you say in one sentence why this picture is that line? If not, the shot is wrong.

The second cut also had a structural error that looked like a creative one. It contained a thirty-second instrumental intro and a verse that, according to our transcript, was never sung. Both were false. The transcription tool is a speech recogniser, and singing over instruments is not speech: on the full mix it found fifteen words and gave up two-thirds of the way through. Isolating the vocal track first and transcribing that gave fifty-three words and reached the end of the song.

The lesson generalises. Absence of a transcript is not absence of singing. Before a single shot is generated, the song's structure has to come from the audio itself — where the sections actually begin and repeat — and it has to be checked against the written lyric sheet. If the measured section count matches the number of stanzas on the sheet, the analysis can be trusted. If it does not, stop.

One more consequence of measuring: captions are keyed to the moment the singer actually starts each line, not to the bar the section starts on. On this song the gap between the two ran from almost two seconds early to two and a half seconds late. For a sing-along, captions that are out of step by that much are fatal.

A song film has a small cast that must look the same for eighty seconds. Video models do not remember characters between shots, so consistency is something you build, not something you get. We generated one clean reference image per character, stripped every trace of label text out of it — any text left in a reference leaks into the generated footage — and wrote a one-paragraph description of each character that was pasted verbatim into every prompt. Paraphrasing the description between shots is the single most common reason a character's face changes halfway through a film.

The reference image and the prompt must also agree with each other. If the picture shows a yellow top and a green skirt, the prompt cannot ask for a yellow kurta with orange flowers. The model does not pick one; it splits the difference, and the character drifts.

The word in the rule is "literally", and that puts a hard requirement on the video model: it has to do what the verb says. We ran the same prompt — rolls toward the camera, does not hop — through three current models. One produced a character that walked and never rotated. One rolled but lost the character's face partway through the shot. One rolled and kept the face. That third model became the default for the whole film, and the choice was made on that evidence rather than on price or reputation.

Two smaller rules came out of the same shoot. Every shot is generated longer than the slot it will fill and then trimmed, never stretched: slowing a clip to fit is slow motion, and most of the rejected second cut was, on inspection, running in slow motion. And every prompt ends with an instruction to render no text, no captions, no logos and no watermark — the model is never allowed to draw words, because words are the one thing it gets subtly and visibly wrong.

The words on screen come from the written lyric sheet, never from the transcript — automatic Tamil spelling is wrong often enough to embarrass a children's channel. Tamil script also needs proper shaping to render, which the usual command-line video tools do not do, so every caption was rendered as an image in a real browser and composited onto the footage afterwards. The same goes for the title card and the end card: real artwork, placed on top, never generated.

Because the captions are their own layer keyed to the measured vocal onsets, the published video also carries proper Tamil and English subtitle tracks built from the same timings. That matters for a channel marked as made for kids, where parents read along.

The film went live on 18 August 2026 on a channel with no audience and no other videos. Seven days later it had 3,463 views and 47 likes, and the channel had 43 subscribers. Roughly five hundred views a day from a brand-new channel is the platform choosing to show it, not luck, and it is also perishable: a channel that promises a new video every week and then goes quiet loses that push quickly. Three vertical cuts of the same film followed in September. The next episode is the real test.

We do not treat the channel as an income line. A made-for-kids video carries no personalised advertising at all, and children's video in India earns very little per view. It is a portfolio piece for the AI video work we do for clients — the same pipeline, the same rules, aimed at a product or a brand instead of a potato — and it sits alongside our other products on the products page.

Bring the song and the written lyrics. The measuring, the shot table, the cast lock, the shoot and the caption layer are the work — and the work scales with the number of shots, which is why a tight 34-shot film is a better film than a lazy 80-shot one. What an AI video costs is written up separately. The channel itself is at youtube.com/@nimkidsofficial if you want to see the result before you read another word.