Dheepan Ratnam's classical dance clip blurs the line between AI and direction
Dheepan Ratnam's 30-second classical Indian dance clip holds one hard spotlight, one costume, its jewelry and one face across about 17 cuts, all from a single Seedance 2.5 generation with the soundtrack made in the same pass. It's a 720p render upscaled for posting, he hasn't published the prompt, and the real test, a whole lineup of dancers, is still to come.
A single hard spotlight falls on a black stage. Inside its circle a dancer in white and gold sinks into a deep half-seated stance, arms raised level with her shoulders, her shadow pooled at her feet. Then the camera is close on her hands, bangles and rings catching the light, then on her face under a gold forehead ornament, then wide again as her braid swings through a turn. Thirty seconds later the light is still on her and nothing has slipped.
The clip was posted on X early on Sunday 18 October by Dheepan Ratnam (@Dheepanratnam), a Sri Lankan AI storyteller and digital artist based in London with about 15,600 followers. He made it with ByteDance's Seedance 2.5 and called classical Indian dance "the ultimate stress test for AI video". It is our Video of the Day.
A test with nowhere to hide
This is probably the first time we have looked at a piece of generative footage and not known whether we were watching AI or human direction. The movement is immaculate, and so are the lighting and the sync. From start to finish it is a beautiful piece of footage, and it is genuinely impressive and captivating to watch.
The staging makes that harder to pull off, not easier. Against a stark black backdrop there is little room for error: one light source means every shadow has to fall where that light says it should, and there is no set dressing to pull the eye away from a wobbling hand. In the frames we checked, the costume, the heavy temple-style jewelry, the jasmine in her hair and her face all hold steady across every framing, and the two-hand close-up shows five fingers on each hand with the rings staying put.
We do feel the real test is a whole lineup of dancers rather than just the one. Ratnam is already heading there: he posted a three-dancer clip the day before and told one reply that a group comes next, then more length.
One generation, many shots
His caption calls the piece a "30s single continuous take" with "zero cutaways", and that wording needs care. The video cuts hard about 17 times, moving between the wide spotlight shot, medium shots, a hand close-up and extreme close-ups of her face. Every shot stays on the dancer, so nothing cuts away from her, but it is not one continuous shot. What it is, is one 30-second generation: Seedance 2.5 can produce up to 30 seconds per pass and organize several connected shots inside it, which ByteDance itself describes as one-take creation. Ratnam uses the phrase the same way, writing of his flamenco clip the same day that "every cut and ankle bell came out of one run."
He says the picture and an original soundtrack were generated together, which the model does natively. In a reply he said the clip came from a text prompt alone, and that it was a 720p generation upscaled for posting, even though the file on X is stored at 2560 by 1440 pixels. He has not published the prompt or said how many tries it took. The poses read as classical Indian dance to a general viewer, but no dancer has checked the choreography, so we leave that label as his.
In practice, for anyone making work this week, the lesson is to pick a setting that shows your mistakes. A single top light on a black floor gives the model one rule to obey and gives viewers one thing to judge, and that is what makes this clip persuasive. Recent research from IIT Kharagpur on generating two-hand dance gestures points to thin training data and loose text descriptions as the reasons hands usually break, which is why close-ups that hold up are worth showing off.
A week of dance tests
This was the fourth of eight Seedance 2.5 dance clips he posted in four days, and the follow-up to his breakout rather than the breakout itself.
- Posted (UTC): 16 Oct. Clip: 30-second high-energy music video; Impressions: 164,265
- Posted (UTC): 17 Oct. Clip: "Bollywood isn't ready for this"; Impressions: 659,218
- Posted (UTC): 17 Oct. Clip: Three dancers in one generation; Impressions: 10,914
- Posted (UTC): 18 Oct. Clip: Classical Indian dance (this clip); Impressions: 63,017
- Posted (UTC): 18 Oct. Clip: Bollywood meets flamenco; Impressions: 12,977
- Posted (UTC): 18 Oct. Clip: Mollywood meets Rio Carnival; Impressions: 71,895
The classical piece had about 63,000 impressions by Monday morning, roughly 116 times his median post, helped by quoting the Bollywood clip. The video itself was played about 5,200 times, so most people saw the post without pressing play. That makes it a craft pick rather than a viral one.
Who made it
Ratnam's feed mixes Seedance 2.5 and Kling 4.0 tests, daily AI news threads and spec ads. He is a member of the Higgsfield Creator Partnership Program, which gives partners a plan and extra credits, and he has said that platforms help him with credits so he can test things. This post names no platform and carries no paid-partnership label.
Watch the clip on Dheepan Ratnam's X post at x.com/Dheepanratnam/status/2111646855096312263, sound on.
Clip and frames by Dheepan Ratnam (@Dheepanratnam), made with Seedance 2.5 and shared on X. They are used here with credit. If you are the maker and want any of it taken down, message @the__generated on X and it comes down.
Your Voice
mudras are such a smart pick for a stress test. every finger position carries meaning, so there's nowhere to hide when a model turns hands to soup.
Do the mudras hold up if you pause on a frame? Fingers are usually where it falls apart for me, even when the face looks perfect.
Quoted from public posts on X and credited by handle. If one of these is yours and you want it removed, message @the__generated on X and it comes down.
Got something to say? Tag @the__generated on X