Even the Choreography Is AI: A Korean Company Moved a Four-Member Girl Group's Entire Production Line Into Its Own Platform
The short version
A four-member Korean girl group with no human members has been reported across Korean media this week — and this time the claim is not just that the faces and voices are synthetic. It is that the choreography is too.
Between November 18 and 19, Maeil Ilbo, Sports Seoul, TenAsia and Seoul Economic Daily all ran versions of the same story: JIRO (지로), a Korean company that operates the AI image-and-video platform Dropshot AI, built a group called SEORIA (서리아) entirely inside its own product. The four members — Riye, Cora, Iseol and Freya — take their names and world-building from Seoljuckhwa, a legendary band of female warriors said to have fought the Khitans during the Goryeo era. Each one embodies an element: fire, earth, water, wind.
The novelty is not "another virtual idol." Virtual idols are already a working business in Korea; PLAVE got there first. The novelty is a single line in the company's own description: choreography now sits on the list of things AI produces. Maeil Ilbo puts it as "not only the music, but the members' appearance, choreography and stage direction are all realized with AI." TenAsia's headline is blunter still: song, choreography and music video, all made by AI.
This article is compiled from multiple media reports. JIRO's official site, dropshot.ai, currently resolves to a domain-parking page behind a Cloudflare challenge, so there is no citable press release; everything below rests on Korean news coverage, and anything that is only the company's own claim is labeled as such.
Why this belongs in a column about AI choreography
For the past two years the division of labor in virtual idols has been stable: humans choreograph, AI renders. Motion capture, face replacement, lip sync, rendering — all of it replaces the shoot, not the act of composing movement.
SEORIA's description moves that line one notch. According to the company, there was no live-action filming, no studio, no outside production house. The world-building, character design, member visuals, concept film and debut music video were all produced inside Dropshot AI. If choreography really is on that list, then what gets compressed is not only shooting cost but part of the upfront work that normally belongs to a choreographer.
That claim deserves a discount. Nobody — not JIRO, not any of the outlets — has published which model generated which routine, or whether the dancing was produced directly from a text-to-video prompt or assembled shot by shot from a storyboard. The accurate reading is this: it is an AI-ified content pipeline, not a music-to-dance motion model. Those are two different technical tracks. The second one — the one covered here before, in systems like Wan-Dancer or AHEADance — takes audio in and puts out driveable skeletal motion. The first takes shot lists and prompts in and puts out finished footage.
How to use it: what keeps the same face from changing
Anyone who has made multi-shot AI video knows the sore spot: change the camera angle or the lighting, and the face changes with it.
JIRO says it solved this with two things. One is its own model orchestration, chaining several generation models in sequence. The other is Canvas, a new Dropshot AI feature that connects image generation and video generation as nodes on a single canvas, so a member's look fixed at the concept-art stage carries through to the final video instead of being re-described for every shot.
In practice that is a three-step order of operations:
- First, lock the person. Fix the four members' faces, hair, wardrobe and color palette at the concept stage. These images get reused constantly later.
- Second, wire the nodes. Use Canvas to connect that image to that video, so every generation starts from the same reference image rather than a freshly written prompt.
- Third, revise shot by shot. Stage angles and group-formation shots are generated separately, so a change touches one node instead of forcing a full re-run.
None of this is magic. It turns consistency from a model-capability problem into a workflow-management problem. For anyone producing serialized content, that idea is more transferable than the model itself.
How big is the company behind it
To avoid mistaking this for a lab demo, some numbers (from ZDNet Korea's November 16 interview with the company):
- Dropshot AI went into beta in October 2025 and launched publicly in January 2026, aggregating more than 70 image and video generation models behind a Korean-language web interface.
- Cumulative content generated: 4.4 million pieces, up from 420,000 in April — a tenfold increase in seven months.
- 510,000 cumulative registered users. 2025 revenue was 4.6 billion won, up 89% year over year.
- An enterprise plan launched in October, marking a push into B2B; a Delaware entity was set up in April, and the company says overseas usage has reached 50%.
So this is not a two-week concept piece. It is a company with revenue, enterprise customers and an overseas push using its own product as a production tool for its own IP. SEORIA is closer to a showroom than a side project.
The timing it landed in
SEORIA's debut single "Burning" came out in August, and its music video has passed 1.16 million views; total video content across the group's channel sits around 2.52 million views, with some outlets citing 2.54 million. The group itself is not new this week.
What made this week the moment for coverage was a separate story: on the afternoon of November 13, a song made with AI assistance — "Have You Become the Adult You Dreamed Of?" by the project MEOMURU — reached No. 1 on Melon's Top 100 for the first time, and held the top spot through November 18. The vocals were generated with Suno; the concept, lyrics and overall production direction came from human creators. The listener breakdown is telling: 36% in their thirties and 30% in their forties, 66% combined.
An AI-assisted song taking the top of Korea's mainstream chart, followed immediately by a group whose choreography is also AI — those two stories landing in the same week is why the coverage clustered.
Who it is a good fit for
- High-volume, short-cycle content: brands and marketing teams that need several short videos a week and cannot justify a standing shoot crew.
- Collaborations and pop-ups: SEORIA has already worked with the athleisure brand andar on video tied to a pop-up in Seoul's Seongsu-dong. One-off activations that still need a star presence fit this cost structure well.
- Multi-language localization: the same character assets move across languages and markets without re-casting.
- Concept validation: test audience reaction with an AI version before committing to a live-action production. Much cheaper than shooting a pilot.
When not to use it
- Anything that needs real human physicality. Muscle engagement, breath, weight transfer — that is where generated video still gives itself away. SEORIA's group-dance footage cuts fast and changes angle often, and that editing style itself avoids long continuous movement.
- Commercial work that needs a clean rights chain. Korea's music copyright association published standards in August for registering the human-led portion of AI-assisted music, then withdrew them amid dispute over where human authorship ends and generation begins. Until rules settle, licensing is guesswork.
- Long-term IP partnerships. A face can be reproduced. Keeping one character coherent across three years is an asset-management and licensing problem, not a generation problem.
- Reading this as "choreographers are replaceable." See below.
Three common mistakes to avoid
One: this is not an AI choreography model. The systems covered here before — Wan-Dancer, AHEADance, ChainDance — take music as input and output skeletal motion that a character or robot can perform. SEORIA's pipeline takes storyboards and prompts as input and outputs finished video. Confusing the two leads to the wrong conclusion about jobs: what shrinks here is production and post-production labor, not the act of composing movement.
Two: an August debut is not a release this week. The debut single went up on Apple Music and Spotify in August. What is new this week is that the production line has been made public, plus the announcement that the process will be shown for the first time on November 24. Keep those two timelines separate and readers will not be misled by headlines.
Three: the company itself is not claiming choreographers are unnecessary. TenAsia's report notes that JIRO intends to work with professional composers and choreographers going forward. That line sits in the same article as "choreography made by AI," which tells you what stage this is at: AI is producing usable dance footage quickly, not replacing the creative judgment behind a routine.
Cost and time: not disclosed
What can be confirmed: per-video cost, model versions and render times were not published in any Korean report, and JIRO has not disclosed them.
What is public is platform-level — 70-plus models, 300-plus templates, an enterprise tier that only launched in October. For anyone evaluating their own project, the practical approach is a 30-second comparison test that counts the human hours spent writing prompts and re-generating, not just the subscription fee.
One date worth watching: November 24. JIRO holds its first press conference in six years, and the same afternoon runs AI CONTENTS SUMMIT 26 at the Fiber Center in Gangnam, co-hosted with Korea's Ministry of Culture, Sports and Tourism and the Korea Creative Content Agency. Co-CEO Lee Jae-seok gives the keynote, and a separate session covers how SEORIA was made. Production process, model choices and cost structure will most likely surface there.
First step
Decide whether you need a person or a shot. If you need a face you can use for years, the work is character assets and a consistency workflow. If you need a steady stream of dance short-form, the work is a shot library and a re-generation budget. Those two investments point in completely different directions, and mixing them wastes both.
Sources
- Maeil Ilbo (Nov 18, 2026): AI 음악 멜론 1위 시대… AI 걸그룹 '서리아' 아이돌 콘텐츠 AI로 구현 — https://www.m-i.kr/news/articleView.html?idxno=2019997
- Sports Seoul (Nov 19, 2026): AI 걸그룹 '서리아', 뮤비부터 '음방' 무대까지 아이돌 콘텐츠 구현 — https://www.sportsseoul.com/news/read/1650852
- TenAsia (Nov 19, 2026): AI 걸그룹 서리아, 노래부터 안무·뮤직비디오까지 AI가 만들었다 — https://www.tenasia.co.kr/article/2026111969514
- Seoul Economic Daily (Nov 18, 2026): AI 노래가 멜론차트 정상에…K팝 제작 공식 바뀌나 — https://www.sedaily.com/article/20116115
- ZDNet Korea (Nov 16, 2026): 지로 '드롭샷 AI' 콘텐츠 생성 7개월 만에 10배 성장 — https://zdnet.co.kr/view/?no=20261116133245
- Madclub (Nov 20, 2026): 지로, AI 콘텐츠 제작 경험 공유하는 'AI CONTENTS SUMMIT 26' 개최 — http://www.madclub.co.kr/news/articleView.html?idxno=16115
- KPOPFIZZ (Nov 19, 2026): SEORIA, JIRO's Fully AI-Built Girl Group, Brings K-Pop Stages to Life — https://kpopfizz.com/news/seoria-jiro-s-fully-ai-built-girl-group-brings-k-pop-stages/