← Back to blog
AiAbout 8 min read

Six Amateurs, 1,113 Episodes, a Thousand Yuan of Compute an Episode: How an AI Drama Got Hand-Built

Published Nov 21, 2026
Six Amateurs, 1,113 Episodes, a Thousand Yuan of Compute an Episode: How an AI Drama Got Hand-Built

When it launched on August 21, nobody noticed. The first four days of data were flat. The team reviewed their own cut and found that hit dramas end every episode on a hook, a five-second cliffhanger that makes you tap for the next one. They dropped their flat, linear editing, and the story began to rise and fall. By November 15, when season eleven premiered, the first ten seasons, roughly a thousand episodes, had passed ten billion cumulative views on the platforms' count.

The show is called Armed Into Song. Five modern special-operations soldiers are dropped into the fourth year of the Xuanhe era of the Northern Song. Ammunition is limited, the gasoline burns out and the vehicles become scrap, and the only way to live is to farm, train troops, improve weapons, and sit the imperial exams. The strangest thing about it isn't the premise. It's the people making it.

A 54-year-old former noodle shop owner, and five people who never studied film

Chief planner Zhang Hansheng is fifty-four this year. He ran a noodle shop, then an internet company; neither worked out. He first touched AI last March, put in a dozen-plus hours a day teaching himself, paid for online courses, and studied screenwriting, video generation, and post-production. In June he formed a small group with a few AI hobbyists he had met online and started testing short dramas.

Among the six there is a stay-at-home mother, the owner of a small wholesale stall, a former ride-hailing driver, a delivery worker, and an ordinary office employee. Not one of them went to film school. Another thirty-odd people collaborate from the outside, spread across the country. There is no office. The core creators work from home and hand off footage over WeChat groups and cloud drives. No salary either; they split revenue by effective runtime. Zhang calls the arrangement OPC: one person, one company, everyone their own boss.

The division of labor is clear enough. Zhang picks scripts and sets the tone. Li Zhaohua handles screenplays and text. He Xiushan leads the sample episodes, sets the standard, breaks scenes out to individuals, then collects and reviews the final cuts. Choosing the people took a full week, watching each person's earlier work and handoffs over and over, so that everyone would share one judgment of what counts as good.

That judgment later became the most expensive thing in the whole production.

One to two thousand yuan an episode, and the million rolled in gradually

The "million-yuan budget" that circulates outside is, in He Xiushan's telling, essentially 100 percent compute once you set labor aside.

Compute is held at one to two thousand yuan per episode. Total spend across the series passed a million, but the money didn't land all at once. It rolled in bit by bit from what the earlier quick-turnaround dramas earned. What they cut was location, lighting, extras, equipment rental, and the heaviest line item on a traditional crew: management.

Laid out plainly, that number is easy to misread. Cheap compute doesn't mean a cheap episode. Script adaptation, character design, prompts, review, rework — none of it goes on the compute bill. The team says prep took half a month, most of it on script and prompts, because "that's the part that saves the most rolls." The rolls you don't spend are the money you save.

Rolls: three for a normal shot, twenty for a battle

What actually decides cost isn't the unit price. It's the discard rate.

A normal shot takes three generations on average before it's usable. Battles — many characters, many setups, complicated staging — routinely go past ten, and past twenty is not rare. Their acceptance test sounds plain: no continuity breaks, the performance reads normally, the camera work reads normally. But if the image is "flat," it's a sixty, treated as a discard.

That standard deserves a paragraph of its own. Most teams get stuck on whether the model can generate a thing at all. This team gets stuck on whether it generated the right thing, and whether it looks good. The first problem the models are solving. The second one they can't. So there is no more efficient route to rework. In He's words, it's pure hand-building.

And it's this standard that splits "AI drama is cheap" in two: the cost of making something went down. The cost of making it well did not.

The asset library: no face repeats

A thousand-plus episodes, a hundred named characters, countless soldiers and passersby. The place most likely to collapse is faces.

Their method is a separate file for every character who appears: hair, face shape, temperament, background, regional features, fixed item by item. Not just the five leads. Ministers with a handful of scenes, soldiers, villagers, the young widow all get their own assets. Whoever builds a segment pulls the character design and props they need from the library. Crowd fight scenes are where duplicate faces show up most, and the library is what holds that line.

The Li Qingzhao figure is another example. The team generated at least two hundred images and liked none of them. Eventually He Xiushan proposed restoring an ancient portrait with AI, and that produced the version viewers later described as someone who had walked out of an old painting.

The approach sounds clumsy, but it is engineering used to cover for what the model can't do. Consistency isn't solved by one sentence in a prompt. It's solved by an asset table you can reuse.

A row of hand-made clay heads, each tagged with a handwritten paper label, arranged on dark wooden shelves.

The fifteen seconds the model gave, stretched to thirty

When the creators talk about model capability, they don't mention image quality. They cite two very specific numbers.

One is reference images. Seedance 2.5 expanded the reference set from nine images to several dozen, which helps most with crowds and wide shots, because you can lock the look and costume of more characters at once. The other is shot length, extended from fifteen seconds to thirty. At fifteen, a single exchange of dialogue needs two cuts, and every cut is another chance of a mismatch. Thirty holds a complete piece of staging, and the rework count drops with it.

Neither is the kind of news you would call a leap in capability. But for a team with a thousand-odd yuan of compute per episode, they change what can be shot. A thirty-second shot means you can play a standoff in one go instead of stitching three fragments together.

Which is also why this show looks steadier than most AI dramas. Not a better model, but a clearer map of where the model's edges are, with all the design kept inside them.

221,900 titles and half a percent

The industry numbers are right there. In the first half of 2026, 221,900 AI dramas went live on Douyin. Fewer than 0.5 percent passed a hundred million views. Production cost for an AI comic drama fell from nearly 15,000 yuan a minute to about 500, and per-title cost dropped from the 500,000 to 800,000 of live-action shorts to somewhere between 50,000 and 240,000. Hongguo's Guoran Plan puts 30 to 100 percent of a project's total budget into selected premium titles, started backing AI micro-dramas in the second half of the year, and has nearly twenty in development.

Supply is already in surplus. Attention hasn't grown. At a moment like that, a six-person group breaking out doesn't come down to cheaper compute. It comes down to locking in what counts as good before anyone starts.

A few things here seem worth taking.

Pick scripts without a cheat. Zhang went through a lot of titles in the rights library. He chose this one not for traffic but because it has no system. No infinite supplies, no resurrection; characters can only get stronger by learning. That choice forced the team to spend its effort on character and detail instead of on making the cheat look cooler.

Make prep heavy, make rolling light. The more precisely you fix character design, sets, and prompts, the fewer discards later. The half month spent up front buys time across the next thousand episodes.

Leave the model a boundary. If you know it holds steady at thirty seconds and can lock faces with a few dozen references, don't send it off to stage a complicated forty-second move. What you save is rework.

Read the numbers carefully. Four billion, 6.6 billion, 12 billion have all been quoted. Those are cumulative views at different points in time, not contradictions. As for the rumored ten million a month in revenue and heavy paid promotion, the creators have denied it, and there is still no standard revenue-share data on the commercial return.

Technology can generate a death. It can't generate a real farewell. In season six, Qin Hongying dies in battle. The funeral in episode thirty-nine drew more than forty thousand comments, and the creator responsible for those two episodes wrote a song for the scene. When Zhang heard the song, he knew the man had been moved.

That line is probably the least methodological part of the whole story, and the hardest to copy.

Related articles