The short drama Chi Xie Ru Song — literally, “bearing arms into the Song” — did not become a top-ten hot search on November 6 because a model turned one sentence into video. It got there because Zhang Hansheng, a 54-year-old former freelancer from Shijiazhuang with a string of failed startups behind him, built a 30-plus-person remote operation that could regenerate the same shot three, ten, or twenty times until it held together [1][2]. Douyin’s vice president Li Liang amplified the series by framing its makers as “grassroots creation” — no film-school training anywhere on the roster [1]. The series topped the AI recommendation chart on Hongguo, ByteDance’s short-drama platform, with 11.41 million favorites, and the team says the earnings have cleared the debts he brought into the trade — and those of at least one teammate who started out owing money [1][2]. The case documents a structural shift: consumer video generation has collapsed the tool threshold, and the bottleneck has moved to human selection and consistency control — the parts a demo reel never shows.
Ten seasons, three view counts
The series behind the story has run ten seasons totaling roughly 1,000 episodes, each under two minutes, since late July 2026 [1][2][3]. Cumulative view counts attached to it are team-disclosed figures relayed in media reports at different dates — 6 billion-plus as of November 3, 4 billion in a November 5 column, 6.6 billion by November 6 [1][2][3]. The team’s make-up is unusual: stay-at-home moms, ride-hailing drivers, van couriers and office workers, all coordinating online, none of them trained in drama production [1][2].
The loop one sentence cannot close
What the production record documents is a loop — generation is one step inside it. Initial preproduction on the series took about half a month: reading the source novel, revising scripts, pinning down character designs, locking a tone-setting sample [2]. The team built dedicated character-asset libraries for main roles and even supporting cast to hold identity steady and suppress face-drift across shots [2]. Generation counts scale with shot complexity: ordinary shots took roughly three regenerations on average, while battle scenes with demanding camera movement and composition often exceeded ten, sometimes twenty [2]. The team rejects the idea that its output is purely algorithmic gacha (blind-box) — shot selection, consistency control and post-production editing remain dense human labor [2]. Inside a 10-person core spanning scriptwriting and directing, editing, production management and IP control, the pipeline runs as a three-stage relay — scriptwriting, storyboarding, AI execution — holding per-episode cost to RMB 1,000–2,000 (about $140–280) [2][3]. The crew started by taking client orders to build skills, then moved to its own IP [2].
What 2024’s platform dramas proved
That loop matches where the technology actually stood as platform productions proved it out in 2024. Douyin and Bona’s Sanxingdui: Future Apocalypse ran on Dreamina — an AI creation platform under ByteDance-owned Douyin — with new control over how far and in which direction the camera moves [4]. Kuaishou showed off a model lineup pairing its Kolors image model with the Kling video model, and separately reported that its five-episode AI drama Strange Mirror of Mountains and Seas cost less than a quarter of a traditional production [4]. For that Kuaishou series, the production team explained, the work skipped text-to-video entirely: images were generated first, then animated, because drama production requires consistent characters, scenes and style that text-to-video could not guarantee — and human authors still wrote the initial script and handled the post-production [4]. Zhang’s team runs the same division of labor — scripts, selection, post-production — at serial scale [2]. No report names the text-to-video, voice or digital-human tools this crew uses, and nothing documented covers how its audio is made; the named models above belong to the 2024 platform productions.
Self-reported money, documented odds
Every money figure attached to this case is a team self-report. The series’ compute bill exceeded RMB 1 million (about $140,000), the chief screenwriter put the commercial return at around RMB 10 million (about $1.4 million), and the team says all of its roughly 40 titles are profitable [1][2]. Caizhongshe (财中社), the outlet with the deepest production detail, states plainly that these are the team’s own claims lacking external verification — no complete per-title revenue, cost or settlement data exists — and that compute is not the full cost (salaries, IP rights, post-production, marketing and platform cuts all sit outside it), while views do not convert to cash at any fixed rate [2].
The wider numbers argue against easy replication. Commission prices for AI short-drama production fell from a 2025 peak of RMB 3,000 per minute (about $420) to RMB 200–300 per minute (about $28–42) becoming common, according to the team’s director He Xiushan and Zhang [2]. Those are quotes clients pay for outsourced production — a different price system from the RMB 1,000–2,000 per episode the crew spends on its own series — and the efficiency gains did not all convert into producer profits, with part ceded to clients through price competition [2]. DataEye counted 221,900 new AI drama and comic-drama titles on Douyin in the first half of 2026, of which 1,055 passed 100 million views — roughly half a percent, with outlets rounding it between 0.47% and 0.48% — and more than 90% of content underperforming [2][5]. The market behind that funnel: more than RMB 22 billion (about $3.1 billion) in January–May, on track for around RMB 40 billion (about $5.6 billion) for the year, up 138% [5]. Volume carries its own failure mode — by season ten, the column noted, opponents had flattened into plot devices, with two-way conflict reduced to one-way steamrolling [3].
Hongguo, Green Bubble, and the new labels
Where this output circulates is shifting as fast as how it is made. Hongguo runs a standalone app built on algorithm-driven recommendation and steady content supply [6]. Tencent’s answer, Green Bubble, is a free WeChat mini-program launched in late September that produces nothing itself — it aggregates third-party production companies and licensed distributors under a free-to-watch, ad-supported model, and from October 16 to December 31 it is paying Q4 incentives of up to 100% of revenue (90% cash share plus a 10% advertising credit) to eligible new dramas that average at least 10 minutes of viewing per user [6].
Regulation is moving in parallel. New measures set out requirements covering filing, review, distribution and disclosure labels for AI-generated content; the labels tell audiences how a work was made, but they cannot replace fact-checking or exempt a production from other content requirements [7]. The scale those rules are catching up to, per the NRTA: 430,000 micro-dramas went online in China in the first eight months of 2026 — thirteen times all of last year — with AI dramas making up more than 90% [8]. The forward signals are concrete: whether Green Bubble’s Q4 incentives continue past December 31, and whether per-minute prices keep sliding. Those two curves, more than any demo clip, will decide whether crews like this one — without a drama-school credit among them — stay a durable part of the supply, with the roughly half-a-percent hit rate as the odds each bet carries.