On the morning of Nov 22, Weibo carried the trending entry “AI剧里的太监都像太子” — in AI dramas, the eunuchs all look like the crown prince. Commentary published that same morning described the screenshots behind the joke: users placing AI-drama eunuchs beside live-action princes, with the eunuch lead of 《帐中香》 — an ancient-costume rebirth-romance AI drama — as the discussion's centerpiece [1].
《帐中香》 is an AI-animated micro-drama (漫剧) on Hongguo, ByteDance's micro-drama platform. Its male lead, 江佑泽, is written as a true eunuch — a former top scholar whose family was destroyed by imperial suspicion, castrated into palace service and risen to 司礼监秉笔太监, a grand eunuch of the palace's Directorate of Ceremonial, known in-story as the “Nine Thousand Years” — while the character is drawn classically handsome. By the count of an Oct 12 creator-economy analysis, the series carried a post-holiday heat score of 63.87 million, sixth on Hongguo's animated-drama chart, with some viewers crediting the fine modeling and calling the plot thin [2].
The joke lands on a documented behavior of generative-video pipelines: when production is built for speed, output converges on one idealized face. The same-morning commentary described how far the sameness goes — AI gave the grand tutor, the princes and the eunuchs one beautified face, and the difference between the grand tutor and the prince came down to costume color and forms of address [1]. That convergence is the price of a volume-first trade-off: standardized faces survive frame cuts, and frame stability keeps the assembly line cheap. Two forces are now repricing that trade-off — platform acceptance rules that turn character distinctiveness into a requirement, and tool releases that claim reference-based identity control. For anyone producing or evaluating AI video series, the working next step is to watch two signals: whether Hongguo's distinctiveness rules reach enforcement, and whether reference-based identity holds across a full series rather than a demo clip.
Cross-frame stability is why the faces converge
An Aug 28 industry feature that interviewed practitioners gives the mechanism: a video model must keep a face consistent from frame to frame. Symmetric, unblemished faces survive shot changes with the fewest artifacts; distinctive angles, facial hollows and skin texture produce cross-frame face drift and misplaced features, so the model defaults to a standardized face [3]. Fine-tuning deepens the convergence: small and mid-size teams build on the same base architectures, and even prompts written against the default — “high cheekbones, freckles” — get softened back into a standard face by the algorithm [3].
The training data pushes the same way. Statistics cited in the feature, without a named source, put 78% of faces in commercial datasets in the 25–35 age band, 85% light-skinned, and standard beauty-portrait imagery above 80%; distinctive looks stay under 20% [3]. Director Mo Zheng (莫争), quoted in the same feature, traces part of the sameness to mainstream video models having distilled large amounts of Korean drama footage, which imprints a Korean beauty template on output [3]. The grey market feeds it too: a pirated pack holding more than 20,000 copyrighted dramas was reportedly sold for RMB 0.85 (about US$0.12; conversions in this piece are approximate, at roughly RMB 7.1 per US$1) as cheap training material [3].
The complaint wave predates the meme. In early June, the Weibo hashtag #AI脸 生理性厌恶# — physiological disgust at AI faces — had reached 52.2 million reads by the count of a July 3 analysis [4]. The same analysis records viewers complaining that different characters within a single drama share one AI face, and traces the root to cost pressure, efficiency bottlenecks and tool limits pushing teams toward a safer, more standardized assembly line [4].
Volume economics keep the default in place
Removing the sameness costs production speed. The Aug 28 feature describes what a deep-refinement pass adds: detail work raises labor and time costs and slows the line, so most volume productions skip that step [3]. The skip buys scale. AI dramas made to look live-action (AI仿真人剧) drew close to 75 billion views in Q1 2026, more than all AI-drama and animated-drama viewing in 2025 combined, while 93% of AI short dramas stayed under 1 million monthly views [4]. The China Netcasting Services Association counted 239,000 micro-dramas launched in Q2 2026, 153,000 of them AI dramas [5].
Platform incentives fund part of that volume. Douyin Group's short-drama copyright center upgraded its incentive activity for dramas combining live action with AI, with a post-verification cash bonus of up to RMB 400,000 (about US$56,000); Kuaishou's short-drama copyright center pays up to RMB 200,000 (about US$28,000) per drama, and its Kling AI China sales lead said a single project can earn up to RMB 500,000 (about US$70,000) in cash plus a separate non-cash “inspiration points” reward [5]. Actor Ye Xuan's (叶璇) cost claim spread with the Nov 21 entry “Ai剧打败真人剧的真正原因”: as reported by Jiupai News (九派新闻), she argued that dramas under RMB 300,000 (about US$42,000) can pass self-review and air. The same post adds, without naming a source, that AI dramas thrive on genres high-cost live-action shows cannot get approved — 《持械入宋》, 《万妖图录传》 and eunuch romance among them. The post cites an earlier Beijing News (新京报) report for the quality contrast: top live-action drama scores 90–100, and AI has yet to reach that level [6].
Platform rules and tool releases are changing what the sameness costs
Hongguo moved in July. A notice on its creator platform on July 21 raised content-entry thresholds for AI仿真人剧 categories including urban emotion, breakthrough growth, family warmth and urban comeback, demanding more on frame transitions, character images, expression changes and audio-visual sync, and advising rights holders to fix substandard titles and re-upload them [7]. On July 22 it issued 《关于规范AI剧角色创作的公告》, a formal announcement on character differentiation and copyright compliance: main characters must carry clear visual distinctiveness; highly similar images must not be reused across different series; and a special-governance campaign targets “high-frequency AI face” reuse, content homogenization and material violations, with a call for self-inspection [7]. The Aug 28 feature reported the same rule independently [3].
Tool vendors moved in October. Kuaishou put Kling 4.0 into full release on Oct 15 and states the upgrade centers on long takes, subject consistency and controllability: a single generation runs up to 30 seconds, supports up to 10 keyframes, and combines up to 15 reference materials to hold character, product and scene consistency [8]. Those specifications come from Kuaishou's announcement as relayed by a securities-media account. Tencent disclosed that its Hunyuan Video model generates 720P video with 8.3 billion parameters, at 1.87 times the inference speed of comparable models [5]. LibTV announced a 23-character virtual library and portrait-texture controls at the end of June, claims carried on a promotional channel [4].
What to watch next: acceptance rules and reference specs
The next concrete signal is whether Hongguo's distinctiveness rules move from announcement to enforcement, and whether reference-based identity — vendor-stated at up to 15 reference inputs per generation on Kling 4.0 [8] — survives contact with a full series. Until then, the meme works as the cheapest screening question: a pipeline that cannot make a eunuch read as a eunuch without the costume is buying frame stability with character distinctiveness, and the next batch of look-alike eunuchs will show which side of that trade a given tool chose.