The Second Half of AI Short Video: From "Can It Generate?" to "Who Directs?"

--- title: The Second Half of AI Short Video: From "Can It Generate?" to "Who Directs?" slug: ai短视频的下半场-从能不能生成到谁来当导演 meta_title: The Second Half of AI Short Video: The Contest Is Over Who Can Direct meta_description: With Kling 4.0 in beta, the 48-episode one-person production Yan Bian Qing Mei passing 100 million views, and Helong winning an award overseas, the competition in AI video has shifted from image quality to performance and narrative control. category: ai tags: AI short video, Kling 4.0, AI short drama, AI comic drama, Kling, likli, AI video generation, controllability, keyframes ---

Over the first few days of October, a handful of things happened in China's AI video scene that are quite interesting when you look at them together.

Kuaishou teased Kling 4.0 at New York Advertising Week and officially announced an October launch. The feedback from people who have already gotten hands-on time with the Fast version focuses not on how smooth the motion is, but on facial expressions, micro-reactions, body language, and sense of rhythm. In the community's words, the bar has shifted from "can it move?" to "can it act?" Around the same time, reports said Kuaishou is beta-testing an AI video creation agent called likli that covers the entire chain from idea to finished film, aimed at short dramas, e-commerce, and brand marketing.

On another front, a creator with the pen name Wanwan single-handedly handled writing, directing, editing, and art direction over two months to produce the 48-episode AI historical short drama Yan Bian Qing Mei. It passed 100 million views on Hongguo, with related topic searches on Douyin exceeding 200 million. The story is about Yan Jiming, nephew of Yan Zhenqing — the young man who is left with nothing but the two characters "met his death" in the Draft of a Requiem to My Nephew.

A few days earlier, at the first Astana AI Film Festival, the 17-minute short film Helong ("Closure") by the Chinese team 2xLabs won Best Story in the open competition section.

The three events belong to product, individual creation, and international jurying respectively, but they point to the same shift: AI video has already cleared the first gate.

The first gate tested image quality; the second tests control

Over the past two years, video model launches have usually revolved around a few quantifiable metrics: resolution, duration, naturalness of motion. These metrics matter, but they answer only one question: can you generate something that looks like a video?

Kling 4.0's specs push those numbers higher: 30 seconds of native generation in a single pass, 4K and 1080p 10-bit HDR, 21:9 ultra-wide framing, multimodal references expanded to 15 items, 10 keyframes, and a prompt limit of 8,000 tokens. In its update, CICC Media characterized it as a "spec war" and noted that 4.0's upgrades in long-take continuity, keyframe control, and multimodal references lower the barrier for high-precision scenarios such as medium- and long-form dramas and film/TV visual effects.

But what's really worth watching is why it was upgraded this way. Is 30 seconds long enough? Is 4K sharp enough? Those questions are pain points for short-film creators, but for a film crew they're just the price of entry. What a crew wants isn't one good-looking shot — it's ten shots that can be cut together.

A one-person production and a seventeen-minute short film prove different things

The value of Yan Bian Qing Mei isn't in the number 48, but in how it was made. Wanwan's approach was to first lock down the historical anchors — Yan Jiming's fate, the loyalty and martyrdom of the Yan clan, the identity of the messenger — and not change a single one of them; where the histories are silent — Yan Jiming's boyhood, how he crossed rebel blockades to deliver messages — she fills in. She sums this up as "being the one who fills in history's blanks."

That approach overlaps with the direction AI video products are evolving in. Generation isn't the hard part; the hard part is making the model know what must not be touched and what needs to be filled in.

The team behind Helong, by contrast, lays out the cost structure more directly. A dozen-plus people, a month and a half, compute costs under 100,000 RMB, of which actual video generation took only a little over a week. The bulk of the remaining time went into polishing the script and settling on a visual style. Executive director Wang Yuchen's remark in an interview is worth writing down: technology updates fast, dazzling visuals are easy to copy, and it's the creator's own thinking and cultural expression that give a work its distinctiveness.

In other words, the part that saves money is generation; the part that costs time is still creation.

Market size and policy are filling in the other half

According to institutional forecasts, the domestic market for AI dramas and AI comic dramas is expected to exceed 40 billion RMB in 2026. Behind that number are moves on the platform side: Douyin, Hongguo, and Kuaishou are pushing interactive AI short dramas, and brands are shifting their budgets this way.

As the content ecosystem grows, so do the problems. AI short drama production capacity is already very high; one tally says 430,000 short dramas were produced in eight months, 90% of them machine-generated. Once capacity outstrips demand, audiences start to distinguish one thing: which films have someone genuinely telling a story. Wanwan's work made it onto the trending lists not because of technology, but because she first figured out how a person should be remembered.

A few thoughts for content teams

For content teams, the real dividing line in the second half of 2026 isn't image quality. Kling 4.0's 4K and 30 seconds will be caught up with; Nano Banana 2.1 halving output prices and Seedream 5.0 Pro making regional editing precise will all become industry standard.

What really separates teams are three capabilities: whether you can break a shot down into reusable keyframes, whether you can keep a character consistent across multiple shots, and whether you can write your creative standards into a workflow so the model executes them over and over. The first two are tool problems; the third is a people problem.

Wanwan produced 48 episodes on her own not because she had mastered a lot of prompt techniques. She had a story worth telling first, and only then did the remaining technical problems have somewhere to land.