---
title: "One Storyboard, 1,000 Frames: When Technology Goes Free, Taste Becomes the Only Hard Currency"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-09-30T14:11:36+00:00"
canonical: "https://ffcap.cn/en/research/one-storyboard-1000-frames-when-technology"
source: "https://uniqueresearch.substack.com/p/one-storyboard-1000-frames-when-technology"
language: "en"
---

# One Storyboard, 1,000 Frames: When Technology Goes Free, Taste Becomes the Only Hard Currency

[![cover](https://substackcdn.com/image/fetch/$s_!bCnS!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc9d4449-40d3-4e2b-bde5-1b3f2b5406f8_2730x1536.jpeg)](https://substackcdn.com/image/fetch/$s_!bCnS!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc9d4449-40d3-4e2b-bde5-1b3f2b5406f8_2730x1536.jpeg)

People, not AI, set the ceiling on content.

For one storyboard, they first let AI generate 1,000 images, pick the one that works, and only then move into real production.

That is the daily routine of Nanqiang, general manager of Shenzhen-based Aimenvise (艾门韦思). At an AI imaging roundtable during the Chengdu AI Entertainment Conference, he sat alongside three other frontline operators: Zhu Leimeng, head of Mengjue Studio, which makes feature films and anime series; Sun Tingyan, founder and chief director of Chengdu Xin Shijue AIGC, a company of 400–500 people working from vertical short dramas up to AI feature films; and Chen Qizhi, a visual artist who trained in fashion design and worked on the front line as a photographer.

All four share one thing in common: they make a living from AI imaging.

But the interesting part is this: across the whole roundtable, the thing they talked about most was everything AI cannot do.

My takeaway was direct: in this line of work, technical problems are disappearing at a visible pace, and human problems are getting more expensive.

The most striking line of the roundtable came from Zhu Leimeng.

Character consistency — the number-one problem the AI imaging industry was collectively fighting last year — got a one-sentence answer from him.

“Character consistency doesn’t need me to solve it; let the model handle it. Last year it was still a big problem. Now that the models have iterated, a lot of it is already resolved.”

Chen Qizhi’s experience is a footnote to that line. From 2023 to 2024, the industry mainly solved character consistency by training LoRAs. In 2024 he personally built a complete workflow covering image compositing, face swapping, background replacement, and color grading. And then? “Within two months the model updated, and the old workflow was basically meaningless.”

Two months.

A technical moat you spend months building may have a shelf life of just two months. Model capability spills over faster than any content company can build workflows. Grinding away at the tooling layer is a win that only lasts until the next release.

So what never expires?

Nanqiang put it bluntly.

“Only polished output counts as a work. What AI mass-produces is raw material — alternatives, or frames that get discarded.”

That’s the story behind the opening number: one storyboard, hundreds or thousands of images tested, one selected. Or 1,000 shots run to test camera transitions, character emotion, light and color. He says it’s very much like live-action production — film sets used to generate mountains of rejected footage, and what AI can mass-produce is exactly that layer: the outtakes and the backup options.

“A lot of people run a video once or twice, maybe two or three times, and stop. Without comparison, you never know which one is best.”

Sun Tingyan’s framing is more analytical. He splits content into two layers: fast-food, punchy, fragmentary products belong to AI’s efficiency; but polish is another matter. “AI plays the role of tool and efficiency. Whether something is polished is defined by people. People set the ceiling on content, not AI.”

The two are saying the same thing: AI has turned “production” in the content industry into “generating candidates.” The human role shifts from maker to judge.

Mass production is no longer scarce. Judgment is.

Narrative was the consensus hard bone of the roundtable.

Zhu Leimeng: “If you’ve used ChatGPT or Claude to write a script, you’ll notice the AI taste is strong — it’s hard for a script like that to generate strong empathy.” Mengjue Studio makes AI films, and screenwriters remain essential. They almost never use AI to write film scripts; at most it serves as a creative assistant in the planning stage.

He didn’t deny what AI can do: re-creations, short pieces, seven-to-ten-minute content — AI may produce something that catches the eye. “But when more and more content reaches the same level, it stops being scarce.”

That sentence punctures the logic of AI-content saturation: technological leveling makes “pretty good” available to everyone, and “pretty good” becomes worthless.

Worse, the audience has changed. Zhu Leimeng described an emotion he has observed: the “flipping the table” mentality.

“I don’t need you to feed me anything. I watch whatever I want to watch, and if I want it, I’ll watch it all the way to 60 million box office. That’s my right.”

Once the audience holds the choice, self-indulgent creation has no future. Nanqiang’s angle starts from character consistency: the industry is no longer satisfied with a stable face. Pre-production now has to define personality, habitual gestures, verbal tics — how a character laughs in act one, at which node their personality shifts in act two. “Fixing only the appearance is useless. Scripts from AI may always be just text piled together, not knowing what the audience wants to watch.”

Sun Tingyan contributed a methodology. They are making a horizontal mid-length drama with 200 scenes; before shooting, costume assets alone must be organized down to each scene and each costume in five shot scales — wide, full, medium, close-up, extreme close-up — plus three-view drawings, with special lighting and motion built as separate assets. That is a huge investment before formal shooting begins.

His other principle is more practical: “Don’t try to brute-force the model’s limits. If it can’t do a 60-second long take, use director’s thinking to find a controllable alternative that doesn’t hurt the narrative too much. I don’t fight the model’s capability head-on.”

He offered a contrast: Wang Luodan is also making AI films and is very picky about one extremely precise shot. The model couldn’t produce it, so she kept pulling cards. Sun Tingyan says there’s no right or wrong — an independent creator can pursue perfection on a few minutes of footage, even with PS cutouts. But as a content company of four to five hundred people, beyond expression you have to count labor efficiency and productivity; he wouldn’t let the whole team work that way.

See — even “should we die on this one shot” is fundamentally a human judgment: who you are, and how you do your math.

After the technical threshold collapsed, hiring standards are the most honest industry barometer.

Zhu Leimeng is currently hiring AI directors and producers. In the requirements for AI director, aesthetics accounts for more than 70%, with narrative ability right behind: you don’t have to know how to write a script, but you must know where a character’s performance goes, where the tension of a scene lives.

A mature AI film crew needs roughly seven roles: director, art, DIT, producer — plus a new job type called the “generator” (generative artist), known in the trade as the “card puller.” A generator doesn’t need full narrative understanding, but must understand basic cinematography language, otherwise the waste-card rate rises and cost control breaks down.

Sun Tingyan prices aesthetics even higher: above 90%. He says that when he scrolls Douyin at night and sees a video with an exceptionally good aesthetic style, he’ll go into the backend to dig out the creator. “Even if the person is only strong on aesthetics, everything else can be handed to the rest of the team.”

A boss of a 400–500-person company spends his evenings scrolling short-video platforms to recruit people. That detail says more than any industry report about how scarce taste is right now.

Chen Qizhi’s explanation is the most interesting. He brought up a word from Hideo Kojima, the producer of _Death Stranding_: “meme.” Genes are determined by innate DNA; memes are accumulated from lived experience. “Many AI directors weren’t originally in the film industry. With AI as a tool that makes ideas easier to realize, the accumulation they carry inside is magnified at this moment.”

Tool leveling erases the gap between “can do” and “can’t do,” and in doing so magnifies the gap between “has seen the world” and “hasn’t.”

Zooming out to the whole industry chain, the bottlenecks the four named were almost never purely technical.

Sun Tingyan summarized three: people who can participate in production, relatively cheap tokens, and better creators. Compute being expensive is a real cost pressure, but what’s scarcer than compute is “people with good minds.” Chen Qizhi added something concrete: optimizing prompts can cut a single image-generation’s consumption from 500 words to 200; small teams can’t move compute prices, but they can optimize algorithms and usage to squeeze more output from the same specs.

Zhu Leimeng, standing on the platform side, watches both ends. Upstream is IP: amid a screen full of AI shorts, which ones are worth developing? The standard is whether the worldview can expand and support a long-form story. Downstream is production capacity: individual creators can polish slowly, but film and TV projects demand going from development to broadcast within three to five months, with cost, schedule, and promotion controlled. “If that person has something going on today, the project has no production capacity today. That doesn’t work.”

So the answer still ends up like a traditional crew: everyone does their own job, with a general director controlling the whole process.

Sun Tingyan also shared a case happening right now. They are adapting the big IP _I Was a Taoist in Those Years_ (《我当道士那些年》). The first review was bounced back — the original involves a lot of metaphysics and ghost elements, making adaptation a technical craft. She herself is a fan of the original: “Some changes I also hate. But sometimes there’s no way around it. What we can do is preserve the spiritual core of the original and still do the compliance work that must be done.”

Nanqiang named a more hidden bottleneck: quality control simply doesn’t align. Upstream and downstream want different things, and there’s no consensus even on “what counts as polished.”

“I’ve received many sample reels and I’ll say it straight: this doesn’t even qualify as B-level or C-level — it can’t be called polished in our book. But the other side thinks that in some markets it is polished.”

When everyone can produce content, “what is good” becomes the thing the industry can least agree on.

Near the end of the roundtable, an audience member asked a very pointed question: what if you submit your work and it sinks without a ripple? Sun Tingyan’s answer was almost zen: “If you think it’s good, then it’s good. Box office has an element of luck to it too — works that are both acclaimed and commercially successful have never been common. If your mindset is off, just make some short videos and play with them for yourself.”

It sounds like consolation, but it’s actually the plain truth. When AI flattens the barrier to production and the cost of making content approaches zero, the one thing that was never leveled is the judgment you’ve accumulated from the films you’ve watched, the books you’ve read, and the mistakes you’ve made over the past decade or more.

Tools iterate. Models fill in the gaps. Technical moats refresh every two months.

The moat that remains in the end is your own meme.

**Moderator**: Huang Jingrui, VP of Unique Capital

**Guests**: - Zhu Leimeng, head of Mengjue Studio - Sun Tingyan, founder & chief director, Chengdu Xin Shijue AIGC - Nanqiang (Rylee), general manager, Shenzhen Aimenvise - Chen Qizhi, visual artist

**Huang Jingrui**: Generative AI is reshaping the way creative work is made, raising both the speed and scale of content supply. Before the panel starts, I’d like each of you to take a minute or two to introduce your company and what you do.

**Zhu Leimeng**: I’m Zhu Leimeng, head of Mengjue Studio. We mainly work on AI-created content, developing long-form films and anime series. Personally I take on the producer and director roles.

**Sun Tingyan**: I’m Sun Tingyan, founder and chief director of Chengdu Xin Shijue AIGC. We’re an AI content production company, currently around 400–500 people. We make vertical short dramas — including overseas vertical short dramas — plus horizontal mid-length dramas and AI feature films.

**Nanqiang (Rylee)**: I’m Nanqiang, head of Shenzhen Aimenvise. We mainly produce AI creative content, including flat visuals and video, spanning domestic and international entertainment content and commercial ad videos.

**Chen Qizhi**: Hello, thanks to Unique Capital for the invitation. I trained in fashion design first, then worked on the front line as a photographer, so I understand the whole pipeline from early-stage creative to post-production. To me, AI is a more efficient tool. I work on image-asset optimization — solving how content assets go from “looking good” to “being useful.” I’m also currently serving the evaluation of an AI platform.

**Huang Jingrui**: AI is no longer just an efficiency tool. First question, starting from the most practical choice a team faces: how do you balance fast, scaled output against the pursuit of polish? Let’s start with Leimeng.

**Zhu Leimeng**: Fast iteration versus polished content — since AI arrived, we’ve been facing this question: can we be fast, good, and cheap all at once? For the platform, we hope to use AI to support more creators and creative ideas. But technology isn’t really the hard part now; there are still many barriers in visual cinematic expression that creators have to keep breaking through to reach what we’d call polish.

How do you define polish? Polish means using long-form narrative logic to attract and hold an audience — both image quality and narrative have to stand out. And these capabilities can’t yet be granted to a person by AI alone. So we have to find talented creators to make polished work.

Fast output can be validated. We have a short-drama channel and a manga-drama channel where creators can produce quickly in a UGC atmosphere. But how to balance is still a hard problem — I can’t offer a ready-made solution. When making AI imaging narrative works, I still put narrative and cinematic thinking first, and put tool-level equalization second. That’s my personal view.

**Huang Jingrui**: You mentioned barriers — what kind of barriers specifically?

**Zhu Leimeng**: Many creators use AI spontaneously: “I make whatever I want to make.”

But the platform ecosystem has to put the audience first. What does the audience want to watch? Some works weren’t made with that in mind. And now audiences have this “flipping the table” mood: I don’t need you to feed me anything; I watch whatever I want, and if I want, I’ll watch it all the way to 60 million box office — that’s my right. So we have to make content with audience emotion in mind, not self-indulgence.

If we only use AI to cut costs and re-serve old wine in new bottles, that’s irresponsible to the medium. We want to make responsible works that give the audience more to expect in narrative and image.

**Huang Jingrui**: Sun, your view?

**Sun Tingyan**: We’re a content production company, so I have to look at this by layers. Fast output — the efficiency layer — has its matching products. Some products are fast-food style, punchy, fragmentary; AI genuinely has an advantage in efficiency there.

But polish is a human problem. AI plays the role of tool and efficiency; whether something is polished is defined by people. Different products have different applications — people set the ceiling on content, not AI. Some products suit fast output, so use AI to nail that part. The key is distinguishing product and content types.

**Huang Jingrui**: So it’s product-driven. Nanqiang?

**Nanqiang (Rylee)**: I don’t think mass production and polish conflict. AI can mass-produce raw material quickly, but you have to separate the two: only polished output counts as a work; mass-produced output is raw material, backup options, or discarded frames.

We may test 1,000 images and pick one as a polished storyboard; we may also run 1,000 shots to test camera transitions, character emotion and performance, light and color. It resembles the live-action process — sets used to generate tons of rejected footage, and what AI can mass-produce quickly is exactly that layer.

Then you pick the most suitable one and turn it into the final polished content — that’s where people come in. So AI mass production and content polish need to combine.

**Huang Jingrui**: As capacity becomes easier to obtain, will the definition of polish change?

**Nanqiang (Rylee)**: Polish first means standing out from the crowd — most people who see it think it’s good, and it’s ahead on at least some dimensions. That’s also why we mass-produce raw material up front: without comparison, you don’t know which is best. Many people run a video once, twice, or two or three times and stop. But we keep testing — one storyboard often produces several hundred or several thousand images before one is chosen.

**Huang Jingrui**: On capacity —

**Sun Tingyan**: — and quality, as a pair.

**Chen Qizhi**: As large models iterate, quality and efficiency gains have been proven in the market. Take product development: before, an internet product needed at least four or five people — front-end, back-end, ops, product manager. Now one product person plus one backend-leaning person, two people, can do it. A demo used to take three months; now it takes one month.

That’s the lift from model iteration. For us, my idea is to reduce wasted effort. We’re not a giant with huge data and compute to develop our own model or tool that further raises efficiency. In this link we should do good tech selection to guarantee more efficient output.

As for quality — especially in film and TV — it’s more about subjective aesthetics and high-spec requirements. AI still can’t handle every detail in one pass; some polished, cinematic-grade texture still needs traditional tools for further refinement. That’s my view.

**Huang Jingrui**: Leimeng just said complex narrative is a difficulty in generative imaging. Second question: character consistency and complex narrative — what problems have you actually hit, and how do you solve them? Start with Leimeng.

**Zhu Leimeng**: Character consistency doesn’t need me to solve it; let the model handle it. Last year it was still a big problem; now that models have iterated, a lot is already resolved.

Narrative does need people. If you’ve used ChatGPT or Claude to write a script, you’ll notice the AI taste is strong and it’s hard to generate strong empathy. We make features more, and feature-film logic has to be clever enough to keep the audience watching. Audiences’ attention and how they consume have changed — beyond TV and cinema there are games and other content. Your narrative has to be clever enough, and the story core precise.

AI may produce something striking on re-creations and short pieces — six or seven minutes, or seven to ten minutes. But when content at the same level multiplies, it stops being scarce. Judged by long-form logic, it’s still hard for it to pull audiences into emotional resonance. So for AI films, screenwriters remain very important.

We almost never use AI to write film scripts, but it can be a small assistant in the planning phase — offering creative points for us to filter, helping in brainstorming. For appearance consistency, at this stage we use face capture and similar methods.

For animation, style consistency may matter even more. We once wanted to make an animation incorporating traditional culture; the texture of a lantern’s edge was a creative focus, but the model may never have trained on that texture, so it couldn’t help us. On real-human faces, if review and compliance issues can be resolved, I believe model iteration will solve consistency too.

**Huang Jingrui**: Sun?

**Sun Tingyan**: The earlier discussion was mostly technical. As Leimeng said, the 2.5 model is stronger than 2.0, and character consistency is much better. But in the horizontal mid-length drama we’re making, if you look closely you still find problems.

Here’s our experience: when scenes are complex, costumes numerous, and performing characters many, a single generation can still go wrong. Our method is to build more assets up front. For example, a drama with 200 scenes means organizing costumes down to each scene and each costume in wide/full/medium/close-up/extreme close-up and three-view drawings; special lighting and motion also get their own assets.

That’s a big investment before formal production. If requirements are very high, consistency problems across scenes still exist — camera movement and light can also create character drift.

Also, creators need method and technique. Model capability, workflow construction, and director thinking all matter. Don’t brute-force the model’s limits. If it can’t do a 60-second long take, use director’s thinking to find a controllable alternative that doesn’t hurt the narrative much. If it doesn’t support a certain shot, don’t insist.

Expression is deeply personal. For art films, extreme pursuit is fine. But we’re a content production company — beyond expression, we have to balance labor efficiency and productivity. I don’t fight model capability head-on.

I just came back from Beijing yesterday and talked with Wang Luodan. She’s making a film now and is very particular about one extremely precise shot; the model couldn’t do it, so she keeps pulling cards. But that’s a different stance and logic from ours as a content company.

If you’re an independent creator chasing perfection on a few minutes of footage, you can try everything, even PS cutouts. But from a content company’s angle, I wouldn’t allow the whole team to work that way — you still need balance. It depends on your identity and role.

**Huang Jingrui**: Thanks. Nanqiang?

**Nanqiang (Rylee)**: A year ago, talking about character consistency meant mostly whether the face stays the same. Now that AI tech is mature, beyond fixing appearance, wardrobe, and recurring props, we emphasize complete pre-production character design — personality, habitual gestures, verbal tics.

Real people have their own features in every smile and gesture. In long-form narrative, you have to watch whether a character’s personality stays consistent; if it shifts, you have to design at which node. Not “laughing this way in act one, then acting like a different person in act two.” Fixing only appearance is useless.

For appearance consistency, besides thorough pre-production and repeated trial, we do lots of stress tests before formal production. Some character settings are naturally prone to model-training-data influence — the output looks like a different person, or the eyes and temperament drift. When that happens, we change the setting from the start, repeat stress tests, and only enter production once we confirm the character stays consistent on screen.

Narrative consistency, I think AI can’t solve. Whether short film, short drama, mid-length, or feature: the most important thing is what narrative I give the audience, what story I tell, and at which stage the audience sees and understands what. Scripts from AI may always be text piles that don’t know what the audience wants to watch or what the director wants. People have to control that, returning to the story itself. Characters, scenes, props, and narrative settings all circle around the story.

**Chen Qizhi**: Nanqiang’s point extends character design from appearance to personality. Those inner things, large models or visual outputs truly can’t see.

Previously, character consistency mainly meant training LoRAs in 2023–2024. In 2024 I also built a complete workflow covering image compositing, face swapping, background replacement, and color grading — and within two months the model updated and the old workflow lost its meaning. So tech selection still matters.

On the basis of analyzing a character’s inner setting, we sediment the method into a reusable workflow, then adjust parameters per use to achieve consistency from the inside out. Roughly that.

**Huang Jingrui**: Third question turns to commercialization: how does IP adaptation satisfy platform standards, respect original fans’ expectations, and still express the IP’s creative identity? Start with Leimeng.

**Zhu Leimeng**: Management of AI content is getting stricter, with more emphasis on compliance. Creators who want their works legal and compliant — beyond images or video they produced themselves — also need authorization for any faces and voices used, to do compliant IP adaptation. Otherwise it’s likely just re-creation: it may earn traffic and attention, but it’s hard to bring direct revenue under film-and-TV project logic.

To adapt an IP, you have to find the characters and relationships, grasp the original’s core narrative and what audiences actually expect — you can’t randomly change things. There was a period when people were wildly mangling _Journey to the West_ and other classics, and that kind of practice got restricted.

AI develops fast, and multi-form IP development is a good thing. Before, adapting an IP into film was too heavy an investment. Now the same IP can simultaneously appear as anime, film, TV series, interactive traffic content, and re-creations — which helps build the IP.

The premise remains compliance, capturing audience expectations, and creators knowing their own positioning. If you want to make films, you need film knowledge; short dramas, short videos, or MVs each demand understanding of their own form. This has little to do with AI itself.

**Huang Jingrui**: Sun?

**Sun Tingyan**: We’ve been involved a lot here. We’ve been adapting IPs since last year, and right now we’re adapting a big one — the first review bounced it back. I read the review comments; the content was indeed too sensitive.

**Huang Jingrui**: What IP is it?

**Sun Tingyan**: _I Was a Taoist in Those Years_ (《我当道士那些年》) — a very big IP. The original involves plenty of metaphysics and ghost elements, so adaptation is a craft; you have to give some things up. This is no longer just platform standards — it’s higher-level institutional requirements.

From a review perspective, adapting a script requires compliant artistic handling, which touches on whether you follow the original. I’m a fan of the original myself; some changes I also hate, but sometimes there’s no way around it. What we can do is preserve the spiritual core of the original and still do the compliance work that must be done.

AI can enhance the work at the aesthetic level. This IP is special — it has high requirements for visual effect and art, and AI can play a role. But in the end it still depends on the creator’s grasp of the original’s spirit. Every reader understands the original and the IP differently; perfection is impossible. You just try your best and express as precisely as you can.

**Huang Jingrui**: Nanqiang?

**Nanqiang (Rylee)**: We’ve touched quite a few IP adaptations too. First, look at the purpose of the adaptation. The others mentioned compliance and reasonableness. We’ve also met traditional IPs whose original plots were very tense but involved ghosts, feudal superstition, or other sensitive elements that needed adjustment. While respecting the original authors, you have to understand why you’re changing it.

Some classic old IPs had many fans back then; now in the short-video era, people want faster content. Old animations that told stories slowly in 15-20 minute episodes may have lost traffic, and authors want to adapt to the market. Some become short mini-episodes or short dramas of a few minutes, some go on self-media, some condense the whole work into a micro-film.

Different adaptation methods depend on how far the original author allows you to go. Do you tear up the old story and rewrite it, use new tech to realize visual effects the old animation couldn’t do back then, or tell a new story? Length also makes a big difference.

**Chen Qizhi**: We’re involved in IP adaptation less. But with or without AI, the core is respecting the original’s spirit. I see three aspects: story expression, character stance, and the visual style that goes with them.

An artist or author may be good at one or two styles, while AI can generate all kinds. When you get a work, you shouldn’t immediately use its assets to generate material. First find, from all styles, the visual style that best fits the original’s temperament, then refine. That’s an approach both technical and creative.

In recent years many works have adapted _Three Kingdoms_, _Journey to the West_, and _Fengshen_. Some are pleasing; some just feel like special-effects stacking. The latter often fail to capture the character core. Take _Journey to the West_: everyone knows the four protagonists’ personas. If you keep their core personalities and repackage them as an internet-style, entertainment-mode office drama, that can be a fresh form of packaging and output. That’s my thought on IP adaptation and AI.

**Huang Jingrui**: An audience member in the back has asked a question. I’ll ask one final question, and then have the four of you answer audience questions. If AI imaging truly achieves efficiency gains and scale, what’s the bottleneck the industry chain upstream and downstream most needs to break through? Please speak from an industry perspective.

**Zhu Leimeng**: For our studio, the upstream source of projects is IP. AI now produces lots of content — the main question is which ones are quality IPs worth developing. We see many creators publishing their own films on short-video and long-video platforms; to a degree those are IPs, but you can’t judge only by whether the picture looks good.

We make longer content. When we see a good five-minute short, we first consider whether its worldview can expand and support further development.

Downstream is the production team. Whether the team has enough production capability also needs consideration. AI lets many independent creators make things they love; they’re inspired and I’m excited too. But I have to consider who the work is for and whether it can hold a long-form audience using long-form logic.

Many individual creators can work slowly and finely and produce great work. But film-and-TV projects have time frames — possibly three to five months from development to broadcast, with cost, schedule, and later promotion to control. I can’t let a creator keep going slowly forever. If they have something on today, the project has no capacity today. That doesn’t work.

So after talking with individual creators, we may still assemble a crew the traditional way — everyone does their own job, with a general director controlling the whole process. If someone suddenly has a problem, someone else can step in. That’s how you’re responsible to the project and get it out on time, in front of the audience. Those are the two things I consider upstream and downstream.

**Sun Tingyan**: Could you repeat the question?

**Huang Jingrui**: If AI imaging truly achieves gains in efficiency, what’s the bottleneck the whole upstream-downstream chain most needs to break? It can be model capability, or production standards.

**Sun Tingyan**: This is really a question of how productive forces and production relations balance. After batch production and efficiency gains, productive forces are greatly enriched. Ordinary labor isn’t scarce — but compute being too expensive is a bottleneck. On the other hand, people with good minds who can create may also be a bottleneck. In my view there are three cores: people who can participate in production, relatively cheap tokens, and better creators. You have to think about productive forces and production relations together.

**Huang Jingrui**: Nanqiang?

**Nanqiang (Rylee)**: The bottleneck is whether the parties can align on the work’s goal and quality control. What upstream and downstream want isn’t necessarily the same. Upstream may want better artistic expression and more commercialization; downstream creators may want to write their own stories rather than stories everyone approves of.

Output quality is also hard to align. Earlier we discussed consistency; content style and finished-film quality equally need strong directors and producers to control. What I want may not be what the audience wants; what downstream hands me may not be what I want. We’ve hit exactly this problem recently.

Many companies now say they make polished work. I’ve received plenty of sample reels and I’ll say it straight: this doesn’t even qualify as B-level or C-level — it can’t be called polished in our book. But the other side thinks that in certain markets it is polished. This shows quality standards and output requirements are very hard to align.

**Chen Qizhi**: The others spoke from the film-making angle about director thinking that controls the whole, and about aligning requirements to finer granularity before execution. That resolves some of the barriers in creation and execution.

From another angle, the industry chain’s fundamental bottleneck, I think, is compute itself and the algorithms that optimize compute usage. Base compute is often superstructure — you usually can’t control it. But on top of it, we can optimize algorithms. More quality-and-efficiency-preserving algorithm optimization raises output under the same spec consumption.

**Huang Jingrui**: Thanks to the four. Let’s look at the audience questions on the big screen. First: how do you define a good work? If a work is released and makes little splash, how should a creator review and optimize in mindset and capability?

**Sun Tingyan**: How to define a good work? If you think it’s good, then it’s good. If box office fails after release, box office has an element of luck. Works that are both acclaimed and commercially successful have never been common. Creators should adjust their mindset. Now AI can mass-produce content and making things is easier — early vertical short-drama platforms had tens of thousands of shows a month, and people had to gradually accept that environment. If your mindset is off, first make some short videos and play with them yourself.

**Huang Jingrui**: When reviewing an applicant’s work, what ability matters most to you? How big a share is aesthetics?

**Zhu Leimeng**: First I look at the position I’m hiring for. If it’s an AI director, aesthetics may account for more than 70%, and narrative ability is also very important. You don’t have to know how to write a script, but you must know where a character’s performance goes and where a scene’s tension is — or if I explain it, you have to understand. We’re indeed hiring AI directors and producers right now.

A mature AI crew needs roughly seven roles, including director, art, DIT, producer, and generator. Some call generators “card pullers.” Different roles have different requirements. A generator doesn’t need full narrative understanding but must understand basic cinematography language and logic, otherwise the waste-card rate rises and cost control breaks down.

Aesthetics is a foundational ability. Judged by film standards, you must have aesthetics; in other content forms, the ability requirements of each role differ.

**Sun Tingyan**: Two more sentences. If a person’s aesthetics are especially good, I’ll value them highly in interviews — possibly more than 90%. I also scroll Douyin at night; when I see a video with an especially good aesthetic style, I go to the backend to find the creator. We badly need people with good taste. Even if the person is only strong on aesthetics, everything else can be handed to the rest of the team.

**Huang Jingrui**: Qizhi, I looked at the material you sent me yesterday and my aesthetics improved. What’s your view? How big is aesthetics’ share?

**Chen Qizhi**: Whether we receive a résumé or send our own work out, the first thing a stranger sees shapes their judgment of you. A work with good aesthetics makes people remember you instantly.

How that perception, that sense, forms is different for everyone — it comes from accumulated work history, the volume of films you’ve watched, and so on, and finally becomes your unique judgment. If the other side happens to need that sense — feels it fits their product or industrial pipeline — there may be room for further cooperation.

I think of the producer of _Death Stranding_ — his works have a very distinct authorship. He once used the word “meme.” If genes are decided by innate DNA, memes are accumulated from lived experience. Many AI directors weren’t originally in the film industry; with AI as a tool to realize ideas, their past accumulation is magnified at this moment. That’s my view on what makes something good.

---

Original publication: https://uniqueresearch.substack.com/p/one-storyboard-1000-frames-when-technology
On-site reading page: https://ffcap.cn/en/research/one-storyboard-1000-frames-when-technology