Hidden Sichuan: 68 cuts

Context

A delivery restaurant has nowhere to put its brand. There’s no sign, no dining room, no staff to remember. The whole impression is a name and a photo in an app. For Hidden Sichuan, I wanted more than that: a story world people could follow before they ever ordered. For a restaurant nobody walks into, the story is the storefront.

The usual way to get an animated series is to hire a studio, and for a business that hadn’t opened yet, that wasn’t realistic. So I made it myself with AI image and video tools.

The problem

The tools turned out to be the easy part. The hard part was getting them to produce the same characters, the same kitchen and the same food, consistently, across dozens of shots.

What I did

Building the world first

Before generating anything, I wrote the world: who the characters are, where the story happens, and who they’re up against. That included a rival brand, Double Dragon, written as Hidden Sichuan’s opposite.

A character bible someone else could use

I made a 36-sheet character bible an outside artist could work from, with every character shown from several angles. I designed each character’s body as a record of their working life rather than a default “attractive” shape, and gave each group in the story its own color system.

Managing only what repeats

My first list of visual assets had 104 items. I cut it to 87, then to 51. The rule: manage only what appears again and again. Anything that shows up once gets described directly when it’s needed.

First list104 items

Second pass87

Final51

Fig. 1 The visual asset list, cut down to only what repeats.

68 cuts on a beat

The first episode is 2 minutes and 16 seconds long: 68 cuts, every one exactly two seconds, edited to a 120 beats-per-minute track. At its center is a 60-second recipe sequence in which each of the six cooking stages adds a new layer to the music, from the kick drum to the melody, with the first kick timed to the moment the flame ignites.

The 60-second recipe sequence, 30 cuts
The rest of the episode

Fig. 2 Episode 1, cut by cut. Every cut is exactly two seconds, so at 120 beats per minute a new cut lands every four beats.

Separately, I generated a full cooking sequence of 86 shots, first as still images and then as video clips.

Doing the research

Content only feels real if it’s accurate. I wrote 39 pages of research notes on mala: the six layers of flavor in the sauce, where each ingredient comes from, the Chinese national standards for key ingredients, the science of fermentation, and food safety. It ends with a list of 20 claims that sound true but shouldn’t be stated as fact.

Making the process cheaper

A failed video costs far more to redo than a still image. So I locked the key frames as images first, and only then generated video from them. I also built a simple tool for locking the visual style: about 75 references across nine decisions, with a checker that flagged problem words in prompts.

writingWrite the world and the character bible
stillsLock the key frames as still imagescheap to redo
videoGenerate video from the locked framesexpensive to redo

Fig. 3 The order of work, set by what each step costs to redo.

What I learned

Two minutes of story is a lot of decisions. Sixty-eight cuts means sixty-eight choices about what to show for two seconds, and each one has to agree with the character bible, the color system, the music and the food research. Doing it myself is how I learned how labor-intensive creative production really is.

AI tools reward precision more than imagination. When I described a pile of food as “crown-shaped,” the model drew an actual crown in the pan. It happened with wings, ribbons and domes too. Once I banned comparisons and described shape and physics directly, wider at the base and narrowing toward the top, the results became consistent. It wasn’t a creative problem. It was a specification problem.

And cost structure decides workflow. Knowing what was cheap to redo and what was expensive changed the order I did everything in.