The Fukushima Prompts

Prompt, submission, generation

These three words sound alike but mean three different things. Unless you keep them apart, the numbers in the analysis won't add up.

Send the same text five times and that's one prompt and five submissions. Each submission can return several results, which is why there are more generations than submissions.

The 3 missing generations are object swaps, too few to matter. You'll also see “365 unique”: 371 is the count per folder, and 6 texts appear in two folders.

A single shot

A shot is what the camera records without a cut. The central idea of this film is to look like one take from start to finish.

You can't generate 3 minutes in one go, because a Seedance take runs 30 seconds at most. So they wrote each scene as one take, and wrote its ending with the next scene in mind. The trick is to block the camera with something physical: coffee, murky water, a person's back. While the image is covered, you can jump somewhere else without anyone noticing.

1. 1
   The seabed. An oarfish flees when the quake hits, and the camera rises through the mud.Join: it comes out of the inside of a coffee cup.
   9 s
2. 2
   Control room. The engineer orders the diesel started. The gauge reaches 800 °C, the power cuts out, the reading drops, the lights return. He pats the panel: “So reliable.” Someone shouts “Tsunami!”Join: a fast turn toward whoever shouted.
   24 s
3. 3
   He runs with the others. He sees the sea through the window, and we read his fear in his reflection. He puts on his helmet. Something hits him and a set of keys falls.Join: a whip down, with the image blurred.
   25 s
4. 4
   A young employee runs against the flow. “You're safer here.” “I can't reach my family.” He nods and lets her go.Join: she steps outside and the light blows out the image.
   27 s
5. 5
   She reaches her red car. He watches from the second floor. The tsunami flips the car and she drowns, trapped by the seatbelt.Join: murky water fills the whole frame.
   37 s
6. 6
   The camera breaks the surface in the basement. He fights a stuck valve, falls into the water, turns on a flashlight and escapes.Join: a door and a turn around a corner.
   22 s
7. 7
   Through the window he sees the red car crushed against a building, also reflected in his glasses. He forces himself to keep going.Join: his wet back covers the entire lens.
   17 s
8. 8
   Opens on black: it's that same back. Control room, total blackout. “And the emergency power?” “It's down.” His confidence breaks.Cut to black. The only declared cut.
   28 s
9. 9
   Reactor. The water level falls and the fuel is exposed. No explosion. I placed it last because of its content: nothing in the data gives its position.
   12 s

The math nearly works out: the latest versions add up to about 201 s, and the final cut runs 203.6 s. That suggests there isn't much else in the film. It doesn't prove these versions were the ones used.

Stitching shots with Seedance

This is the most unusual technique and the most reusable one. Seedance normally generates video from text. Here it's also a post-production tool, used to fill the gap between two takes that don't match.

1. 1
   You have the end of take A and the start of take B.
2. 2
   You build one video: A, a stretch of black, and B. The prompts confirm the black, but don't say what software assembled it.
3. 3
   You give it to Seedance as an input video. In the prompt it appears as <<<video_1>>>.
4. 4
   You ask it to generate only the black part and keep everything else identical, frame by frame.

The video runs at 24 frames per second, which is why the prompts talk in frames. Try it with real examples:

Why so precise

The prompt gives the same limits in seconds and in frames, and adds one key rule:

“At 3 seconds, frame 73 of the input must return, not a similar reconstruction.”3.000秒必须恢复输入的第73帧，而不是相似的重建画面。

It also makes clear that the black is a gap, because this film has real blackouts too:

“Black is the area to replace. It is not a blackout in the story or a transition.”黑色是待替换区域，不是剧情中的停电或转场。

The model's tricks

When the model can't solve something, it fakes it. Each prompt version corrects one way of faking. What the model was doing is my deduction from those corrections.

The black mask

This is the same idea applied to an area of the frame instead of a stretch of time. If part of the image goes wrong, you cover it with black and ask for just that part to be rebuilt.

They used it on a temperature gauge the model couldn't get right. It's the most repeated operation in the cleanup folder.

Schematic drawing. The prompts never fix how many segments the gauge has.

The text kept shrinking

The early versions described everything in words, in more and more detail. The solution was to supply three images: the correct first frame, what the gauge looks like, and how it looks in the dark. With those, the last prompt is 123 characters, against 1,037 for the longest.

The model also misread the mask, and they had to spell it out: “The black mask is only the area to repair. It doesn't mean the area should be filled with yellow.”黑色遮罩只是需要修复的区域，不代表要把整个遮罩区域填成发光的黄色。

Back to simple

When a result doesn't work, the instinct is to write more. That happened in three scenes, and the last prompt ended up being the first one with a minimal change. I checked this by comparing the full texts.

Red marks the last version in each folder. I don't know if it made it into the film, but it shows a habit: when the text grows without improving, go back to the base and change one thing.

A related habit: change the reference image before changing the text. Sc_06 has two versions with text of identical length where only the corridor reference changes. The reactor does the same.

The image work

Before generating video, they solved everything in images: characters, locations, objects. Those images then become the video's references. The process has three steps.

1. 1
   Long prompt. Describes the place or object, in English and in sections: light, color, camera, and a list of things to avoid.
2. 2
   Targeted edit. EDIT — change ONLY X, plus a long list of what must not change. Whatever the long prompt can't achieve gets fixed here. For example, they couldn't get the right number of fish in the basement, so they ended up asking to “remove half”.
3. 3
   Mark and erase. They paint a colored area on the image by hand and send a short order: “удали фиолетовую зону, остально сохрани” (“erase the violet area, keep the rest”). There are 16 of these very short prompts.

Every edit can break something else. They gave the young employee the engineer's uniform and her shirt disappeared. The next prompt had to open with “KEEP HER SHIRT”.

AVOID lists work like scars. Each new mistake joins the list of things to avoid and stays there for good. Read one of those lists and you can see everything that went wrong before.

A single red. They set down in writing what the most saturated thing in the frame is, with the phrase “Saturation hierarchy” (35 prompts). The film has grey skies and flat light, and red is saved for the car, the valve and the beacons. That's why your eye goes to the car.

Where the credits went

Video costs far more to generate: one Seedance generation costs about 46 times as much as one from Seedream.

That's why they fight their problems in images, where mistakes are cheap, and reach video with everything solved. It's also why the text of a video prompt matters so much.

The last day

Of the 174 video submissions, 130 are from September 14, and the last prompt came less than 3 hours before the festival submission. More than half of all video submissions (99) don't generate a new take at all: they edit one that already exists.

Why so many languages

The spoken dialogue is in Japanese, written as is inside the prompt. So is the text you read on screen, like the license plate 福島 580.

What I infer, with no confirmation: the team speaks Russian and Ukrainian, because those short orders have the typos of someone in a hurry. And the long Chinese texts probably went through an assistant like ChatGPT or Claude, because they're very uniform and free of errors. There's no way to tell why they chose Chinese.

What we know and what we infer

Higgsfield doesn't report which take made it into the film. To find out, you'd have to watch it and compare.

What you can copy

1. Decide how the shots connect before you write them. Write the end of each take with the next one in mind.
2. Save time for stitching. Editing existing footage took 57% of video submissions, all of it on the last day.
3. Ask for the physical event. “Film the camera crossing the surface” works better than “make the cup appear”.
4. If a rule isn't followed, make it concrete. “Everyone runs right” failed, so they replaced it with two employees, each with a position.
5. If the text grows without improving, go back to the base. Try a different reference first.
6. Solve details in images. It's much cheaper, and a targeted edit is often enough.

Glossary

Shot or take

What the camera records without a cut.

Join

The point where one take ends and another begins.

Occlusion

Something that briefly blocks the camera. Here it acts as a curtain that hides the join.

Frame

Each still image of the video. There are 24 per second here.

Locked frames

The ones the model must leave identical.

Mask

An area covered on purpose to show what has to be rebuilt.

Inpainting

Rebuilding a covered area of an image or video.

Asset

A finished piece saved for reuse: a character, a location or an object.

Reference

An image or video you give the model to base its work on.

<<<video_1>>>

How a prompt cites an attached file. <<<image_1>>> works the same way.

uuid

The long code that identifies an asset, like 152846bb-dbbe-…. It's the asset's internal name.

Negative prompt or AVOID

A list of what you don't want to appear. In this film the video prompts don't use one: the prohibitions go inside each sentence.

Re-roll

Generating again with the same text, hoping for a better result.

Pan and tilt

Turning the camera sideways, or up and down.

Saturation

How intense a color is. A saturated red grabs attention; a muted one barely does.

LLM

A text model, like ChatGPT or Claude.