A film written as one continuous shot and stitched together on the last day

Eight scenes, each written as a single take, joined by physical occlusions such as coffee, water and a man's back. On the final day, Seedance was used as a post tool to fill the joins frame by frame.

Original project

Source: Higgsfield project by @aist (workspace "AIST for Festival"). 3m 24s, 2206×946, delivered to the ai_film_festival. Prompts are public.

At Fukushima Daiichi, a veteran engineer believes the plant can withstand the disaster — until the catastrophe proves otherwise.

Coverage: complete. 985 submissions produced 371 unique prompts per folder, and 365 unique prompts across the whole project (6 appear in two folders). 89 are video and 282 are image. All of them were read, with repeated phrases collapsed.

What was not done. The film was not watched and no generation was downloaded. This analysis covers what was asked for, not what came out. When a problem is said to "reveal" a model failure, that is inferred from what the prompt corrects.

There is also an interactive explainer with a frame-by-frame view of the stitching.

1. In one line

A film conceived as a single continuous shot of 3m24s: 8 scenes of 9 to 28 s, each written as one take, joined to each other by physical occlusion (coffee, water, a back covering the lens). On the last day, Seedance was used as a post tool to fill those joins frame by frame.

Where the budget went (according to the project's published data):

82% of submissions are images (811 of 985), and only 174 are video. Images are generated many times because they are cheap. Each video costs about 46 times what a Seedream image costs, and it is reached with everything already settled.

Compared with ANERNEQ

2. The story according to the prompts

The dramatic arc is in two phrases: 「頼もしいな」 in Sc_02 and 「落ちました」 in Sc_07. It is exactly the logline.

The sum adds up. Using the last version of each scene gives about 201 s, against 203.6 s for the final cut. In Sc_04 the 4 s of the input video are subtracted. This hints that the film has little beyond these takes. It does not prove that those versions were used.

Links that only show up when the batches are combined:

• The employee owns the car. The GIRL character sheet says "ABOUT 24 YEARS OLD, A JUNIOR ENGINEER". The car interior says "TRACES OF THE OWNER — a neat, sweet, intelligent 24-year-old Japanese woman… She is not in the picture". It is a character told through objects: omamori charm, plush toy, angel keychain.
• The oarfish. It opens alive in Sc_01. In the basement it appears dead: in the asset, only its head, Regalecus glesne; in Sc_05, "水边死鱼…银色躯体和破损鳍" (dead fish at the water's edge, silver body and broken fin). The link between the two is an inference, and so is reading the fish as an earthquake omen from Japanese folklore. No prompt says either.
• Only two characters have their own design, and neither has a name. Everyone else is "WORKER".

3. The main finding: one shot, stitched in post

The joins are written into the prompts

The end of each scene is written for the next one, and closing any other way is almost always prohibited:

• Sc_03_b_01: "不停成钥匙静物，不抬回人物" (do not stop on a still life of keys, do not tilt back up to the character).
• Sc_04: "没有淡出、黑屏或剪辑" (no fade, black screen or cut).
• Sc_06: "结束在真实衣料全遮挡与持续运动中，不停步、不定格、不人工淡出，不提前显示控制室" (end with real fabric covering everything and still moving; no stopping, no freezing, no artificial fade, do not show the control room early).
• Sc_07: "START ON COMPLETE BLACK. The black screen is actually created by the ENGINEER'S dark back and shoulder".

The technique is to use something opaque and physical as a curtain: dark liquid, muddy water, a back, a vertical sweep. The step between shots stays inside the occlusion. Sc_01 says it directly: "海底到咖啡的空间变化只发生在完全遮蔽视线的深色流体内部" (the change of space happens only inside the dark fluid that completely blocks the view).

The 130 submissions of the last day

57% of video submissions (99 of 174) carry a <<<video_1>>> input: they do not generate a take, they edit one that already exists. Transition and Cleanup start at 06:45 on Sep 14 and the last prompt is at 18:31, less than 3 hours before delivery. This is the most useful fact about the film for anyone who wants to borrow the method: stitching takes more submissions than shooting, so time has to be reserved at the end. Here it was all done in one day, and it cannot be known whether that was a plan or the rush of the deadline.

Transition template: fill in with locked frames

The input video is built with shot A, a stretch of pure black, and shot B, and Seedance is asked to generate only the black stretch. 22 of the 24 Transition prompts number the frames. Reconstructed from one of them:

以<<<video_1>>>为唯一视频依据，只填补原视频{A}秒至{B}秒之间的黑色视频占位段。
黑色是待替换区域，不是剧情中的停电或转场。
成片保持原视频{N}秒总时长、24fps、{F}帧及原始画幅。

最高优先级：0.000≤t<{A}秒和{B}≤t<{N}秒为锁定的原始视频，直接保留原帧，
不重新生成、不重绘、不美化、不调色、不裁切、不变速。
以第一帧编号为1：第1—{a-1}帧和第{b+1}—{F}帧必须与输入逐帧完全一致，每帧保持原时间位置；
只生成第{a}—{b}帧。不得插入、删除、重复或移动锁定帧。
{B}秒必须恢复输入的第{b+1}帧，而不是相似的重建画面。

{A}—{B}秒的新动作：承接黑场前最后原始画面…
新片段的起点匹配第{a-1}帧的构图、…；终点自然接入第{b+1}帧的…，不把边界帧复制成停格。
整个{N}秒看起来是同一台摄影机…的一次实拍。
新增区间内不使用黑场、叠化、形变转场、跳切、瞬移，不增加额外剧情。
原视频已有声音保持原时间位置，不增加对白、音乐或镜头whoosh音效。

What each part does: it states that the black is a hole and not a blackout in the story; it gives the same limits in seconds and in frames; it demands a return to the actual input frame and not a similar one; and it forbids the model from covering the work with a fade.

Each correction reveals a way the model cheated:

Cleanup template: repair with a black mask

It is the same idea in space and not in time: an area of the frame (the gauge) is covered with black and reconstructed. Mature version:

局部修复 <<<video_1>>> 中被黑色遮罩覆盖的温度表。

<<<image_1>>> — 视频正确的第一帧。
<<<image_2>>> — 温度表外观、外壳、刻度及上部关闭分段的详细参考。
<<<image_3>>> — 黄色分段逐渐熄灭时，温度表在黑暗中的外观。

保留正确的第一帧，并将同一个温度表延续到后续画面。去除黑色遮罩…
图片仅用作参考，不插入静帧或转场。保留 <<<video_1>>> 原始的{N}秒时长、构图、镜头运动、人物及其动作、照明和声音。
只修改指定的温度表及其局部辉光。

The text gets shorter as the images do the work. The Cleanup video prompts measure, in order: 701 → 958 → 1,037 → 876 → 482 → 346 → 332 → 631 → 694 → 438 → 366 → 123 characters. The fight was against a model that filled the whole column with yellow, until it reached "准确地只有两格" (exactly two segments). They also found that it read the mask as an area to light up: "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" (the black mask is only the area to repair, it does not mean the whole masked area should be filled with glowing yellow).

The case was solved with a correct first frame + an appearance reference + a reference of the dark state, not with more description.

4. How a scene is written

There is no single template. No single skeleton is shared across scenes. Each scene has its own, with different time formats and a different verb for citing references (提供, 规定, 锁定, 控制, 确定). Within a scene, the skeleton holds across versions. The usual order is:

Format and one-take rule                写实灾难电影，21:9宽银幕，一个连续长镜头。没有剪辑…
References with a role verb             <<<uuid>>>规定控制室环境；<<<uuid>>>锁定工程师身份…
Global rules (quake, light, language)   所有人物只说日语…
Camera route in one sentence            主摄影路线只有一条：从咖啡中上升、俯拍杯子、仰向面部…
Action in prose, step by step           (no timecodes in Sc_02, Sc_03, Sc_05 and Sc_06)
Lens and focus, brief                   约50毫米高质量变形宽银幕镜头
Sound                                   自然现场同期声音…没有音乐或字幕

Sc_07 comes from a different hand or process: English, 12 thousand characters, uppercase headings, timecodes, dialogue in script format and typos that read as hand-typed ("engoineer", "insicator"). It was written on the last day and is the only scene prompt in this style.

Techniques that repeat

Measured across the full text, EN+ZH:

Some techniques cannot be counted with a pattern, but they appear in every scene:

• The cause before the effect. It does not describe "shakes": it describes what moves first. "地板先晃动，他的身体随支撑面移动…不是身体自行抽搐" (the floor shakes first and his body moves with the surface; it is not the body convulsing by itself).
• A legibility test. "震动在关闭声音观看时也必须清楚可见" (the shaking must be clearly visible even when watching with the sound off).
• Denial attached to affirmation. "这只是摄影机在原位逐渐低头看向沙地，不是起重机下降" (this is only the camera gradually lowering its head in place to look at the sand, not a crane descending).
• Geography through objects. "方向盘正后方始终是驾驶座" (directly behind the steering wheel is always the driver's seat) and "不镜像环境，不交换车头车尾，不生成第二辆红色汽车" (do not mirror the environment, do not swap the front and back of the car, do not generate a second red car).
• An object's state as a reference. Two photos of the same gauge and the clarification that they are "一个物体在不同时间的状态" (one object at different times), with "不生成第二只" (do not generate a second one).
• Limiting how a reference is used. "不作为必须逐像素复现的首帧" (not a first frame to be reproduced pixel by pixel); "草图中的线条…不进入生成画面" (the lines of the sketch do not enter the generated image).
• Keeping the model from "explaining". "不为解释接触而扩展构图" (do not widen the frame to explain the collision).

Performance

It is written with physical symptoms and with what the character does not do:

• "下颌紧绷，呼吸短促，表情克制，没有微笑" (jaw tight, breath short, expression restrained, no smile).
• "不用嘴型说谢谢" (thanks without mouthing "thank you").
• In Sc_04, three declared phases: tense, frozen on seeing the wave, and desperate struggle.
• In the poster, anatomy instead of adjectives: "NOT bulging, NOT wide-open… brows drawn together… lips slightly parted as if a breath has stopped halfway".

5. How it iterates

Going back to the simple version

In 3 of the 8 scene folders, the last prompt is the first one with a minimal change. Checked by diffing the raw text:

Something similar happens in the reactor: the version that "starts underwater" is abandoned and the plan goes back to v1, at 12 s. And in the Sc_04 shot the camera goes from optical zoom to a physical route and back to zoom.

This has to be read with care: it is not known which take went into the cut. What does show is that when the text grew fat, the work went back to the base and changed one thing.

Change the asset before the text

Sc_06 v3 and v4 have the same length (2,732) and only loc_cor_car_01 is swapped for loc_cor_car_02. Reactor v1 and v2 are the same (loc_r2 for loc_r1). The same happens in Sc_03_b_01. The same text is resubmitted with another image.

From general rule to concrete case

In Sc_03_b_02 there is a single problem, attacked three times: the background crowd runs in any direction. "Everyone goes right" is not enough, so the final version puts two concrete employees: one in the background layer and one in the foreground. It also limits the scope of the rule ("only in the initial run") so it does not spoil the rest of the shot.

In images: negatives as scars

87 of 282 image prompts carry an AVOID list, and each new error adds to it. The best example is the control room ceiling: the model merges panels and ceiling → heights in meters are requested → a huge hall comes out → "not a hall" → it merges again → "hall-like". Two versions in a row ask for opposite things.

Generate, edit and erase

Half of the image craft is this three-step flow:

1. Long prompt in English, by sections.
2. EDIT — change ONLY X + "Preserve everything else exactly: [long list]" (87 prompts, 31% of images). What the prompt cannot achieve is fixed here: the number of fish, a beacon turned into a valve (×10), a fish replaced by an oarfish.
3. Mark and erase: a colored zone is painted over the image and removal is requested. "удали фиолетовую зону, остально сохрани" (remove the purple zone, keep the rest). There are 16 very short prompts and 33 submissions. There are also hand-drawn outlines used as a mask: "hand-drawn GREEN OUTLINE".

Each edit can break something else. The engineer's uniform is transplanted onto the GIRL and her shirt is lost, so the next prompt starts "KEEP HER SHIRT".

Another clear pattern is the requests about poster text: a single quantity word is swept ("в рази більше", "трохи більше", "більше", meaning "many times more", "a bit more", "more"). In the end the text is removed and added as a separate step, on a clean image.

6. Assets and art direction

A color canon

It appears in all the long image prompts, in almost the same words:

It is a film of gray sky, flat light and a single red: the car, the valve, the beacons, the emblem. The parking lot prompt says it: "the red car stays the single strongest colour". The internal language speaks of "the plant's material canon" and "the project's grade".

Asset decisions

• Chimneys: they go from red and white to "WHITE, WITH A PALE BLUE HEAD… NO stripes, NO bands, NO red". The AVOID list prohibits what was requested before, and the poster corrects them by hand afterward.
• Brand: from the real TEPCO patch to an invented four-circle emblem, with "must NOT resemble any real-world brand". The keychain goes through something similar: "менее узнаваемый" (less recognizable). The legal motive is an inference. "TEPCO" is still written in 27 prompts.
• Engineer: he is 33 on the character sheet and "early 50s" on the poster. The logline says veteran.
• Gauge: from a dial with a needle to a segment bar, which allows states. The states are produced by editing and cropped into separate images.
• Car: a red Suzuki Wagon R MH23S, plate 福島 580 / あ 19-17, reused up to the tsunami scene.
• Reversed shots: the corridor reverse shot is requested as "REVERSE SHOT of Image 1", with the walls swapped and a negative that avoids returning to the original: "windows on the left, doors on the right, the same composition as Image 1 un-mirrored".

Research: a lot, but uneven

Well documented: the two-level ceiling of the room, the 3号機 (Unit 3) sign, the blue camouflage of the buildings, tetrapods, BWR vocabulary in the reactor: SCRAM, decay heat, core shroud; "NO Chernobyl-style graphite blocks"; "No blue Cherenkov glow".

What is missing: "generator" never appears. The diesel is only a spoken line in Sc_02 and is not seen in any asset.

The time of day is a license. 39 prompts fix "about 16:00" for any state: "Minutes after the earthquake, before the tsunami" (corridor) and "minutes after the tsunami" (basement) are the same hour. Outside the prompts, and from memory without verification, the real earthquake was at 14:46 and the tsunami arrived around 15:30. A full historical review was not done.

Models: tendencies, not rules

Seedream does the volume and most of the edits. GPT Image appears mostly for legible text, edits with two references and the helmet. There are minimal prompts ("A photo, 2011, …") tried on several models, and sometimes the same text is resubmitted to another model to compare. No distinct skeleton per model was found: what changes is the task.

7. Language and authorship

Reasonable inferences, none stated in the prompts:

• The short image commands are in Russian and Ukrainian and appear typed by hand ("удили", "остально", "логотимоп"). An "фронт" slipped into one English prompt.
• The long texts read like LLM output. What suggests it is the uniform Chinese next to short commands in Russian and Ukrainian, the exact frame arithmetic, and phrases such as "in the previous version", which read as written from a conversation.
• Why almost all video is in Chinese and Sc_07 in English cannot be known from the prompts: it is the same model, and the two styles coexist on the last day.
• More than one person: it is likely (British and American spelling, styles that change within the same day), but it cannot be settled. A single bilingual person also cannot be told apart from two people.

8. What is worth copying

1. Decide how the shots will join before writing them. Write how each take ends with the next one in mind, and use something opaque and physical as a curtain.
2. Save time for stitching. Editing existing footage took 57% of video submissions (99 of 174), all of it on the last day.
3. Use the fill templates from section 3. Build A + black + B, number the frames, require the real return frame and forbid the fade.
4. Ask for the physical event, not the result. "Film the crossing of the surface", not "make the cup appear". The model delivers the result by cheating.
5. If the text grows without improving, go back to the base and change a single thing. Try another asset first.
6. Make the rule that is not followed concrete: two employees with positions instead of "everyone goes right".
7. Resolve the world in images, where it is cheap. Here images were 92% of generations and 28% of credits.
8. Do not fight the generation over a detail: move it to an EDIT — change ONLY X, or paint it and erase it.
9. Fix a saturation hierarchy and write it the same way in every prompt.

9. What remains open

• Which takes went into the cut. The public data does not say. Confirming it requires watching the film.
• Whether the film really looks like a single shot. The prompts ask for it; the sum of durations supports it.
• The gauge. Sc_07 asks for two red segments over the yellow ones, and Cleanup prohibits red and asks for two yellow ones. It is not known whether they are the same shot.
• Where the reactor belongs. It sits in the Sc_07 folder and was placed last by content, not because anything indicates it.
• The valve. The vent folder designs a conical horizontal handwheel, and a different one appears in the basement.