Моя серия · Seedance 2.5

Один кадр из вагона — пять миров за окном

Камера прикручена внутри вагона и не двигается все 30 секунд. За окном поезд на скорости проходит пять миров, между ними — четыре подъёма в облака. Герой стоит у стекла мелко, справа, в профиль. Ниже — ровно те промты, картинки и настройки, которыми сделан ролик сверху, включая промт на 27 тысяч знаков, который и делает всю работу.

Модели: GPT Image 2 + Seedance 2.5 + Topaz (всё в Higgsfield)Референсы: 7 картинокСложность: высокаяМой результат: в видео ↓
Мой результат. Один непрерывный кадр на 30 секунд: летний луг с низким солнцем → розово-кварцевая долина → изумрудные холмы над облачным морем → пещера с фонарями → сакура-остров над облаками. Сгенерировано в Seedance 2.5, поднято до 4K в Topaz.

01Что нужно

02Шаг 1 — карта персонажа

Карта персонажа — то, по чему модель держит лицо и одежду все 30 секунд. Делается в GPT Image 2 из двух-трёх своих фото при ровном свете: лицо крупно и фигура в рост. Промт — четыре ракурса на сером фоне:

Промт · карта персонажа, 4 ракурсаGPT Image 2 · Higgsfield · 16:9 · 4K · + 2 своих фото
Izobrazheniye odnogo i togo zhe cheloveka. Litso dolzhno byt' gipperrealistichnym bez platikogo litsa. List s opisaniyem personazha — chetyre rakursa na neytral'nom serom fone: [RAKURS 1 — POLNYY RAZMER, SPEREDI] Polnyy vid personazha speredi pod uglom tri chetverti, vidna vsya figura ot golovy do nog. [RAKURS 2 — POLNYY RAZMER, SZADI] Polnyy vid personazha szadi, pryamo szadi. Vidna vsya figura ot golovy do nog. [RAKURS 3 — KRUPNYY PLAN SPEREDI] Krupnyy plan golovy i plech, pryamoy vid speredi. Chetkaya detalizatsiya tekstury kozhi, aksessuarov i detaley kostyuma. V nizhney chasti kadra vidny nagrudnaya i naplechnaya bronya/odezhda. [RAKURS 4 — KRUPNYY PLAN PROFILYA] Krupnyy plan golovy i plech, vid sleva pod uglom 90 gradusov. Vidny sheya i verkhnyaya chast' plech. Osveshcheniye i prezentatsiya: Chistoye studiynoye osveshcheniye — myagkiy osnovnoy svet v levom verkhnem uglu, myagkoye zapolnyayushcheye osveshcheniye sprava. Yedinaya identichnost' personazha, proportsii i detali kostyuma na vsekh chetyrekh rakursakh. Bez teksta, vodyanykh znakov, dopolnitel'nykh figur, fona. V stile, predstavlennom nizhe... Kinematograficheskiy fotorealizm, kontseptual'noye izobrazheniye personazhey, rezkaya fokusirovka po vsem rakursam, malaya glubina rezkosti tol'ko na fone.

An image of the same person. The face must be hyper-realistic without the appearance of plastic. Character Description Sheet - Four angles on a neutral grey background: [ANGLE 1 - FULL SIZE FRONT] A full front view of the character at a three-quarter angle, showing the entire figure from head to toe. [ANGLE 2 - FULL SIZE REAR] A full back view of the character, from directly behind. Showing the entire figure from head to toe. [ANGLE 3 - CLOSE SHOT FRONT] A close-up of the head and shoulders, viewed from the front. Crisp detail in skin texture, accessories, and costume details. Chest and shoulder armor/clothing are visible at the bottom of the frame. [ANGLE 4 - CLOSE SHOT PROFILE] A close-up of the head and shoulders, viewed from the left at a 90 degree angle. Showing the neck and upper shoulders. Lighting and Presentation: Clean studio lighting—soft key light in the upper left corner, soft fill light on the right. Consistent character identity, proportions, and costume details across all four angles. No text, watermarks, additional figures, or background. In the style presented below... Cinematic photorealism, conceptual character art, sharp focus across all angles, shallow depth of field only in the background.

Промт написан по-русски латиницей и продублирован по-английски — модель одинаково хорошо читает оба. Главное в нём — четыре ракурса: фигура спереди и сзади, лицо крупно и профиль. Крупное лицо на карте — то, что потом держит идентичность в видео, а для этого ролика особенно важен профиль: герой все 30 секунд стоит к камере боком.

Одежду модель с первого раза придумывает сама — мне она выдала чёрную броню. Меняется вторым шагом, тем же GPT Image 2 в режиме редактирования: карта первым референсом, фото с нужным аутфитом вторым, и один промт:

Промт · замена одежды на картеGPT Image 2 · эдит · референсы: карта + фото аутфита
поменяй одежду на первом фото как на втором

Так на карте появился кремовый свитер с воротником, светлые брюки и белые кроссовки — в этой одежде герой и едет в поезде. Наушников на карте нет: их дорисовывает промт видео, там прямо сказано, что проводные наушники в референсе отсутствуют и их надо добавить. Так можно докладывать мелкие детали, не перегенерируя карту.

после первого промта: 4 ракурса и лицо на месте, одежду модель выбрала сама
@Image 2 · после эдита одежды — эта карта пошла в видео

Почему GPT Image 2, а не Nano Banana

Для карты персонажа важно одно — лицо. На моих прогонах Nano Banana делает лица глянцевыми, «пластиковыми»: сглаженная кожа без пор, стеклянный блеск, и эта пластмасса потом переезжает в видео — Seedance честно копирует то, что видит на карте. GPT Image 2 держит фактуру: поры, лёгкую асимметрию, живые глаза — лицо выглядит как фотография, а не как рендер. Nano Banana хорош для быстрых правок и сцен, но карту героя я делаю только в GPT Image 2.

03Шаг 2 — интерьер вагона

Вагон — первый референс, и он важнее, чем кажется: по нему модель берёт форму окна, резиновый уплотнитель, металлический кант, полку под стеклом, цвет стен и базовый свет. Окно на картинке — чистый белый прямоугольник, заглушка: в видео модель заполнит его мирами. Два решения в промте: камера стоит чуть левее окна под 45°, а справа у поручня оставлено пустое место. Туда встанет герой.

Промт · интерьер вагонаGPT Image 2 · Higgsfield · 16:9 · 1k · high · без референсов
Photorealistic interior of a modern Chinese high-speed train carriage (CR400 Fuxing style), vertical 9:16. Sleek premium design: glossy white and light-grey curved panels, seamless joins, soft warm-white LED strip lighting along the ceiling, clean light floor, a slim brushed-metal vertical handrail pole. The camera stands slightly to the LEFT of a large panoramic window and looks at it from about a 45-degree angle, so the window recedes in gentle perspective across the left and centre of the frame, and on the RIGHT side, directly in front of the glass, there is clear EMPTY standing space by the handrail — room reserved for a person, but the carriage is completely empty, no people. The window glass is a BLANK PURE WHITE VOID — no scenery, no reflections, just seamless white, as a placeholder. Around the glass: a dark rubber window seal, slim clean metal trim, a narrow ledge below. Everything new and well-kept, only the faintest signs of real use — a few micro-scratches on the ledge, faint smudges on the glass edge. Neutral soft interior lighting, real materials: glossy composite panels, brushed metal, glass. Documentary photographic realism, fine natural grain, no CG gloss over-shine, no text, no logos, no people.
первая попытка: окно во весь кадр, герою негде стоять — в ролик не пошла
@Image 1 · пошёл в видео
Скачать ↓

Первый вариант я делал «по учебнику» — окно на 85% кадра, как требует видео-промт. Но на плоской картинке это не оставляет места человеку, и модель потом сажала его поверх стекла. Ракурс под углом решил вопрос: окно читается целиком, а справа есть куда встать.

04Шаг 3 — пять миров

Каждый мир — отдельная картинка. В промте у каждой свой абзац с описанием и свои секунды, а главное правило одно: картинка мира — это то, что видно в окне, а не весь кадр. Порядок слотов совпадает с порядком в промте.

@Image 3 · мир 1, летний луг
Скачать ↓
@Image 4 · мир 2, розово-кварцевая долина
Скачать ↓
@Image 5 · мир 3, холмы в облачном море
Скачать ↓
@Image 6 · мир 4, пещера с фонарями
Скачать ↓
@Image 7 · мир 5, сакура-остров
Скачать ↓

Скачать все референсы одним архивом ↓ 6 файлов: вагон + 5 миров

Четыре мира фотореалистичные, пещера с фонарями — рисованная. Промт требует фотореализм, и модель её «пересняла», но в кадре она всё равно самая мультяшная из пяти. Хочешь ровный стиль — подбери пещере фотореалистичный аналог.

05Шаг 4 — видео в Seedance 2.5

В Higgsfield: Cinema Studio → Videos → Seedance 2.5. Референсы строго по слотам — слот в интерфейсе и есть номер в промте:

  1. интерьер вагона;
  2. карта героя;
  3. с третьего по седьмой — миры в порядке галереи выше.

Параметры ролика сверху: 9:16 · 30 секунд · 480p · звук включён · bitrate high · seed не фиксировать. 480p — сознательно: проверять, что миры встают по секундам, дешевле в низком разрешении, а финал потом поднимается апскейлом.

Промт · видео, копипастомSeedance 2.5 · Higgsfield · 30 сек · 9:16 · 480p · звук
**SCENE CONTEXT**
Single continuous 30-second shot. Vertical 9:16. Photoreal live-action. The camera is bolted inside a moving train carriage and never moves, never pans, never zooms, never cuts. FIVE worlds, connected by FOUR cloud passages. No stations, no tunnels bored through rock, no space.
**FRAMING — THE WINDOW IS ALMOST THE WHOLE FRAME**
The camera is turned nearly square-on to the glass so the window is seen head-on and its full rectangle is visible edge to edge. The window occupies roughly 85% of the total frame — from the top edge down to the bottom sixth, filling the entire left and centre. The young man is SMALL and pushed to the far right edge, occupying only the right 15% of the frame width, a narrow vertical slice, seen in profile and cropped by the frame edge so we see him from the shoulder forward. Visible interior is minimal: the rubber seal, the metal trim, the ledge, a sliver of ceiling panel at the very top. If he takes up more than a sixth of the frame, the framing is wrong. Every landscape must be readable across its full width at all times.
**ACTIVE REFERENCES**
<<<image_1>>> — TRAIN INTERIOR. Controls the window's shape, rubber seal, metal trim, ledge, wall colour and wear, ceiling panel, base interior lighting. The blank white void in this reference's window is a placeholder and must be FILLED with the worlds below. Does not control camera position — we sit closer and far more square-on to the glass than this reference.
<<<image_2>>> — THE YOUNG MAN. Controls face, facial structure, skin texture, eyes, brows, lips, hair, and wardrobe: a cream ribbed-knit half-zip sweater with a stand collar over a white t-shirt, cream wide-leg trousers, white low-top sneakers. 100% match, zero drift. He additionally wears white wired earbuds in both ears with a thin cable running down — the earbuds are NOT in the reference sheet, add them naturally. Does not control pose, lighting or background — the grey studio backdrop must never appear.
<<<image_4>>> — WORLD TWO, the rose-quartz valley: towering translucent pink crystal spires rising from green meadow slopes, a winding pink glass-smooth river, banks crowded with large pale-pink blossoms, soft green mountain ridges behind, silver-grey daylight sky. Window content only, 6.5–11.5s.
<<<image_5>>> — WORLD THREE, the cloud-sea hills: colossal round emerald hills breaking through a rolling white fog ocean, small white houses with dark roofs on each summit, a narrow grey road winding up the slope with a few tiny cars, streetlamps, deep blue sky above the mist. Window content only, 13–17.5s.
<<<image_6>>> — WORLD FOUR, the lantern cavern: an immense open cavern world of dark violet-and-blue rock arches, glowing turquoise-white waterfalls of light pouring off the stone, floating orange paper lanterns everywhere, a wooden market street of stalls and jetties running along a luminous turquoise river speckled with glowing shallows. Window content only, 19–23.5s.
<<<image_7>>> — WORLD FIVE, the sakura island over the cloud sea: a huge floating rock island carried above an endless cream cloud ocean, crowned by one enormous weeping cherry tree in full pink bloom, cascades of blossom trailing off every branch and off the island's mossy pink-draped cliffs, a small dark-roofed wooden shrine with a red torii gate under the canopy, drifting petals, soft blue sky with scattered cumulus. Window content only, 25–30s.
**CRITICAL REFERENCE RULE.** Every landscape reference is a WINDOW PLATE. None ever becomes the full frame, none dictates the camera, none replaces the interior. We stay inside this carriage, at this exact angle, for all 30 seconds.
**SPEED AND DIRECTION — LOCKED**
The train travels forward in one single direction for all 30 seconds and it is FAST. It never turns, never banks, never reverses. Static scenery always drifts across the glass in ONE unchanging direction: entering at the window's forward edge (frame left) and exiting at its rear edge (frame right). Never the opposite.
ABSOLUTELY NO SLOW MOTION. Every frame runs at real-time speed. We pass these worlds — we do not float through them. Foreground detail streaks with genuine motion blur. Wind moves at real wind speed. Water falls at real speed. Petals and lanterns drift at real physical speed. If any element looks dreamy, floaty or retimed, it is wrong.
Every transition is the same manoeuvre: the nose pitches up to roughly 45° and the train flies FORWARD and UPWARD along its own axis into a cloud bank, gaining altitude while still travelling ahead. Never a vertical lift, never hovering. This reads through the world outside and through the slant of the light rectangle on the interior wall — never through a tilted camera. The camera horizon stays level. After each cloud the train levels back to horizontal.
**THE SUN — MANDATORY, TREATED AS A CHARACTER**
The sun is a real physical light source in every open-sky world and must be visibly working in the frame.
- It is low, hard and bright, sitting within or just outside the window's field so its light rakes directly through the glass into the carriage.
- Because the train moves forward and the sun is fixed, its position shifts across the window in real time and its beam sweeps continuously through the interior — a hard bright rectangle travelling across the ledge, up the wall, over his face and hair, then moving on. This sweep never stops in the daylight worlds.
- Every time the sun passes behind something outside — a crystal spire, a hill summit, a cloud tower, a blossom canopy, the floating island itself — the whole interior darkens for a beat then blows bright as it clears. This flicker of light and shadow is one of the main signs of speed.
- ONE EXCEPTION: inside World Four, the lantern cavern, there is no sun. There, the interior is lit only by the passing lanterns and the glowing falls — see LIGHTING. The sun's return at the mouth of the cavern is a lighting event in itself.
- Real optical behaviour on the glass: the sun blooms across the dirt and scratches on the window, veiling glare washing across the pane at certain angles, faint chromatic fringing at the frame edges. No streak flares, no star flares, no lens-flare overlays.
- The sun glints hard off water: sharp moving specular sparkles on the pink river, a bright shifting sheen on the fog sea and cloud-sea tops, a burning rim along the edge of every cloud it stands behind.
- Sun shafts break through gaps in the clouds and through the cherry canopy as long visible beams that sweep past our window as we travel.
- His face is lit almost entirely by this light through the glass — hard warm key on the window side, deep soft shadow on the carriage side, bright rim along his cheek, nose bridge and the edge of his hair. His eyes narrow slightly when the beam crosses them and relax when it passes.
**THE WIND — MANDATORY, STRONG AND VISIBLE**
Strong wind, not a breeze, and every organic surface in every world reacts on its own layer and its own timing.
- Grass is laid over in travelling gusts: waves race across whole meadow slopes and hill faces, flattening blades in a moving band that crosses the frame in under a second, springing back behind.
- Pink flowers along the crystal river whip and nod violently on thin stems, petals torn loose and streaking past the glass.
- The cherry tree bends as a whole mass — trunk flexing, the enormous blossom canopy shoved sideways and shivering, hanging blossom-cascades whipping and lagging hard, trailing, never still. Blossoms are torn loose continuously and driven downwind at real speed — never floating.
- Loose material rides the air: petals, seed fluff, pollen, dust and torn grass streaking past close to the glass.
- Fog and cloud are sheared by the wind — edges tearing and re-forming, thin veils pulled off the hill summits and off the cloud-sea tops, whole shapes deforming as they move.
- Inside the cavern the wind becomes a river-draft: lanterns swing and drift on it with real pendulum lag, stall awnings ripple, hanging vines and market cloths stir, the glowing falls' spray is blown sideways in drifting sheets.
- Chimney and stall smoke is bent almost horizontal.
- Wind noise bleeds through the window seal — a constant pressure hiss rising and falling with the gusts.
**FIRST FRAME AND SPATIAL BLOCKING**
[FRAME] Vertical. Window seen almost square-on, filling the frame from top edge to lower sixth, across the entire left and centre. The young man at the far right, x≈90%, small, cropped by the frame edge, in profile.
[ORIENTATION] Turned toward the window, back and right shoulder toward the lens, roughly 25% of his face readable in profile, lit entirely by the sun through the glass.
[HEAD] Level, chin neutral, eyes DOWN on the phone in his left hand. Never at the camera at any point.
[ARMS] Right arm raised to the pole at shoulder height, hand wrapped around the metal, fingertips not meeting. Left arm bent at the waist, phone at chest level, thumb resting on it. Earbud cable in a soft slack loop.
[TORSO] Upright, weight settled slightly on the right leg.
[LEGS] Feet shoulder-width, braced against the carriage's motion, right foot half a step forward, knees soft.
[BANS] No eye contact with the lens. No head rotation past 90°. No arms above the head. No other passengers, ever. He never covers more than a sixth of the window.
**CAMERA**
Fixed rigid mount. Wide-angle spherical lens, roughly 80° horizontal field of view in vertical orientation, mild natural barrel distortion. Focus on the glass plane and on him.
The camera is a passenger with weight: continuous low-frequency sway for all 30 seconds — 2–4 pixels of vertical bob at about 1.5 Hz plus a slow 1–2 pixel roll — never stopping, never smooth, never perfectly still. At each pitch-up the shake spikes to 10–16 pixels of high-frequency vibration for 0.3 seconds and decays over 0.6 seconds; dust motes in the sunbeam jitter. No pans, no zooms, no dollies, no dutch-angle changes, ever.
**OPTICS**
Photoreal digital cinema capture. Fine organic grain. Real specular behaviour on metal and glass. Genuine dirt, water-spots and hairline scratches on the window that light up whenever the sun rakes across them. No anamorphic squeeze, no lens-flare overlays, no floating bokeh circles, no glow, no CG gloss, no beauty filter, no plastic skin.
**ACTION TIMING**
**0.0–1.0s — WORLD ONE, THE MEADOW, ground level.** Ordinary sunlit summer meadow tearing past at speed — long grass laid over in racing wind-waves, dandelion clocks shredding, buttercups whipping, the low treeline bending, pale blue sky with torn cirrus. Near grass is a hard horizontal smear; mid-field readable; treeline drifting with real parallax. Seed fluff streaks past the glass. <<<image_3>>> Everything flows left to right. Hard low sun rakes through the window, its bright rectangle lying across the ledge. The young man looks down at his phone, uninterested. One small thumb scroll.
**1.0–1.6s — PA CHIME.** A thin two-tone station chime from small ceiling speakers, tinny and compressed. He does not react — his eyes flick up two millimetres and drop back.
**1.6–3.4s — ANNOUNCEMENT.** A calm female voice through the overhead speakers, compressed, speaker-thin, faint echo down the empty carriage: {"Next station… Dream Core."} English, unhurried, exactly like a real transit announcement. Outside the meadow keeps tearing past and the sunbeam sweeps up the wall.
**3.4–4.0s — CHIME OUT.** The same two-tone chime closes the announcement. Under it the drive-note climbs.
**4.0–4.6s — FIRST PITCH-UP.** A hard mechanical lurch as the nose rises to 45°. The camera snaps 14 pixels and vibrates. His knees absorb it, his torso pitches forward 15°, his pole grip tightens — tendons raised, knuckles pale — and his left arm swings out, the phone tipping in his palm. His hair lifts and falls. He catches himself. THEN his head turns 40° toward the window, and his EYES GO THE SAME WAY as his head.
**4.6–5.5s — CLIMBING.** Flying forward and upward at 45°. The ground falls away fast: the treeline slides down out of the window's lower edge, the meadow shrinks to a patchwork with cloud shadows racing across it, the horizon bows, the sky expands. The sun swings across the window and its rectangle skews and climbs the wall. He slides the phone into his trouser pocket without hurrying — the earbud cable pulling into a shorter loop, swinging once against his chest — and brings both hands onto the pole.
**5.5–6.5s — CLOUD PASSAGE ONE.** Grey cumulus undersides tower up past the glass, their edges burning where the sun stands behind them, then we punch into the layer: for about one second the glass is a luminous white rectangle with filaments streaking past at speed, and the interior goes flat, shadowless, diffuse white — every hard shadow dies and the sunbeam vanishes. The train levels back to horizontal.
**6.5–11.5s — WORLD TWO, THE ROSE-QUARTZ VALLEY (<<<image_4>>>).** The white tears open and cool pink-silver daylight slams back across the ledge, the wall and his face, hard low sun burning through with strong veiling glare across the pane.
- Towering translucent pink crystal spires rise from the green valley floor in clusters, hundreds of metres tall, passing with heavy parallax — near spires sweeping by fast, far clusters drifting. Light REFRACTS through them: as each spire crosses the sun, the carriage dims and then a soft rose-tinted flash washes the interior as the sun burns through the crystal. Hard specular glints slide across their facets as our angle changes.
- The pink river winds glass-smooth through the valley, throwing sharp moving sun-sparkles and long pink reflections of the spires that ripple as gusts rake the surface.
- Both banks are crowded with large pale-pink blossoms whipping and nodding violently on thin stems, petals torn loose and streaking past the glass.
- The meadow grass is laid over in racing wind-waves, whole slopes rippling in travelling bands, the green killed and restored by racing cloud shadows.
- Smaller crystal boulders sit half-sunk in the grass like standing stones, still and heavy while everything around them moves.
- Soft green mountain ridges roll behind, nearly motionless with distance, torn cloud dragging across their shoulders.
- Birds cross in flocks against the pale sky.
At about 8.0s he takes half a step closer and rests both palms flat on the glass, arms nearly straight, fingers splayed, the flesh flattening and paling at the pressure points. His breath faintly fogs a small oval and clears. From here on he watches — captivated, curious, absorbed, with a small soft closed-mouth smile that barely moves.
**11.5–13.0s — SECOND PITCH-UP AND CLOUD PASSAGE.** The nose rises to 45° again with a gentler bump, the camera shaking 8 pixels, his hands staying on the glass. The crystal valley drops below the window's edge and we fly up into cloud: dense white for about one second, the interior flattening to shadowless white, the light rectangle skewing with the climb. Then level.
**13.0–17.5s — WORLD THREE, THE CLOUD-SEA HILLS (<<<image_5>>>).** The cloud tears open and hard blue-white daylight floods the carriage. We are flying level just above a rolling ocean of white fog, and colossal round emerald hills break through it like green whale-backs, everything in motion:
- The fog sea itself rolls and pours between the hills — thin veils torn off its surface by the wind, whole banks sliding down the slopes and re-forming, edges shearing continuously. Where the sun catches the fog tops they burn bright with a shifting sheen.
- Each hill summit carries one or two small white houses with dark pitched roofs, their windows flashing hard as they catch the low sun in passing. A wisp of chimney smoke off one of them is bent horizontal.
- The nearest hill sweeps past close and fast with violent parallax — its grass laid over in racing gust-waves, whole bright-green bands flattening and springing back — while the far hills drift slowly, half-drowned in fog.
- The narrow grey road winds up out of the fog and over a saddle between hills: two tiny cars crawl along it with real slow distance-speed, a streetlamp or two standing dead in the daylight. The road is a fast pale streak where it passes near, a thin thread where it climbs far.
- Every time a hill summit crosses the sun, the whole carriage dims and blows bright again — hill after hill, a slow heavy strobe.
- Birds cross in loose flocks low over the fog, banking and flashing in the light.
- Above everything, clean deep-blue sky with a few sheared cirrus.
His face goes from pink-silver to hard blue-white here, the hill-strobe throwing slow bars of shadow and light across his profile.
**17.5–19.0s — THIRD PITCH-UP AND CLOUD PASSAGE.** Nose to 45°, a bump, the camera shaking 8 pixels. The green hills fall away below and we climb into cloud: one second of dense white, interior flat and shadowless. Then level — but this time the white does not open onto sky.
**19.0–23.5s — WORLD FOUR, THE LANTERN CAVERN (<<<image_6>>>).** The cloud tears open into an immense open cavern world and daylight dies completely. The interior of the carriage instantly drops into deep blue-violet dusk lit only by what passes outside — this is the dark heart of the shot.
- Colossal arches and vaults of dark violet-blue rock sweep past overhead and to both sides with violent close parallax, each arch briefly swallowing the light as it crosses the window.
- Glowing turquoise-white falls of light pour off the rock faces in long luminous ribbons, their spray blown sideways in drifting sheets; each fall throws a moving cold-blue wash through the carriage as it passes.
- Hundreds of orange paper lanterns hang and drift through the air at different depths, swinging on the wind with real pendulum lag. Near lanterns sweep past the glass fast and hot, throwing quick warm pulses of orange light across the ledge, the wall and his face; far lanterns crawl. This alternating orange-warm / turquoise-cold strobing is the dominant lighting event of this world.
- Below, a long wooden market street runs along the river: stall after stall glowing from within, shelves of bottles and fruit catching lantern light, awnings rippling, hanging vines stirring, cloth signs illegible and flapping. A jetty walkway of worn planks streaks past as a warm blur where it comes near.
- The river itself glows turquoise, its shallows speckled with drifting motes of green-gold light, its surface rippling and throwing moving reflections of every lantern up onto the rock.
- Far down the cavern, a tall bright opening — a luminous pale doorway of daylight at the end of the world — visible as an intense glowing rectangle we are travelling toward, growing very slightly.
- Steam and smoke from the stalls bent horizontal by the draft; motes and embers streaking past close to the glass.
His face goes from hard daylight to this pulsing warm-cold flicker; his eyes track individual lanterns as they whip past. His faint reflection ghosts over the lanterns in the dark glass.
**23.5–25.0s — FOURTH PITCH-UP AND CLOUD PASSAGE.** Nose to 45°, a bump, the camera shaking 8 pixels. We climb toward the bright opening at the cavern's end; the market falls away below, the pale doorway blooms across the whole glass, and for one second we are inside dense luminous white — the interior flat and shadowless, hard daylight flooding back. Then level.
**25.0–30.0s — WORLD FIVE, THE SAKURA ISLAND (<<<image_7>>>).** The cloud tears open onto the widest, calmest vista of all and warm daylight floods back in. We are flying above an endless cream-white cloud ocean under soft blue sky.
- The great floating island approaches from frame left and passes majestically close: a colossal weeping cherry tree in full pink bloom crowning a raw rock island, its underside jagged and hanging with long pink blossom-vines that swing with real pendulum lag.
- The enormous canopy is shoved and shivering in the wind, whole blossom masses leaning as gusts move through them, cascades of hanging bloom whipping and trailing, never still. A continuous stream of petals is torn off and driven downwind, streaking past our glass at real speed.
- Under the canopy, the small dark-roofed wooden shrine and its red torii gate slide past, their edges catching the low sun; the mossy ground around them is carpeted pink, petals skittering across it in gusts.
- Blossom drapery pours over the island's cliff edges like pink waterfalls, stirring and lifting in the wind, the sun burning a bright rim along the island's edge whenever it stands behind it — and when the island itself crosses the sun, the whole carriage darkens for a long beat, then blows warm and bright as it clears.
- The cloud sea below rolls and billows, tops lit hard, hollows in soft shadow, veils sheared off and dragged sideways; its sheen shifts continuously as we travel.
- A few smaller cloud towers pass at different depths with clean parallax, edges burning.
- Distant flocks of birds wheel around the island, catching light as they bank. Long sun shafts break through the cloud gaps and sweep past our window.
From about 27.0s the drive note drops in pitch and the rail rhythm spaces out — clack… clack… clack — the world outside decelerating with real inertia, not slow motion, over three full seconds. Petals keep streaming past the slowing glass. At about 29.0s the train settles to a stop with a soft hydraulic sigh and one last gentle rock forward and back; His body absorbs that final sway through his knees. The last frame holds: the glass, the blossoming island beyond it drifting in the cloud sea, and his small silhouette at the edge.
**PHYSICS**
Everything inside obeys the vehicle. The carriage is always gently rocking, all 30 seconds. Dust motes hang in every light beam and jitter with the vibration. He always moves BECAUSE of the train — one frame after the jolt, never before it, never in perfect sync. His knit sweater compresses at the elbows and shoulders and holds its creases. The earbud cable swings with true pendulum lag, always a beat behind his torso. His hair is lifted and dropped by every impact.
Outside, scale governs speed: nearest ground streaks into blur, mid-ground moves fast with clean parallax, and the enormous distant elements — the far mountain ridges, the far hills, the cavern vaults, the floating island's far side, the cloud layers — move almost not at all. Nothing gigantic moves quickly. Nothing distant moves at the speed of something near. Falling water and falling light always fall downward and always throw mist or motes. Lanterns swing, they never teleport. The floating island drifts and rotates almost imperceptibly, it never bobs or wobbles.
**MATERIALS**
Painted metal window trim with real chips and grime in the corners. Worn ledge with dirt in the seams. Window glass with faint water-spots, one hairline scratch and dust along the rubber seal, all of it lighting up when the sun rakes across. His knit sweater: soft, matte, absorbent — light sinks into it, it never shines. Trousers: dense cream cotton with real seam puckering. Skin: subsurface scatter, visible pores, faint down on the cheek edge glowing where the sun rims him. Outside: real translucent mineral crystal with internal fractures and refraction, real grass, real fog, real rock, real weathered wood and paper lanterns, real bark and blossom, real water — never plastic, never CG-smooth.
**LIGHTING**
Motivated entirely by what comes through the window plus a weak neutral ceiling strip. The interior's colour and intensity change continuously: hard neutral daylight in the meadow → golden haze on each climb → shadowless white flood inside every cloud → cool pink-silver light with rose refraction flashes in the crystal valley → hard blue-white daylight with slow hill-strobing over the cloud-sea hills → deep blue-violet dusk with fast warm-orange / cold-turquoise lantern-strobing in the cavern → warm soft daylight in the sakura finale. The light rectangle must physically travel and skew across the ledge, the wall and his face in every sunlit world — it is never static — and it must DIE completely inside the cavern, replaced by moving lantern pulses. If the interior brightness stays constant across these worlds, the shot is wrong.
**INTEGRATION LOCK**
The young man is physically inside this carriage. His shadow falls on the ledge and wall and moves as the light moves. He is reflected faintly in the glass — most visibly inside the dark cavern, where his reflection ghosts over the lanterns. He shares the carriage's exposure, grain, haze and lens softness — never sharper, never brighter, never cleaner, never flatter than the trim beside him. The window frame and seal overlap in front of him where the geometry demands it.
**EMOTIONAL REGISTER**
Extremely restrained throughout. He is captivated and curious — absorbed by what's outside, with a small soft closed-mouth smile that barely changes for the entire video. His interest reads through his eyes tracking the view, slow blinks, a two-millimetre brow lift, and his body drawing closer to the glass. Banned completely: wide grins, open mouth, gasps, hands over the mouth, theatrical shock, giggling, tears, any big expression.
**AUDIO — SFX ONLY, NO MUSIC**
No music, no score, no ambient pads, no swells, nothing melodic anywhere. Only real diegetic sound.
Continuous wheel rumble and rail clatter changing character with each stage: hard steel-on-steel at ground level, thinning into wind rush and a deep drive-hum in the air, muffled and close inside each cloud, opening out again in each new world, slowing to spaced-out clacks at the end. Two-tone PA chime at 1.0s and again at 3.4s with the announcement voice between them. Metallic clang and bass thump at each of the four pitch-ups, the first hardest. Constant wind hiss through the window seal, rising and falling with the gusts. In the crystal valley: a faint high glassy resonance as the spires pass, river hiss, wind through the blossoms. Over the hills: open wind, distant house shutters knocking, birds. Inside the cavern: a vast hollow reverberant air, the roar of the glowing falls, faint market murmur and creaking wood and jetty planks, paper lanterns knocking softly, water lapping. In the finale: deep organ-like wind through the great cherry canopy, petals ticking faintly against the glass, distant birds. Carriage creaks and joints ticking throughout. Hydraulic sigh and settling creak at the stop. His only vocal sound is one soft exhale.
**POSITIVE CONSTRAINTS**
Window fills roughly 85% of every frame; the young man occupies only the right 15%. Exactly five worlds in this exact order: meadow, rose-quartz valley, cloud-sea hills, lantern cavern, sakura island. Exactly four cloud passages, each a 45° nose-up forward climb. One continuous take, one camera position, one lens, all 30 seconds. Forward travel only, no turns, no banking, no reversal, never a vertical lift. No slow motion anywhere. No horses, no aircraft, no spacecraft; the only vehicles ever visible are the tiny distant cars on the hill road in World Three. No other passengers. No stations or space visible; the cavern is an open world seen through the window, never a bored rail tunnel around the train. All text illegible — signage, stall labels and his phone screen blank, blurred or turned away. No brand logos. No exterior shot of the train. Full photorealism — real skin, real fabric, real metal, real water, real crystal, real blossom. No CG gloss, no plastic sheen, no beauty filter, no over-saturation, no orange-teal grade, no glow overlay. Real highlights, true contrast, no grey washed veil.

06Как читать этот промт

27 тысяч знаков — не потому что я люблю писать. Это сценарий по секундам, и каждый блок держит одну конкретную ошибку модели:

Заменить мир — подменить картинку в слоте и абзац с её описанием в двух местах: в ACTIVE REFERENCES и в таймлайне. Объявление можно сделать русским: Seedance озвучивает русскую речь, если написать её латиницей и как звучит — вместо "Next station… Dream Core" поставить "Sleduyushchaya stantsiya… Dream Core" и добавить в блок SOUND строку, что реплика — русская речь в латинской транслитерации, произносить как носитель.

07Шаг 5 — апскейл до 4K

Seedance отдаёт 480p — для ленты этого мало. Финальный файл прошёл через Topaz Video в том же Higgsfield: модель улучшения thd-3 плюс интерполяция кадров apo-8 до 30 fps. На выходе 2160×3840.

Порядок именно такой: сначала добиться правильного ролика в 480p, потом один раз апскейлить. Апскейлить каждую попытку — выбрасывать кредиты.

08Где ломается