June 27

Как снять POV-видео с драконом с помощью Seedance 2.0

Я всё делал в SYNTX AI.

Использовал нейросети:

  • Seedance 2.0 Pro
  • ChatGPT Image 2.0

В этом гайде разберём видео из двух отдельных генераций.

  1. Первый отрезок — горизонтальное BTS-видео, где камера на штативе снимает вас возле замка, а огромный дракон разрушает башню.
  2. Второй отрезок — вертикальное видео уже с телефона, где вы снимаете дракона и его загадочного хозяина вдали.

1. Подготовка референсов

Для проекта понадобится 4 изображения и 1 видеореференс.

Подготовка лица

Для начала необходимо подготовить фото лица в виде сетки 3×3 на белом фоне.

Это позволит нейросети максимально хорошо сохранять вашу внешность во время генерации.

Этот файл загрузим как:

@Image1 — Face Reference


Подготовка одежды

Теперь понадобится отдельный референс одежды.

В моём примере используется определённый комплект одежды, однако вы можете использовать абсолютно любую одежду.

Главное, чтобы она была хорошо видна на фотографии.

Этот файл загрузим как:

@Image2 — Clothing Reference


Подготовка локации

Следующий референс — сама локация.

В этом проекте используется старый каменный замок, однако его также можно заменить практически любой другой локацией, если вы хотите изменить сюжет ролика.

Этот файл загрузим как:

@Image3 — Location Reference


Подготовка дракона

Теперь понадобится изображение самого дракона.

Именно этот референс отвечает за внешний вид существа в обоих видео.

Этот файл загружаем как:

@Image4 — Dragon Reference


Подготовка видеореференса

Для второго ролика дополнительно понадобится короткий видеореференс.

Он используется исключительно для движения камеры.

Нейросеть повторит:

  • движение рук;
  • переключение между фронтальной и основной камерой;
  • цифровой зум;
  • темп съёмки.

При этом внешний вид человека, одежда, освещение и окружение из видео использоваться не будут.

Этот файл загружаем как:

@Video1 — Motion Reference


2. Первый отрезок — возле замка

Настройки:

Видео → Seedance 2.0 → Режим Pro → 720p → Соотношение сторон 16:9 → Длительность 11 сек.

Загружаем файлы по порядку:

  1. @Image1 — Face Reference
  2. @Image2 — Clothing Reference
  3. @Image3 — Location Reference
  4. @Image4 — Dragon Reference

⚠️ ВАЖНО!

Перед запуском генерации обязательно убедитесь, что каждый референс привязан к правильному тегу внутри промпта. Если перепутать хотя бы один файл, результат получится совершенно другим.

ПРОМПТ ДЛЯ ПЕРВОГО ОТРЕЗКА

  • Total: 11s / Single continuous static smartphone shot / 16:9.
    Style: raw ungraded amateur smartphone footage, real behind-the-scenes phone recording. Everyday realism, natural imperfections, mild auto-exposure shifts, autofocus hunting, motion blur on fast movement, light compression noise. 35mm normal lens, no fisheye. Heavy overcast cold gray flat light, light rain, damp hazy air, blown-out white sky, slightly underexposed foreground. ZERO cinematic look, ZERO CGI look — like a clip someone shot on a phone by accident. Lens droplets: at the very start only one or two tiny soft out-of-focus rain droplets sit on the lens glass, subtle and blurred. Then partway through the clip ONE more small droplet lands on the lens — it appears mid-shot as if rain just hit the glass, staying small, soft and blurred. Never more than a few tiny droplets total, never covering the frame. Use @image1 ONLY for the man's exact facial identity. Use @image2 ONLY for his clothing: copy his full outfit exactly as in the reference, every layer and detail, no changes. Use @image3 ONLY for the castle: realistic old stone castle near Paris on an open GRASS FIELD, overcast, light rain. A real physical place, not a backdrop. Use @image4 ONLY for the wyvern's shape/appearance. Scene: an open wet GRASS FIELD in front of the castle. The man stands ON THE GRASS — green wet grass with puddles, no stone, no pavement. A horizontal backstage phone is locked on a tripod several meters away, filming him from the front. Closer to him, a modern smartphone is mounted VERTICALLY on a tall tripod. CRITICAL FRAMING — READ CAREFULLY: the vertical phone is turned so its GLOWING FLAT GLASS SCREEN faces the backstage camera, i.e. the screen faces US / the viewer. We clearly see the bright glowing screen, NOT the camera lenses. The phone's REAR camera lenses point AWAY from us, toward the man, so the rear camera is what films him. DO NOT show the triple-lens rear camera block facing us. DO NOT turn the phone screen toward the man. The side facing the viewer is the SCREEN; the side facing the man is the LENSES. This vertical tripod phone stays clearly visible the whole time. A gear bag sits on the wet grass nearby. Camera: completely static, locked on tripod. No pan, handheld move or zoom. Fixed frame the entire clip, except a tiny natural shake at the very end when the phone is grabbed. [0-4s] The man (face @image1, outfit @image2) poses casually for the vertical tripod phone — steps back into pose, shifts his weight, hands into pockets, rolls his shoulders, slightly cold from the rain, small idle natural movements. He looks toward the tripod phone, never at the backstage camera. He does NOT notice anything yet. [4-7s] Behind and above him, from BEHIND the right turret, the wyvern (@image4) appears for only a moment: a small, plain dark distant silhouette far off in the sky, low-detail and heavily motion-blurred through the rain, with almost no visible wing-flapping — it simply glides fast and stiff across the gray sky high near the top of the frame, far above and behind the man, then is gone out to the LEFT. It stays small and distant the whole time, NOT the focus, NOT clean, NOT smoothly animated, NOT sharp — just a quick blurry shape a phone barely caught. As it clips the right turret, a wing strikes the top: a sudden hard impact knocks off a visible chunk of stone — part of the turret top breaks away and is missing, blocks tumble down the wall, dust and rain mist burst out with a loud crash. The man's reaction is triggered by THAT CRASH behind him, NOT by the wyvern: the instant the stone breaks and the loud noise hits, he flinches and twists around to look back over his shoulder toward the collapsing turret. His head and upper body turn toward the falling stone behind him, hands coming out of pockets, eyebrows up, eyes a bit wider — a moderate, believable "what was that" reaction, brisk and alert but contained, clearly more than idle and well short of panic. By the time he is turned, the wyvern is basically already gone; he is reacting to the damaged turret and the noise, not tracking the creature. This damage MUST REMAIN for the rest of the clip: turret stays broken, dust drifting, debris settling on the grass — the castle must NOT snap back to intact. The wyvern is NOT knocked down and does NOT return. [7-11s] Still half-turned toward the broken turret, he then turns back and moves to the vertical tripod phone at a brisk, purposeful walk-into-jog, with light urgency but not a frantic sprint. He grabs ONLY THE PHONE with one hand — fingers wrap its body and pull it OUT of the tripod clamp in one quick motion. The phone comes free while the tripod stays standing and just wobbles/rattles; he does NOT grab or lift the tripod. The backstage camera gets a tiny shake. End the instant the phone is pulled free — phone in hand, empty tripod wobbling, action unfinished, about to continue into a separate vertical 9:16 segment. Do NOT show the phone's screen change. Audio: diegetic only — light rain, damp outdoor ambience, distant wind, wet footsteps on grass, a loud sudden crash of breaking stone and falling debris, tripod clamp rattle, a brief low wing whoosh as the wyvern passes. Muffled low-quality phone-mic sound, wind rumble, slight auto-gain pumping, peaking like a cheap mic. No music, no dialogue. Negative: phone rear camera facing the viewer, triple-lens camera block facing us, lenses visible on the side facing the backstage camera, phone screen facing the man, dark phone back facing us, no glowing screen visible, CGI/3D-render look, clean sharp detailed creature, smoothly animated creature, over-animated flapping wings, slow-motion creature, large creature, big creature, creature in focus, creature close to the man, low-flying creature, creature as the main subject, epic or cinematic framing, glossy fantasy, cartoon, dragon hero shot, drone shot, pan/zoom/fisheye/distortion, handheld backstage move, cuts or transitions, extra creatures, background people, crowd, walking silhouettes, sunny weather, dry ground, stone floor, pavement, cobblestone, frozen mannequin posing, theatrical or exaggerated acting, calm slow reaction, no reaction, frozen stare, reacting to the dragon instead of the crash, turning only when the dragon flies overhead, hesitation before moving, frantic panic, wild flailing, big jump, ducking, arms thrown up, dramatic recoil, lens fully covered in water, heavy rain droplets on lens, all droplets present from the very start, turret intact after impact, castle resetting to undamaged, grabbing or lifting the tripod. The wyvern must NOT fall, crash or tumble or return — it passes once and is gone.

3. Второй отрезок — съёмка дракона

Настройки:

Видео → Seedance 2.0 → Режим Pro → 720p → Соотношение сторон 9:16 → Длительность 10 сек.

Загружаем файлы:

  1. @Image1 — Face Reference
  2. @Image2 — Clothing Reference
  3. @Image3 — Location Reference
  4. @Image4 — Dragon Reference
  5. @Video1 — Motion Reference

ПРОМПТ ДЛЯ ВТОРОГО ОТРЕЗКА

Total: 10s / single continuous handheld VERTICAL smartphone shot / 9:16.

STYLE: raw ungraded amateur smartphone footage, real one-hand phone recording. Natural imperfections, mild auto-exposure shifts, motion blur on fast moves, light compression noise. Phone wide lens (~35mm), no fisheye. The image is simply sharp — NO focus hunting, NO cinematic rack focus. Zero CGI look.

MOTION REFERENCE — @Video1: reference @Video1 STRICTLY for camera movement, handheld behavior, the selfie-to-rear camera switch, the pinch-zoom, panning, pacing and framing. Use @Video1 ONLY as a motion reference. Do NOT copy the person, face, clothing, background, lighting, weather or environment from @Video1.

IMPORTANT — ONE PHONE ONLY: the footage IS the recording from the man's phone. There is NO phone visible anywhere in the frame, NO second phone in his hand. His free hand is empty. Never show a phone being held or flipped.

ATMOSPHERE — READ @Image3: @Image3 (the castle) is the master reference for the whole mood — copy its overcast flat cold gray light, washed-out white sky, muted desaturated color, dampness and FAINT barely-visible drizzle. Apply this look to the entire frame, including where the castle is not visible. Keep the rain as weak as in @Image3 — almost imperceptible damp haze, NO downpour, NO rain streaks. Do not shift warmer/cooler, do not boost saturation or contrast.

GEOMETRY: the rear camera faces AWAY from the castle, across an open wet grass field fading into rain haze and a blown-out white horizon. The castle is only behind the man for a split second at the very start; for the rest of the clip it is NOT in frame. Invent the empty field naturally in the same gray look.

REFERENCES:

@Image1 = the man's exact face ONLY.

@Image2 = his clothing ONLY — copy the full outfit exactly.

@Image3 = the already-damaged castle (broken turret, never intact) AND the atmosphere source above.

@Image4 = the wyvern's shape ONLY.

The hooded figure has NO reference: a tall person in a long dark wet rain cloak, hood up, face hidden in shadow. Plain and grounded — NOT glowing, no magic, no staff light.

[0-3s] Handheld selfie/front-camera POV, matching @Video1. Opens immediately on the man's face — his face (@Image1) fills the upper frame, outfit (@Image2) visible, damaged castle behind him in the rain. Sharp from frame one. Brief handheld wobble, then his hand settles. He looks tense, alert, breath visible in the cold.

[3-4s] CAMERA SWITCH, matching @Video1 — like double-tapping the screen in TikTok: the feed INSTANTLY flips from the front camera to the rear camera. This is a digital camera switch, NOT a physical phone flip — the phone is NOT turned around in front of us, nothing rotates, no phone appears in frame. One frame it's his selfie face, the next frame it's the rear view across the open field, AWAY from the castle. Only gray grassland in rain haze, no castle in this direction.

[4-6s] Phone PINCH-ZOOM toward the far field, matching @Video1. CRITICAL: the wide shot looks clean, but the moment he zooms the image quality visibly DROPS like real smartphone digital zoom — softer, grainier, mushy, pixelated, with heavy compression artifacts and shakier framing. The farther it zooms, the noisier it gets. Through this awful zoomed-in mush, far away in the rain, the wyvern (@Image4) is LANDING right beside the hooded figure — coming down onto the wet grass, wings flaring and folding, legs reaching for the ground, kicking up spray as it settles. SCALE IS CRITICAL: this is a COLOSSAL dragon, the same beast that smashed a castle turret with its wing — it absolutely TOWERS over the hooded figure, many times the figure's height. The human-sized figure beside it looks TINY, barely up to the wyvern's leg. There IS motion here, not a static creature. But it stays barely readable: dark, low-detail, you can only just make out its huge shape through the grain. The hooded figure (its master) beside it is even harder to see — a tiny dark smudge in the pixelated haze, dwarfed by the beast.

[6-8s] The huge wyvern has settled on the ground next to the tiny figure. Slow quiet beat: the colossal creature lowers its enormous head all the way down toward its master; the small figure has to REACH UP, raising one hand, and TOUCHES the wyvern's snout, slow and almost tender, like calming a horse far bigger than itself. The wyvern stays still and obedient. Still seen through the same degraded zoom — soft, grainy, barely-there detail, only just distinguishable in the mush. Eerie — this giant thing smashed the turret, now gentle under the hand of a person a fraction of its size.

[8-10s] Far away on the horizon, a single faint DISTANT LIGHTNING FLASH in the clouds (see below); right after it the hooded figure slowly LIFTS its head and looks straight at the camera across the field, as if it knows it is filmed. The man's hand jolts slightly. Hold on the figure's hidden face turned at the lens, then abrupt cut — moment unfinished.

LIGHTNING — MUST BE BARELY NOTICEABLE, FAR AWAY, REAL NOT CGI: NO visible bolt, NO zigzag, NO strike hitting the ground, NOT near the figure. Only a faint distant sheet-lightning glow deep in the clouds on the far horizon — a soft flicker of light inside the overcast, very small and far. It lasts a SINGLE FRAME — an instantaneous blink you'd barely catch out of the corner of your eye, NOT clear, NOT distinct, NOT a 1-2 second flash, no fade, no afterglow, no brightening of the whole scene. A low delayed thunder rumble follows a beat later.

AUDIO: diegetic only — light rain, damp ambience, distant wind, the man's tense breathing, low delayed thunder after the flash. Muffled low-quality phone-mic sound, wind on the mic. No music, no dialogue.

CONSTRAINTS: use @Video1 ONLY for camera movement, pacing, framing and the selfie→rear switch and zoom choreography — do not copy anything visual from it. Front-to-rear camera switch is a double-tap digital switch, never a physical flip. Keep the man's face (@Image1) and outfit (@Image2) consistent, castle (@Image3) always damaged. The wyvern must be COLOSSAL, towering many times over the hooded figure. No second phone, no phone in frame, free hand empty. No lens droplets. No tripod/gimbal smoothness, no fisheye, no cuts mid-clip, no transitions, no slow motion, no horizontal framing. Grounded realistic physics.

NEGATIVE: small wyvern, tiny dragon, wyvern same size as person, human-sized creature, wyvern shorter than figure, undersized dragon, figure taller than wyvern; second phone, phone in hand, visible phone, holding a phone, flipping the phone, hand rotating; lens droplets, water on lens; sharp creature detail, clear wyvern, crisp dragon, readable figure in zoom, high-detail zoom, creature already grounded at start of zoom, static landed creature; jagged/zigzag/glowing/CGI/cartoon lightning, lightning bolt, lightning hitting ground, lightning near figure, clear or distinct lightning, lingering or multi-second flash, slow fade, sustained bright sky, strobing; glowing wizard, magic glow, staff light, energy aura; focus hunting, rack focus, clean/sharp/optical zoom; gimbal smoothness, fisheye; heavy rain, downpour, rain streaks; warm grade, saturated, high contrast; intact castle, repaired turret, castle in the field; creature flying/attacking/close/large in frame; figure walking or running toward camera; slow motion, horizontal framing, cuts mid-clip.


4. Монтаж

После генерации сохраняем оба ролика и соединяем их в любом видеоредакторе.

Сначала ставим первый горизонтальный отрезок.

В момент, когда персонаж снимает телефон со штатива, сразу делаем склейку на второй вертикальный ролик.

Именно благодаря этому переходу создаётся ощущение, что всё происходило непрерывно и было снято одним телефоном.

При желании можно добавить небольшую тряску камеры, компрессию и лёгкую телефонную цветокоррекцию для ещё большей реалистичности.