ai21.net
语言

行车记录仪首帧锁死:Image 2.5→Seedance 公路短片

先用 Image 2.5 把 dashcam 光学真实感钉死(挡风玻璃脏点、A柱、宽角JPEG),再整帧喂 Seedance 2.5;回复含完整 DIRECTIVE + 秒表分镜。谁还在瞎抽视频首帧,先把参照图焊死。

演示视频

原帖未提供效果图

提示词

First, lock the initial frame to a realistic one. I've used GPT Image 2.5 and the prompt below:

DIRECTIVE:
Produce one still that reads as a real in-car dashcam frame grabbed from a moving car on a highway. Optical dashcam capture, wide windshield view, windshield glass, A-pillar, a slice of dashboard/hood — lived-in dashcam JPEG, not cinema, not HDR.

BEAT / COMPOSITION:
Looking forward through the windshield. On the LEFT side of the road (left lane or left shoulder, clearly in frame): TWO branded trucks close together.

COCA-COLA TRUCK (LEFT, CRITICAL):
A full-size Coca-Cola tanker / delivery truck, official Coca-Cola red livery and logos readable. It has a breakdown: a torn ragged hole in the tank wall. A thick, heavy jet of dark-brown Coca-Cola is blasting out of that hole onto the asphalt — lots of liquid, puddle spreading, foam, spray in the air, wet road shine. The truck is stopped or crawling, hazard situation.

MENTOS TRUCK (BESIDE IT, CRITICAL):
Right next to the Coca-Cola truck (same left cluster, slightly ahead or alongside): a closed box truck / delivery truck with large, unmistakable MENTOS branding on the side (Mentos logo, candy rolls artwork). Rear and side doors CLOSED. No candy spilling. You can clearly read that it is a Mentos truck.

CAMERA PACK:
Fixed dashcam behind the windshield, slight barrel wide, dashboard or hood bottom of frame, windshield dirt/reflections, timestamp overlay optional, daytime road, other traffic farther ahead. Real consumer dashcam still.

LIGHT:
Daylight, overcast or sun, real road color, cola looking like dark soda not black oil.

PHOTOGRAPHIC CHARACTER:
Unstaged dashcam grab — physically grounded trucks, readable brands, the leak is the event.

Once it's ready, use it as reference for Seedance 2.5 and this prompt:

[STYLE + CAMERA + ATMOSPHERE]
Vertical 3:4 smartphone video filmed FROM INSIDE a moving car, passenger or driver seat. Phone held handheld (not a bolted dashcam): arms, slight breathing shake, windshield and A-pillar in frame, a slice of dashboard/hood at the bottom, maybe a finger edge. We are ARRIVING — forward motion first, trucks growing larger, then we slow and stop short of the cola lake. CONTINUOUS SINGLE TAKE the entire 30s. Digital pinch-zooms only (no cuts): punch into the Coke jet, the jug fill, the Mentos doors, the launch, then zoom out as he shrinks to the horizon. Daylight, overcast, wet asphalt, rain specks on glass. Consumer phone JPEG, autofocus hunting, rolling shutter. Photoreal live physics: cola viscosity, foam, Mentos cases, one adult body launched like a Mentos rocket.

LOCKED SET / VEHICLES:
- LEFT ahead through the windshield: red Coca-Cola tanker with a ragged hole, dark-brown Coca-Cola ALREADY blasting onto the lane, spreading cola lake.
- BESIDE it: Mentos box truck, rear doors CLOSED at arrival, Mentos livery readable ("mentos / The Freshmaker").
- Highway continuing past them toward a clear vanishing-point horizon.
- NO other truck for the guy. He does not exit a vehicle.

LOCKED CAST:
- THE GUY: one very large man alone — sweat-dark tank top, cargo shorts, stained baseball cap, Alabama redneck energy. No second person. He ENTERS FRAME FROM THE RIGHT (shoulder / right edge of the windshield view) already on foot — he does not get out of any truck, van, or car.
- HERO OBJECT: one giant gas-station plastic jug / Big Gulp–scale cup in his right hand. Starts EMPTY (daylight visible through clear plastic), stays in that hand until the blast.

Diegetic audio: road noise while arriving, cola hiss, gulping, metal latch, volcanic foam roar, Doppler scream as he flies away, phone mic clipping.

[IMAGE REFERENCES]
[ref_image] = locked opening phone still from inside the car looking through the windshield at the Coke tanker + Mentos truck cluster. Tanker, hole, cola jet, Mentos truck, and in-car foreground are LAW. Clip BEGINS with us still rolling toward that cluster.

[TIMELINE SECOND BY SECOND]

0-3s: [HOOK — ARRIVING + ZOOM] Phone already rolling. Road noise. Through the windshield the Coca-Cola tanker and Mentos truck are already ahead on the left; cola ALREADY jetting from the hole. We close distance — trucks grow. Digital zoom punches into the Coke stream and the hole. By ~2.5s brakes dip. Several events already: motion, the leak, closing in, first zoom.

3-5s: [STOP + HE ENTERS FROM RIGHT] We fully stop. Idle phone shake. Zoom out a notch. THE GUY walks INTO FRAME FROM THE RIGHT — already on foot from the shoulder / right edge, giant EMPTY jug held high, clear plastic with daylight through it. He does not exit any vehicle. He steps straight into the cola lake toward the tanker jet.

5-9s: [FILL — HOLD IT] He shoves the empty jug INTO the Coke stream. ONE CONTACT: jug into the jet. Phone slam-zooms the fill and HOLDS. For a clear ~2 full seconds (through ~7–8s) the jug is STILL filling under the blast — brown liquid climbing the plastic, foam boiling over the rim, overflowing his forearm, stream hammering the mouth of the jug the whole time. Only then does he pull the full heavy jug out and raise it.

9-12s: [CHUG] He tilts and CHUGS — huge gulps, cola down the tank top. Empties most of it, gasps, still holding the jug. Turns to the Mentos rear doors. Phone zooms out enough to keep tanker jet + Mentos bumper + him in frame.

12-16s: [OPEN + DUMP] ONE CONTACT: he yanks both Mentos rear doors open. Cases stacked high. He dumps a case — Mentos rolls cascade into the cola lake. Packs split, white mints into the brown river. Digital zoom on the mint waterfall hitting the puddle. He still stands in the lake with the jug.

16-20s: [IT GOES] Hard fizz, then a WHITE FOAM COLUMN erupts between the trucks. ONE CONTACT: the fountain hits him from below like a human Mentos rocket. Feet leave asphalt. Jug flies. He launches UP and AWAY down the highway axis. Phone slam-zooms his body leaving the ground, then struggles to track him.

20-24s: [HUMAN FOUNTAIN] He is a dark silhouette riding the white column, spinning once, cap gone. The geyser leans him toward the open highway horizon. Foam hits our windshield — phone jerks but stays inside the car, still filming through glass. Digital zoom hunting his shape in the spray.

24-26s: [SHRINKS] Zoom out through foam streaks on the glass. He is airborne over the far lanes, a tiny spinning figure on a white spray trail, Mentos truck hopping, tanker still feeding the column. Phone digital-zooms into the shrinking speck, then out again.

26-30s: [PAYOFF — GONE] He dwindles to a speck and DISAPPEARS into the horizon haze past the vanishing point (~1s earlier than before). Phone holds on empty sky/road where he vanished, foam still climbing between the trucks, Mentos doors still open, Coke still dumping. Final tiny digital zoom into that empty vanishing point. Loop-open: empty horizon, roar still going.

[STYLE & QUALITY BOOSTERS]
Photoreal handheld phone from inside a car, 3:4 vertical. Arriving motion then stop. Digital pinch-zooms only, continuous take. One man enters from the RIGHT on foot — never exits a vehicle. Empty jug fills on camera for a clear multi-second hold. Body launched to the vanishing point. Coherent liquid physics. Identity lock to [ref_image] for both trucks, the hole, and the in-car windshield framing.

中文译文

首先,把首帧锁死在一张写实的画面上。我用了 GPT Image 2.5 和下面这段 prompt:

指令(DIRECTIVE):
生成一张静帧,读起来要像从高速公路上一辆行驶中的车里抓取的真实行车记录仪画面。行车记录仪光学拍摄、宽广的挡风玻璃视角、挡风玻璃、A 柱、一小截仪表台/引擎盖——充满使用痕迹的行车记录仪 JPEG,不是电影感,不是 HDR。

节拍 / 构图:
透过挡风玻璃向前看。在道路的左侧(左车道或左路肩,清晰地出现在画面内):两辆带品牌的卡车紧挨在一起。

可口可乐卡车(左侧,关键):
一辆全尺寸的可口可乐罐车 / 配送卡车,官方可口可乐红色涂装,logo 清晰可读。它抛锚了:罐体壁上有一个撕裂的破洞。一股粗壮厚重的深棕色可口可乐正从那个洞里喷涌到沥青路面上——大量液体、水洼不断扩散、泡沫、空气中有飞溅、湿漉漉的路面反光。卡车已经停下或正在缓慢爬行,处于危险状况。

曼妥思卡车(就在它旁边,关键):
紧挨着可口可乐卡车(同一个左侧车组,稍微靠前或并排):一辆关闭的厢式卡车 / 配送卡车,侧面有巨大且不容错认的 MENTOS 品牌标识(Mentos logo、糖果卷图案)。后门和侧门都关闭。没有糖果洒出。你能清楚读出这是一辆 Mentos 卡车。

摄影机配置:
固定在挡风玻璃后的行车记录仪,轻微桶形畸变的广角,仪表台或引擎盖位于画面底部,挡风玻璃有污渍/反光,时间戳叠加可选,白天道路,更远处还有别的车流。真实的消费级行车记录仪静帧。

光线:
日光,阴天或晴天,真实的道路颜色,可乐看起来像深色汽水而不是黑色机油。

摄影质感:
未经摆拍的行车记录仪抓拍——卡车的物理质感扎实落地,品牌清晰可读,泄漏本身才是事件。

出图之后,把它作为 Seedance 2.5 的参考图,并使用这段 prompt:

[风格 + 摄影机 + 氛围]
竖屏 3:4 手机视频,从一辆行驶中的车内部拍摄,乘客座或驾驶座。手机手持(不是固定死的行车记录仪):手臂、轻微的呼吸抖动,挡风玻璃和 A 柱在画面内,底部有一小截仪表台/引擎盖,也许还有手指边缘。我们正在抵达——先是向前运动,卡车逐渐变大,然后我们减速并在可乐湖前停下。整整 30 秒一镜到底。只用数码推拉变焦(无剪辑):推近可乐喷射、桶的灌装、Mentos 车门、起飞,然后在他缩小成地平线上的小点时拉远。日光、阴天、湿沥青、玻璃上的雨点。消费级手机 JPEG、自动对焦反复寻找、滚动快门。照片级实拍物理:可乐黏度、泡沫、Mentos 箱子、一具成年人的身体像 Mentos 火箭一样被发射出去。

锁定场景 / 车辆:
- 透过挡风玻璃,左前方:红色可口可乐罐车,罐体上有一个撕裂的破洞,深棕色可口可乐已经在喷涌到车道上,可乐湖正在扩散。
- 在它旁边:Mentos 厢式卡车,抵达时后门关闭,Mentos 涂装清晰可读("mentos / The Freshmaker")。
- 高速公路从它们旁边继续延伸,通向清晰的地平线消失点。
- 没有别的卡车给这个家伙。他不是从任何车辆里下来的。

锁定人物:
- 那个家伙:只有一个体型极其庞大的男人——被汗浸深的背心、工装短裤、有污渍的棒球帽、阿拉巴马红脖子气质。没有第二个人。他从右侧入画(路肩 / 挡风玻璃视角的右边缘),已经是步行状态——他不是从任何卡车、厢式车或轿车里下来的。
- 关键道具:他右手拿着一个加油站巨型塑料桶 / Big Gulp 尺寸的大杯。一开始是空的(日光能透过透明塑料看到),一直握在那只手里,直到爆炸发生。

画内音:抵达时的路面噪音、可乐的嘶嘶声、大口吞咽声、金属门闩声、火山般的泡沫轰鸣、他飞走时的多普勒尖啸、手机麦克风爆音。

[图像参考]
[ref_image] = 锁定的开场手机静帧,从车内透过挡风玻璃看向可乐罐车 + Mentos 卡车车组。罐车、破洞、可乐喷射、Mentos 卡车以及车内前景都是铁律。片段以我们仍在朝那个车组行驶开场。

[逐秒时间线]

0-3s:[钩子——抵达 + 变焦] 手机已经在录制。路面噪音。透过挡风玻璃,可口可乐罐车和 Mentos 卡车已经在左前方;可乐已经在从洞里喷射而出。我们缩短距离——卡车变大。数码变焦猛地推近可乐液流和破洞。到约 2.5 秒时刹车下沉。已经发生好几个事件:运动、泄漏、逼近、第一次变焦。

3-5s:[停车 + 他从右侧入画] 我们完全停下。手机怠速抖动。变焦拉远一档。那个家伙从右侧走进画面——已经是从路肩 / 右边缘步行而来,高举着巨型空桶,透明塑料里透着日光。他不是从任何车辆里下来的。他径直踏进可乐湖,朝罐车的喷射流走去。

5-9s:[灌装——保持住] 他把空桶塞进可乐射流里。唯一一次接触:桶进入喷射流。手机猛推变焦到灌装并保持住。在清晰的大约整整 2 秒里(约 7–8 秒期间),桶仍在喷涌下持续灌装——棕色液体顺着塑料往上爬、泡沫沸腾溢出桶口、漫过他的小臂,液流自始至终猛砸着桶口。直到这时他才把灌满的沉重桶抽出并举起。

9-12s:[豪饮] 他仰头豪饮——大口吞咽,可乐顺着背心往下流。喝掉大部分,喘了口气,仍然拿着桶。转向 Mentos 的后门。手机拉远变焦,足以让罐车喷射 + Mentos 车尾保险杠 + 他本人同时留在画面内。

12-16s:[开门 + 倾倒] 唯一一次接触:他猛地把 Mentos 的两扇后门拉开。箱子堆得很高。他倒下一箱——Mentos 糖果卷成串倾泻进可乐湖。包装裂开,白色薄荷糖落入棕色河流。数码变焦对准砸进水洼的薄荷糖瀑布。他仍然拿着桶站在湖里。

16-20s:[它爆发了] 剧烈起泡,随后一道白色泡沫柱在卡车之间喷发而出。唯一一次接触:喷泉从下方击中他,像一枚人体 Mentos 火箭。双脚离开沥青路面。桶飞了出去。他沿着高速公路轴线向上、向外发射出去。手机猛推变焦对准他离地的身体,随后吃力地跟拍他。

20-24s:[人体喷泉] 他成为骑在白色泡沫柱上的深色剪影,旋转了一圈,帽子不见了。喷柱把他倾向前方开阔的高速公路地平线。泡沫打到我们的挡风玻璃上——手机猛地一晃但仍在车内,继续透过玻璃拍摄。数码变焦在飞沫中搜寻他的身形。

24-26s:[缩小] 透过玻璃上的泡沫条纹拉远变焦。他悬在远处车道上空,是白色水雾尾迹上一个旋转的小小身影,Mentos 卡车在跳动,罐车仍在为泡沫柱供给。手机数码变焦推近那个不断缩小的斑点,然后又拉远。

26-30s:[回报——消失] 他缩小成一个斑点,消失在地平线雾霭中、越过消失点(比之前早约 1 秒)。手机停留在空无一物的天空/道路上——他消失的地方,泡沫仍在卡车之间攀升,Mentos 车门仍然敞开,可乐仍在倾泻。最后对着那个空荡的消失点做一次很小的数码变焦。循环开放结尾:空荡的地平线,轰鸣仍在继续。

[风格与质量增强]
照片级手持手机拍摄,从车内取景,3:4 竖屏。先是抵达的运动,然后停下。只用数码推拉变焦,一镜到底。一个男人从右侧步行入画——绝不是从车辆里下车。空桶在镜头前灌装,并保持清晰的多秒画面。身体被发射到消失点。连贯的液体物理。两辆卡车、破洞以及车内挡风玻璃取景都锁定到 [ref_image] 保持身份一致。

溯源

原作者
techhalla
发布日期
2026-09-17
原始出处
查看原帖 ↗
收录来源
xianyu110/awesome-gpt-image2.5 · 2100526482212712584 · MIT