Das Ziel war eine einzige Einstellung in einem speziellen Raum eines fotorealistischen Miniatur-Puppenhauses, die mit einer Kamerabewegung in den nächsten Raum endet. Sie dauert etwa neun Sekunden. Ein winziger Händler isst Nudeln vor drei Monitoren. Ein riesiger Mann - ich - beobachtet ihn durch das hintere Fenster. Der Händler schnauzt: „Lass mich in Ruhe!“ und wirft etwas gegen das Glas. Der Riese sagt „Sorry!“, und die Kamera gleitet weiter in den benachbarten Yoga-Raum, in dem eine Frau auf ihr Tablet starrt.

Die ursprüngliche kurze Einstellung wirkte einfach, erwies sich aber in jedem Produktionsschritt als schwierig: zwei sprechende Charaktere mit Lippensynchronität, ein Wurf mit echter Physik, eine kontinuierliche Kamerabewegung zwischen zwei Räumen und ein großes Gesicht, das den Frame für Sekunden ausfüllt. Hier sind die Ergebnisse im direkten Vergleich, zusammen mit den Prompts.

Der Plan: zuerst drei Bilder

Die Methode bestand darin, die Schlüsselbilder zu genehmigen, bevor Videos generiert wurden: die Eröffnung des Trader-Zimmers, der Riese, der ihn beobachtet, und der Schlussframe im Yoga-Raum.

Start

Das Zimmer des Traders

Der Miniatur-Trader an seinen drei Monitoren, eine Nudelschale haltend; ein schmaler Streifen des Gesichts des Riesen am Fensterrand; eine dunkle Trennwand auf der linken Seite.

Ein Ausschnitt des vorherigen Raums, die Trennwand und nur der Rand des Gesichts des Riesen.

Mitte

Der Riese beobachtet

Der riesige blonde Mann füllt das hintere Fenster aus und blickt auf den Händler herab.

Sein Gesicht füllt das Fenster aus, die Augen sind auf den Trader gerichtet. Dieser Frame wird später zum gemeinsamen Schnittpunkt.

Ende

Der Yoga-Raum

Der Yoga-Raum nebenan: eine Frau auf einer Matte mit einem Tablet, der Riese am Fenster, die Wand des Trading-Raums auf der linken Seite.

Gleiche Fenstergröße, Bodenhöhe und Gesichtsskalierung, ein Raum weiter rechts.

Bilder, Runde 1: 33 Bilder in einem ChatGPT-Chat

Der erste Versuch war ein gewöhnliches ChatGPT-Gespräch in einem Plus-Plan: Beschreiben Sie den Raum, fügen Sie ein Foto meines Gesichts bei, bitten Sie um Änderungen. Ich benötigte 26 Nachrichten und es wurden 33 Bilder erzeugt, wobei etwa 18 meiner Nachrichten Korrekturen waren. Jede Antwort baute die Szene von Grund auf neu auf, sodass die Korrektur einer Sache eine andere beschädigte.

Ein leerer Raum mit einem Fenster und einem Schreibtisch, ohne Monitore und ohne Person.
Keine Monitore, kein Trader.
Ein unordentlicher Trading-Raum bei Nacht mit Stadtblick und ohne Riesen am Fenster.
Zu unordentlich, und der Riese fehlt.
Der Yoga-Raum, wobei der Händler aus dem vorherigen Raum links noch sichtbar ist.
Der Trader aus dem alten Raum ist noch im Bild.
Ein Porträt eines blonden Mannes mit einer im Vergleich zur Referenz deutlich anderen Kopfform.
„Eine völlig andere Person.“
Das Gesicht des Riesen verteilt sich auf zwei Fenster, als wären es zwei verschiedene Personen.
Ich, aufgeteilt auf zwei Fenster.
Der Yoga-Raum mit einem Fenster, das sichtbar kleiner als das Referenzbild ist.
Fenster bei 78% der angeforderten Größe.

Gegen Ende habe ich die exakte Pixelgeometrie für das Fenster und beide Personen angegeben und den Chat gebeten, sein eigenes Ergebnis zu messen. Er meldete ein Fenster von 485 × 644 px im Vergleich zu den 621 × 811 px im Prompt: 136 px zu schmal und 167 px zu kurz. Ein Modell, das ein Bild aus Worten neu zeichnet, kann keine exakten Pixelwerte treffen, egal wie präzise die Worte sind.

Bilder, Runde 2: 11 Bilder, meist beim ersten Versuch

Die zweite Runde startete vom besten Bild aus Runde 1 und lief in Codex verbunden mit der easyai.video-Plattform, mit unseren installierten Bildtechniken. Acht Anfragen erzeugten elf Bilder. Sechs der acht Anfragen waren neue Änderungen, die jeweils beim ersten Versuch genehmigt wurden; die anderen zwei waren Korrekturen für die Wandposition und die Bodenhöhe.

Das Bild des Trader-Raums aus Runde 1, mit einer leuchtenden Decke und einem unerklärlichen Licht auf dem Autoposter.
Start: Das beste Bild aus Runde 1.
Derselbe Raum, neu beleuchtet durch die Deckenpendelleuchte, mit einem schärferen Gesicht am Fenster.
1 · eine Lichtquelle, schärferes Gesicht.
Derselbe Raum, wobei der Riese den Blick nach unten zum Händler senkt.
2 · Fokus auf den Trader.
Derselbe Raum, der direkt hinter einer dunklen Trennwand liegt, wobei der Riese nur am Fensterrand zu sehen ist.
3 · die Wand zwischen den Räumen.
Der Yoga-Raum mit der gleichen Geometrie und dem Riesen am Fenster.
4 · der Yoga-Raum.

Zwischen den Runden hat sich nichts Magisches geändert. Die Art der Fragestellung hingegen schon:

  1. Bearbeiten, nicht neu generieren. Jede Anfrage basierte auf dem zuletzt genehmigten Bild und behielt alles bei, was bereits korrekt war.
  2. Eine Änderung pro Anfrage. Licht, dann das Gesicht, dann der Blick, dann das Layout. Niemals vier Korrekturen in einer Nachricht.
  3. Ein Auftrag pro Eingabebild. „Bild 1 ist das maßgebliche Ziel für die Bearbeitung. Bild 2 liefert nur den linken Streifen.“
  4. Messen, dann verankern. Anstatt „mache den Raum schmaler“ zu sagen, gab der Prompt an, wo sich die Dinge in der Quelle befanden und wohin sie verschoben werden mussten: Übersetze den Raum um 275 px nach rechts bei 100% Skalierung, der Fensterquerschnitt bleibt bei y = 626.
  5. Nutzen Sie eine Technik, keine Improvisation. Die Bildtechniken gaben den Anwendungsfall an (Compositing, Identitätserhalt) und schrieben den Prompt jedes Mal in derselben Struktur.

Sie können dasselbe in Ihrem eigenen Setup tun. Codex und Claude Code laden beide Skills: einen Ordner mit Anweisungen und optional Skripte, die der Agent vor der Ausführung liest. Eine breite Palette an Bild- und Videoskills ist auf GitHub und anderen öffentlichen Ressourcen verfügbar. Wählen Sie Skills aus, die zu Ihrem Modell und Ihrer Schnittaufgabe passen, und nutzen Sie diese, um Ihre Prompts automatisch zu verbessern. Die easyai.video-Plattform enthält diese Skills sowie die Prompt-Formate, die das jeweilige Videomodell erwartet.

Video, Take 1: Die gesamte Szene als eine 8-sekündige Einstellung

Nachdem die drei Bilder genehmigt wurden, verlangten die ersten Video-Versuche alles in einer Generierung: essen, bemerken, rufen, werfen, sich entschuldigen, in den nächsten Raum gleiten. Für ein Experiment begann ich mit kostenlosen Google Flow-Generierungen. Google Flow erhielt die drei Bilder als Referenzen; Take 2 nutzte einen umgeschriebenen Prompt, der den Sprecher jeder Zeile benannte und die Tasse zum hinteren Fenster verfolgte. Seedance 2.5 erhielt die Bilder als Startframe, Bildreferenz und Endframe in seinem Omni-Referenz-Modus.

Take 1

Google Flow

Beginnt in unserem Frame, dann sagt der Riese beide Zeilen, bevor etwas geworfen wird. Der Becher fliegt auf die Kamera, nicht zum Fenster, und der Händler behält seinen eigenen Becher. Der Yoga-Raum blendet ein, anstatt hineinzuschieben.

Take 2

Google Flow

Unser erster Frame wird ignoriert: ein Torbogen und ein hellerer Raum. Zwei Riesen erscheinen, dann tritt ein Mann in Originalgröße ein und wirft den Becher. Niemand sagt „Sorry“. Es endet in einem anderen Yoga-Raum.

Take 3

Seedance 2.5 · 480p Entwurf

Die nächste der drei: Der Becher fliegt von uns weg, prallt gegen die Rückscheibe und die Kamera gleitet in den Yoga-Raum. Gerendert als einfacher 480p-Entwurf.

Die Bilder waren nicht das Problem. Die Einstellung verlangte zu viele Dinge gleichzeitig, woraufhin das Modell einige wegließ und andere umordnete.

Video, Take 2: Zwei kurze Einstellungen durch einen gemeinsamen Frame

Die Lösung bestand darin, die Szene dort aufzuteilen, wo sie natürlich bricht. Einstellung A läuft vom Startframe bis zum mittleren Frame in 3 Sekunden: Der Trader isst und die Riesenfigur ragt in das Fenster hinein. Einstellung B läuft von diesem mittleren Frame bis zum Endframe in 6 Sekunden: Die Linie, der Wurf, die Entschuldigung und der Wechsel in den Yogaraum. Die beiden Einstellungen teilen sich einen Frame, was ihnen beim Schnitt eine gemeinsame Komposition verleiht. In dieser Version wirft der Trader seine Kopfhörer statt des Nudelbechers: Die über das Fenster gespritzten Nudeln waren unangenehm anzusehen, und Seedance stellte sie viel zu groß dar.

Beide Einstellungen wurden mit Wan 3.0 auf Higgsfield erstellt, in nativen 1080 × 1920 bei 30 fps mit generiertem Sound.

Einstellung A · 3 s

Wan 3.0

Der Händler isst weiter und schaut nie auf; der Riese gleitet hinein und positioniert sich im mittleren Frame.

Einstellung B · 6 s

Wan 3.0

„Lass mich in Ruhe!“, die Kopfhörer treffen auf das Glas und fallen, „Entschuldigung.“, dann der Übergang in den Yogaraum, endend auf dem letzten Frame. Am Ende wurden einige Musiknoten hinzugefügt, was der Prompt eigentlich ausschloss.

A + B · 9 s

Verbunden am gemeinsamen Frame

Das Gesicht hat im zweiten Teil eine leicht andere Hauttextur, weshalb der Schnitt sichtbar ist. Ein Modell, das eine einzige ununterbrochene Einstellung erzeugen kann, wie etwa Seedance, ist für die Kontinuität vorzuziehen.

Dieselbe Einstellung auf MiniMax H3: Full-Modell vs. Turbo

Wir haben Einstellung B anschließend erneut auf MiniMax H3 gerendert, mit unserer eigenen Kopie des Modells innerhalb des täglichen GPU-Kontingents von Hugging Face, mit denselben zwei Frames, derselben Geschichte und demselben Seed (42). Die easyai.video-Plattform schreibt den Prompt zuerst in das Format um, das H3 erwartet; der folgende Prompt ist das, was nach dieser Umformulierung in das Modell eingegeben wurde.

Einstellung B · Wan 3.0

1080 × 1920 · 30 fps

Die Referenz: jeder Beat in der richtigen Reihenfolge, endend beim letzten Frame.

Einstellung B · H3 vollständig

768 × 1344 · 24 fps · 28 Schritte

Vergleichbar mit Wan 3.0 in dieser Einstellung: Die Linie, der Wurf, die Entschuldigung und die Bewegung sitzen alle. Niedrigere Auflösung; verbraucht einen Teil des täglichen GPU-Kontingents.

Einstellung B · H3 Turbo

768 × 1344 · 24 fps · 6 Schritte

Erfüllt die erzählerischen Punkte, aber das große Gesicht, die Kopfhörer und der Übergang zwischen den Räumen verraten die KI.

Das Gesicht des Riesen im selben Moment: H3 Full links und H3 Turbo rechts; die Turbo-Version weist dunkle, fleckige Stellen auf der Haut auf.
Im selben Moment, 1.6 s in Einstellung B, in voller Größe. Links: H3 voll. Rechts: H3 turbo, mit dunklen Flecken auf der Haut.
H3 Turbo-Frame, in dem die geworfenen Kopfhörer als flaches schwarzes Icon auf der Nase des Riesen erscheinen.
Turbo, 2.0 s: Die Kopfhörer wirken wie ein flaches Icon, das auf das Gesicht geklebt wurde, nicht wie ein Objekt im Flug.
H3 Turbo-Frame während der Bewegung zwischen den Räumen, mit Blättern im Vordergrund, die niemand angefordert hat.
Turbo, 4.3 s: Es wurde niemand gefragt, warum während der Bewegung ein Drift durch den Vordergrund entsteht.

Turbo ist für Entwürfe gedacht. Es eignet sich gut, um eine Kamerabewegung zu planen oder zu testen, ob die Frames funktionieren. Wenn ein Gesicht das Bild ausfüllt, rendern Sie den Take mit dem vollen Modell.

Kostenübersicht pro Modell

Rote Kreuze markieren inakzeptable Takes. Gelbe Häkchen markieren Takes, die dem Ziel nahekommen, aber noch Unstimmigkeiten aufweisen; bei Seedance betrifft dies beispielsweise überdimensionierte Nudeln. Jede Tabellenzeile deckt ein Modell und eine Plattform ab, wobei die Gesamtkosten des Materials aufgeführt sind.

Modell · PlattformMaterialAusgabeKostenbasisca. EUR
Google Flow2 × 8 s720 × 1280Kostenlose tägliche Credits€0
Seedance 2.5 Entwurf · Higgsfield8 s480 × 85424 Credits€1.05
Wan 3.0 · Higgsfield3 s + 6 s1080 × 1920 · 30 fps31.5 Credits (3.5 pro Sekunde)€1.38
MiniMax H3 voll · Hugging Face6.58 s768 × 1344 · 24 fps≈ 10.9 / 40 täglich GPU Minuten (27.3%)€0.073
MiniMax H3 Turbo · Hugging Face6.58 s768 × 1344 · 24 fps≈ 4.2 / 40 täglich GPU Minuten (10.5%)€0.028

Higgsfield Schätzungen basieren auf dem gebuchten $49 / 1,000-Credit-Plan, wobei alle Credits aufgebraucht sind. Hugging Face Schätzungen verteilen das $9 Monatsabonnement auf 30 Tage und multiplizieren diesen Tagesanteil dann mit dem Quotenprozentsatz in den Run-Logs, einschließlich Conditioning. Dies sind zugewiesene Abonnementkosten, keine Rechnungen pro Run; eine nicht genutzte Quote erhöht die effektiven Kosten.

Konvertierung: $1 ≈ €0.893, gerundet nach dem EZB-Referenzkurs. Pläne und Kontingente können sich ändern; prüfen Sie Higgsfield-Preise und Hugging Face PRO für Ihr Konto.

Was man daraus lernt

  1. Entwerfen Sie zuerst die Schlüsselbilder und animieren Sie dann zwischen ihnen. Gemeinsame Frames helfen dabei, eine Szene mit mehreren Einstellungen auszurichten, aber überprüfen Sie Hauttextur und Beleuchtung beim Schnitt.
  2. Bearbeiten Sie das genehmigte Bild, anstatt erneut anzufragen. Eine Änderung pro Anfrage, ein Auftrag pro Eingabebild, reale Positionen statt Adjektive.
  3. Turbo-Modelle sind für Entwürfe gedacht. Rendern Sie die finalen Takes mit dem vollen Modell, wenn ein Gesicht groß im Bild zu sehen ist.
  4. Ein Open-Source-Modell kann bei einer bestimmten Einstellung mit einem kostenpflichtigen Modell mithalten. H3 im Vergleich zu Wan 3.0 hier. Das Bereitstellen eines OSS-Modells auf einer GPU kann je nach Hardwarenutzung und Hosting-Kosten ein ähnliches Ergebnis bei deutlich geringeren Kosten als ein kostenpflichtiges Modell über eine API liefern.
  5. Nutzen Sie Fähigkeiten, um Bild- und Videoprompts automatisch zu verbessern. Spezielle Anweisungen übersetzen eine kurze Anfrage in die Struktur und die Einschränkungen, die jedes Modell benötigt.

Runde 1 · ChatGPT Chat

Einer der Prompts mit exakter Pixelgeometrie

Use gpt_mainref_shot1.png as the strict geometry and scale reference. Recreate the same dollhouse room proportions, camera position, perspective, window size, giant-person scale, and miniature-person scale. Change only the room content from the trader bedroom into a yoga/living room.
Final canvas: exactly 941 × 1672 px, vertical.
Do not change the camera position, focal length, room depth, wall proportions, floor perspective, or window proportions from gpt_mainref_shot1.png.
Main window geometry must match gpt_mainref_shot1.png exactly:
outer window frame: x=169 to 789, y=197 to 1007
outer window size: 621 × 811 px
visible glass area: x=193 to 767, y=230 to 988
glass size: 575 × 759 px
horizontal window divider: x=181 to 779, y=620 to 637
window center approximately x=479
the window must occupy about 66% of the entire image width
do not make the window smaller
do not move the bottom edge of the window upward
Behind the window is one giant adult man. Use the supplied male portrait reference for identity and pose. Preserve the same face proportions, head shape, nose, jaw, forehead, hairline, eye spacing, skin texture, light blond hair, stubble, light-blue T-shirt, and body proportions.
The giant man must have the same scale as the person in gpt_mainref_shot1.png:
head from top of hair to bottom of chin approximately x=292 to 710, y=228 to 824
head size approximately 419 × 597 px
visible shoulders and upper torso extend across almost the entire 575 px width of the window
the giant person must visually fill the window, not appear distant
his head and shoulders must remain much larger than the miniature person inside the room
Pose of giant man:
body turned approximately 25 to 30 degrees toward screen-right
shoulders turned by the same angle as the head
head slightly turned toward screen-right
eyes looking downward and inward into the room at the woman
expression curious and observant
do not make him look toward the camera
do not reduce his face size
do not shorten or round his head
do not create a second copy of him anywhere
Room geometry must preserve the same apparent size as gpt_mainref_shot1.png. The floor, ceiling, left wall, right wall, and back wall must occupy the same proportions. The camera remains outside the dollhouse, looking slightly downward into the room from the same elevated frontal angle.
On the extreme left, include a narrow slice of the previous room followed by a full-height dark dividing wall. The dividing wall must run continuously from the top edge to the bottom edge of the image and clearly separate the two rooms. The previous room must occupy only a very narrow strip and must not contain any visible person.
Replace the trader bedroom with a bright, realistic yoga/living room.
Inside the new room:
one young adult woman doing yoga on a mat
one tablet standing in front of her, showing a yoga instruction video
one comfortable chair
a soft dress and light fabric casually draped over the chair
one small dressing table on the right side with a few cosmetic bottles
a vase with flowers
minimal tasteful plants
no bed
no desk with computers
no paper
no garbage bin
no cluttered trash
The miniature woman must use approximately the same scale as the trader in gpt_mainref_shot1.png.
Target miniature-person scale:
overall body region approximately 265 × 438 px
top of head around y=925 to 950
lowest visible body point around y=1360
do not place her higher in the frame
do not enlarge her beyond the trader’s scale
she must remain clearly miniature compared with the giant man outside
Place the yoga woman lower in the composition, around the lower central area of the room. Her head should begin approximately where the trader’s head begins in the reference. The majority of her body must stay below the bottom edge of the window.
Yoga mat:
place it low in the frame
approximately x=185 to 710
approximately y=1180 to 1510
roughly 525 × 330 px
keep perspective aligned with the floor exactly like the rug area in gpt_mainref_shot1.png
The woman is seated in a clear seated spinal twist yoga pose. Her torso remains upright. One hand supports her on the mat. Her legs are folded. The tablet is positioned in front of the mat near the lower-left edge of the mat, angled toward her.
Lighting:
natural daylight only, entering through the giant window
neutral soft daylight
ceiling lamp present but switched off
no warm yellow glow
no wall lamps
no hidden artificial lighting
no cinematic orange lighting
room illumination must come primarily from the window
Keep the room realistic and naturally lived-in, but clean.
Critical constraints:
window size must remain exactly equivalent to gpt_mainref_shot1.png
giant man scale must remain exactly equivalent to gpt_mainref_shot1.png
miniature woman scale must remain equivalent to the trader scale in gpt_mainref_shot1.png
woman must be positioned low in the composition
camera must not zoom out
camera must not move backward
do not increase visible wall space around the window
do not shrink the window
do not shrink the giant face
do not create duplicated body parts or duplicated people
do not show any part of the giant man inside the narrow previous-room strip on the far left
preserve one continuous architectural dividing wall from ceiling to floor
photorealistic, realistic lens perspective, consistent scale, high detail, 9:16 composition.

Runde 2 · Codex mit Bildtechniken

Die funktionierende Korrektur der Geometrie

Use case: compositing. Generate one corrected photorealistic 9:16 dollhouse room-reveal keyframe, 941x1672 composition. Image 1 (mf1_sharp_reference_face) is the authoritative edit target for exact room geometry, object sizes, vertical positions, furnishings, person, face identity, textures and light. Image 2 (mf1_partition_reveal) supplies ONLY the left foreground strip and the giant man's partial-face placement. Image 3 (pasted plate) demonstrates the intended right-side cropping of the original room, keeping its proportions.

GEOMETRY IS THE PRIMARY EDIT. The left foreground strip in image 2 measures approximately 275 pixels of a 941-pixel width: neighboring-room glimpse at x=0..159, dark vertical dividing wall at x=160..274. Preserve this strip in the corrected frame. The remaining x=275..940 must show the LEFTMOST 666 pixels of image 1 at precisely its ORIGINAL SCALE and ORIGINAL HEIGHT. Translate image 1's room scene 275 pixels to the RIGHT and allow its rightmost 275 pixels to fall beyond the right border. This is a lateral camera reveal across a continuous dollhouse set. Keep the original room width and object proportions, expressed through partial offscreen framing. Use translation and clipping, with scale fixed at 100 percent. Maintain the source vertical coordinates throughout. Concrete anchors: original pendant center at x about 470 becomes x about 745, staying at y about 150; original window left glazing edge x about 193 becomes x about 468, with window top staying y about 230 and horizontal crossbar staying y about 626. The original window extends past the right frame border. Original bed left edge x about 593 becomes x about 868, so ONLY a narrow portion of the bed is visible at the extreme right. The original trader, chair and three monitors keep their exact source sizes and shift right by 275 pixels. The car poster remains its original size and shifts right identically. The original right wall artwork is beyond the frame edge. The rug and floor keep their original scale, pattern and vertical positions. The foreground strip aligns naturally with the room's open front.

FACE CHANGE: keep the same sharp blond man's identity, realistic skin, hair and original head scale from image 1, but move him laterally OUTSIDE the window to its left, as in image 2. Just a narrow sliver of blond hair, forehead, one eye, cheek and jaw peeks into the glazing along its LEFT edge, approximately x=468..525 in the final composition. The rest of the visible window is blue evening sky. Preserve the window frame and crossbar foreground occlusion. This partial visibility is achieved by the window edge clipping his head; preserve realistic head proportions.

World contract: same seated young man with glasses, black hoodie, dark trousers and white trainers at the same crypto workstation; same wood furniture and warm floor, dark rumpled bedding wherever visible, warm-white downward-facing pendant illumination with physically consistent falloff and shadows, softly dark ceiling, blue evening sky outdoors, crisp photorealistic facial detail. Preserve the source lighting and colors. Deliver exactly one corrected still with the room at full original scale and the right-side framing naturally clipped.

Google Flow · Take 1 · 8 s, drei Referenzbilder

Prompt für Take 1

Create an 8-second photorealistic cinematic video with synchronized English dialogue, in vertical 9:16.

REFERENCE ROLES
Opening: @generated/images/mf1_partition_reveal_corrected.png. Begin on this exact trader-room composition, including the foreground partition, neighboring-room sliver, original room proportions, floor height, tall window, pendant, workstation and cropped bed.
Intermediate appearance reference, around 2 seconds: @generated/images/mf1_looking_at_guy.png. Match the giant observer’s recognizable face, blond hair, blue eyes, realistic skin texture, light-blue T-shirt, head scale and downward gaze toward the seated trader. This guides his appearance during the encounter; camera movement determines the framing.
Ending: @generated/images/lf1_partition_reveal_corrected.png. Land on this exact yoga-room composition and final observer pose.

CAMERA AND GEOGRAPHY
One continuous lateral dolly to camera-right along the open front of a physical dollhouse. The trader room is left of the yoga room. As the camera travels right, the fixed rooms slide left across the image and their dividing wall passes through the foreground. Keep camera height, viewing direction and focal length constant. Preserve the original width and height of each room: furniture and architecture retain their dimensions and leave the image through natural cropping. Maintain a common floor level and consistent miniature scale. The observer remains a single giant man outside the windows throughout.

0.0-2.2 SECONDS
The miniature trader, wearing glasses, a black hoodie, dark trousers and white trainers, eats instant noodles from a cup with a fork while watching his three monitors. The camera glides slowly right from the opening reveal toward a clearer view of the workstation. Outside, the giant blond man leans into the window from its left edge until his face is clearly visible. He tilts his head and looks directly down at the trader, matching the intermediate appearance reference. His expression is curious and mildly intrusive.

2.2-3.4 SECONDS
The trader notices him, pauses with the fork raised, turns his head and upper body toward the window and snaps angrily in English: “Leave me alone!” His mouth movement matches the words. The observer listens.

3.4-4.2 SECONDS
Immediately after finishing the line, the trader throws the noodle cup from his hand toward the observer’s face. The cup travels away from the trader and strikes the INSIDE surface of the window glass. Broth and noodles splatter on the interior glass directly in front of the giant’s face. The glass stays intact and separates the observer from the cup and splatter.

4.2-4.8 SECONDS
The observer flinches, blinks and replies apologetically in English: “Sorry!” His mouth movement matches the word. He then withdraws from the trader’s window and moves physically right outside the building toward the neighboring yoga-room window.

4.8-7.6 SECONDS
The camera continues its rightward dolly, smoothly traversing the foreground dividing wall and revealing the brighter yoga room. The trader room and its splattered window slide out to the left. The woman wears a mauve sports bra and leggings, with dark hair in a bun. She remains seated low on her mat in the reference pose, calmly following the tablet on the floor. Her gaze stays on the tablet. Preserve the cream chair, pink throw, plants, bottle, rug and visible furniture from the ending reference. The house poster and EasyAI artwork remain in the narrow neighboring-room strip left of the partition.

7.6-8.0 SECONDS
The camera gently settles into the ending reference. Match its floor height, window dimensions and placement, woman’s position and pose, furniture scale, partition thickness and framing. The giant has reached the exact final head position, orientation and window-relative scale shown there, wearing the same light-blue T-shirt. He holds that pose while the woman continues focusing on her tablet.

LOOK AND SOUND
Keep the trader room’s warm pendant lighting and the yoga room’s soft natural daylight, with stable exposure and reference-matched colors. Natural facial animation, grounded body motion and believable miniature materials. Maintain continuous spatial movement throughout, with transitions produced by the camera passing the physical wall. Sound consists of quiet computer hum, fork and noodle sounds, the trader’s shout, a cup impact and wet glass splatter, the observer’s brief apology, then calm yoga-room ambience. Only the two specified spoken lines. Ambience and effects form the soundtrack; the image remains free of captions and graphic overlays.

Google Flow · Take 2 · 8 s, drei Referenzbilder

Verwendeter Prompt für Take 2

Create an 8-second photorealistic vertical 9:16 video with synchronized English dialogue.

REFERENCES
First frame: @generated/images/mf1_partition_reveal_corrected.png - opening trader-room composition.
Intermediate appearance reference: @generated/images/mf1_looking_at_guy.png - giant observer’s identity, scale and downward gaze toward the trader.
Last frame: @generated/images/lf1_partition_reveal_corrected.png - final yoga-room composition and observer pose.

CHARACTER AND SPEAKER ASSIGNMENTS
TRADER: the miniature young man INSIDE the bedroom, wearing glasses, a black hoodie, dark trousers and white trainers. He alone says: “Leave me alone!” He throws the noodle cup.
OBSERVER: the giant blond man OUTSIDE the rear window, wearing a light-blue T-shirt. He quietly watches the trader and trading monitors. His sole spoken line is: “Sorry!”

GEOGRAPHY
The camera views the miniature room through its open front. The trader is inside, between the camera and the rear window. The giant stands outside, behind that window. The window glass physically separates them. The cup travels from the trader toward the BACK of the room, away from the camera, and strikes the interior surface of the rear window directly in front of the giant’s face.

0.0-2.5 SECONDS - SILENT OBSERVATION
Begin on the first reference. The trader eats noodles while watching his three monitors. The giant gradually leans into view from the window’s left edge, matching the intermediate reference. His eyes move naturally between the trader and the screens below. His lips remain together, his expression quietly curious, his hands below the window sill.

2.5-3.3 SECONDS - THE TRADER SPEAKS
The trader notices the giant, puts his fork on the desk, and turns his head and shoulders toward the rear window. His mouth is visibly readable in profile as he angrily shouts in English: “Leave me alone!” Use a young male voice originating inside the room. During this entire line, the giant keeps his lips together and listens with a surprised expression.

3.3-4.1 SECONDS - THE TRADER THROWS
The trader transfers the SAME cup he was eating from into his right hand, draws that arm back, then swings toward the rear window and releases it. Show the release and continuous flight clearly. His throwing hand is visibly empty afterward. The separate cup already resting on the desk stays there.

The thrown cup remains miniature, matching the trader’s hand. It recedes through the room toward the rear glass, becoming slightly smaller with distance. It strikes the pane in front of the observer’s face. Broth and noodles splash onto that rear pane and cling to it; the cup falls inside beneath the window. The glass remains intact. The foreground viewing opening and camera view remain clean. The giant stays outside throughout.

4.1-4.6 SECONDS - THE OBSERVER REPLIES
After the impact, the giant gives one startled blink. Only now his lips move as he softly says in English: “Sorry!” Use a distinct mature male voice originating outside the window. The trader’s mouth stays closed during this reply.

4.6-8.0 SECONDS - NEXT ROOM
The giant withdraws and moves right outside the building. The camera continues its smooth lateral dolly right along the dollhouse front. The fixed trader room slides left; the physical dividing wall passes through the foreground, revealing the yoga room. Maintain camera height, focal length, room dimensions, window geometry, shared floor level and furniture scale.

Gently settle into the last reference by 7.6 seconds and preserve that composition through 8.0 seconds. The woman in mauve yoga clothes remains focused on her tablet in the reference pose. The giant reaches the reference’s exact window-relative position and head orientation.

STYLE AND AUDIO
One continuous shot. Preserve each reference room’s lighting, materials and architectural proportions. Natural restrained acting. Sound: computer hum, eating sounds, the trader’s “Leave me alone!”, cup release, rear-window impact, wet splatter, the observer’s “Sorry!”, then quiet yoga-room ambience. The two lines occur separately, with the visible speaker’s lips synchronized to each line. Ambience and effects provide the soundtrack.

Seedance 2.5 · 480p Entwurf · Omni-Referenz, Startbild + Bildreferenz + Endbild

Verwendeter Prompt für diesen Entwurf

8-second contemporary digital comedy scene inside a physical miniature dollhouse. One continuous take: a giant’s quiet curiosity provokes a miniature trader, then the camera reveals the calm neighboring yoga room.

MODEL: Seedance 2.5, omni-reference.
FORMAT: 9:16, 8.0 seconds, photoreal digital cinema.

REFERENCES
@Image1 - generated/images/mf1_partition_reveal_corrected.png, supplied as START FRAME. Preserve its opening composition, architectural proportions, floor height, window geometry, furniture and lighting. Use it for the opening only; the camera subsequently travels right.
@Image2 - generated/images/mf1_looking_at_guy.png, supplied as IMAGE REFERENCE. Take the giant observer’s exact identity, blond hair, blue eyes, light-blue T-shirt, face scale and downward gaze. Use this appearance during the trader encounter; do not insert this still as a separate shot.
@Image3 - generated/images/lf1_partition_reveal_corrected.png, supplied as END FRAME. Match its final composition, yoga-room geometry, woman’s pose, furniture and observer orientation. Its interior belongs only to the neighboring yoga room.

SUBJECTS AND SPEAKERS
TRADER: miniature young man inside the room, glasses, black hoodie, dark trousers, white trainers. He alone says “Leave me alone!” and throws the noodle cup he is holding.
OBSERVER: giant blond man outside the rear window, matching @Image2, wearing a light-blue T-shirt. His only line is “Sorry!”
WOMAN: miniature woman in the yoga room, dark hair in a bun, mauve sports bra and leggings. She remains silent and focused on her tablet.

STYLE: reference-matched lens perspective, deep enough focus to read the trader and observer together, restrained facial acting, natural skin texture and cloth movement. Must feel like a photographed miniature set with a real person outside, not generated footage.
COLOR: warm brown and charcoal in the trader room; cream, muted mauve and oak in the yoga room; blue daylight outside. Stable reference-matched exposure.
ENVIRONMENT: trader room left, yoga room right, connected by a common floor level and separated by a thick structural wall. Warm overhead pendant illuminates the trader; daylight enters through the rear windows. Architecture and furniture remain fixed.

CAMERA: physical lateral dolly to camera-right along the open dollhouse front. Constant height, viewing direction and focal length. Rooms move left across the image through parallax and natural cropping. Motion preset: none. The following numbered beats are phases within the same shot, not cuts.

Shot 1 [0.0s-2.2s] - wide room view from the opening reference: the trader continues eating and watching his three monitors. The observer leans into view until his appearance matches @Image2, quietly looking between the trader and monitors. His lips stay closed and his hands remain below the sill.
Camera: slow rightward dolly.
SFX: faint computer hum and fork contact.

Shot 2 [2.2s-3.3s] - same wide view and angle: the trader puts down his fork and turns toward the rear window while retaining his cup. The miniature trader angrily says, “Leave me alone!” His lips alone articulate these words; the observer listens with his lips closed.
Camera: continue the slow rightward dolly, keeping both faces visible.
SFX: the trader’s young male voice originates inside the room. Realistic lip articulation, no exaggerated mouth opening, no head turns while speaking.

Shot 3 [3.3s-4.1s] - same room view: the trader draws back his cup-holding arm, swings toward the rear window and releases that same miniature cup. Show hand, release and flight as one connected action. His hand is empty afterward. The cup travels away from the camera into the back of the room, becoming smaller with distance, and strikes the interior glass directly in front of the observer’s face. Broth and noodles stick to the rear pane; the cup drops inside below it. The separate cup on the desk stays in place. Glass remains intact; the observer stays outside behind it.
Camera: continue rightward, retaining the throwing hand and rear pane in view.
SFX: cup impact and wet splatter.

Shot 4 [4.1s-4.6s] - same room view: the observer blinks at the impact, then softly says, “Sorry!” Only the giant observer’s lips articulate this word. The trader’s mouth stays closed.
Camera: continue the rightward dolly.
SFX: distinct mature male voice from outside, brief apologetic delivery with realistic lip articulation.

Shot 5 [4.6s-7.5s] - wide view along the same dollhouse front: the observer withdraws from the trader window and moves right outside the building. The camera passes the physical dividing wall, revealing the yoga room. The woman stays seated in her reference pose, eyes on the tablet. The observer arrives at the yoga window and settles into @Image3’s exact head position and orientation.
Camera: continuous rightward translation, gently decelerating into the end-frame composition.
SFX: computer hum fades with distance into quiet room ambience.

Shot 6 [7.5s-8.0s] - held wide composition matching @Image3: preserve floor height, window size and placement, partition thickness, woman’s position, furniture and observer scale. Everyone holds the established ending.
Camera: settled and locked.
SFX: silence holds the final 0.5 seconds.

AUDIO: the two assigned English lines are clear and prominent; synchronized effects are secondary, room ambience subtle.
- No music, no library audio, no voiceover narration, no on-screen text, no subtitles, no added logo.

CONSTRAINTS: one unbroken continuous take. No cuts, dissolves or black frames. No zoom or focal-length change. The trader alone delivers the angry line and throws his own cup. The observer watches silently until his apologetic reply. Preserve each person’s identity and wardrobe. Preserve miniature-to-giant scale, room dimensions and window geometry. The woman’s attention remains on the tablet. Existing wall artwork stays part of the set.
Stable face throughout. No morphing. No deformation. Photoreal cinematic finish.

Wan 3.0 · Einstellung A · Startframe bis Mittelframe, 3 s

Verwendeter Prompt für Einstellung A

3-second contemporary digital comedy shot inside a physical miniature dollhouse. One continuous take: a miniature trader eats noodles at his desk while a giant, unnoticed, slowly rises into view behind his rear window.

FORMAT: 9:16, 3.0 seconds, photoreal digital cinema with native sound.

START AND END: open exactly on the supplied start frame (trader room; trader holding his noodle cup near his mouth; only a sliver of the giant observer's face visible at the far left edge of the rear window) and end exactly on the supplied end frame (same room; the observer's whole face filling the rear window, looking down at the trader). Keep both frames' composition, proportions, floor height, window geometry, furniture, desk items and lighting.

TRADER: miniature young man at the desk, glasses, black hoodie, dark trousers, white trainers. He eats continuously for the whole shot: cup held near his mouth in one hand, lifting forkfuls of noodles, chewing, eyes on his three monitors. He never notices the observer, never turns, never speaks. Black over-ear headphones lie untouched on the desk beside the keyboard.
OBSERVER: giant blond man, blue eyes, light-blue T-shirt, outside the rear window; exactly the man in the end frame. He moves from almost out of view at the left edge across and into the window, curious, settling into the end frame's head position, angle and downward gaze. Lips closed, silent, hands below the sill.

STYLE: frame-matched lens perspective, deep focus holding trader and observer together, restrained acting, natural skin and cloth. A photographed miniature set with a real person outside, not generated footage.
COLOR: warm brown and charcoal room under a warm pendant; blue daylight outside the window.
ENVIRONMENT: the trader room only: desk with three monitors, keyboard, headphones, bed, rug, wall artwork. Architecture and furniture stay fixed.

CAMERA: very slight physical dolly to camera-right along the open dollhouse front, matching the small shift between the start and end frames; constant height, direction and focal length. No zoom.

[0.0s-2.5s] The trader keeps eating - forkfuls lifted and chewed, eyes on the monitors. The observer glides in from the left edge of the rear window, his face rising and sliding across until it fills the window, quietly watching the trader eat. SFX: faint computer hum, fork against the cup, soft slurping, chewing.

[2.5s-3.0s] Held composition matching the end frame: the observer's face filling the rear window, eyes on the trader; the trader still eating, unaware. Camera settled.

AUDIO: eating sounds and computer hum only; subtle room ambience. No dialogue, no music, no narration, no on-screen text, no subtitles, no logo.

CONSTRAINTS: one unbroken take; no cuts, dissolves, black frames or zoom. Nobody speaks. The trader keeps eating and never looks at the window. The headphones stay on the desk. Preserve identities, wardrobe, miniature-to-giant scale, room dimensions, window geometry and desk items. Stable faces throughout, no morphing, no deformation, photoreal cinematic finish.

Wan 3.0 · Einstellung B · mittlerer Frame bis Endframe, 6 s

Prompt für Einstellung B

contemporary digital comedy shot inside a physical miniature dollhouse. One continuous take: a miniature trader, interrupted while eating, snaps at the giant watching him through his window and throws his headphones at him; the giant apologizes and moves on, and the camera reveals the calm neighboring yoga room.
FORMAT: photoreal digital cinema with native sound.
START AND END: open exactly on the supplied start frame @[Image 1](image_1) (trader room; trader eating; the giant observer's whole face filling the rear window, looking down at him) and end exactly on the supplied end frame @[Image 2](image_2) (yoga room; observer full-face at its window). Keep both frames' composition, proportions, floor height, window geometry, furniture and lighting.
TRADER: miniature young man at the desk, glasses, black hoodie, dark trousers, white trainers, noodle cup in hand. Black over-ear headphones lie on the desk beside the keyboard, at his left. He alone says "Leave me alone!" and throws those headphones.
OBSERVER: giant blond man, blue eyes, light-blue T-shirt, outside the rear windows; exactly the man in both frames. His only line is "Sorry!"
WOMAN: miniature woman in the yoga room, dark hair in a bun, mauve sports bra and leggings; silent, focused on her tablet.
STYLE: frame-matched lens perspective, deep focus holding trader and observer together, restrained acting, natural skin and cloth. A photographed miniature set with a real person outside, not generated footage.
COLOR: warm brown and charcoal trader room; cream, muted mauve and oak yoga room; blue daylight outside.
ENVIRONMENT: trader room left, yoga room right, one floor level, a thick structural wall between them. Warm pendant over the trader; daylight through the rear windows. Architecture and furniture stay fixed.
CAMERA: physical lateral dolly to camera-right along the open dollhouse front; constant height, direction and focal length; rooms slide left through parallax. The beats below are phases of one shot, not cuts.
[0.0s-1.4s] The trader stops eating, sets the cup down on the desk, turns to the rear window and angrily says "Leave me alone!" Only his lips move; the observer listens, lips closed. Slow rightward dolly, both faces visible. Young male voice from inside the room, realistic lip sync.
[1.4s-2.3s] The trader snatches the black headphones from beside the keyboard and throws them hard at the rear window, straight at the observer's face - grab, swing, release and flight as one connected action; his hand is empty afterward. The headphones fly away from camera, shrinking with distance, hit the inside of the glass in front of the observer's face and drop to the floor below the window. The cup stays on the desk; the glass stays intact; the observer stays outside. SFX: hard plastic thud on glass, headphones clattering on the floor.
[2.3s-2.9s] The observer blinks at the impact and softly says "Sorry!" Only his lips move; the trader's mouth stays closed. Mature male voice from outside, realistic lip sync.
[2.9s-5.5s] The observer withdraws and moves right outside the building. The camera passes the dividing wall, revealing the yoga room; the woman stays seated, eyes on the tablet. The observer arrives at the yoga window and settles into the end frame's exact head position. Continuous rightward translation, decelerating into the end-frame composition. SFX: computer hum fades into quiet room ambience.
[5.5s-6.0s] Held composition matching the end frame - floor height, window placement, wall thickness, the woman's position, observer scale. Camera settled. Silence holds the final 0.5 seconds.
AUDIO: the two English lines clear and prominent; effects secondary; subtle ambience. No music, no narration, no on-screen text, no subtitles, no logo.
CONSTRAINTS: one unbroken take; no cuts, dissolves, black frames or zoom. The trader alone speaks the angry line and throws the headphones; the noodle cup is never thrown. The observer is silent until "Sorry!". Preserve identities, wardrobe, miniature-to-giant scale, room dimensions and window geometry. Existing wall artwork stays. Stable faces throughout, no morphing, no deformation, photoreal cinematic finish.

MiniMax H3 · Einstellung B · gleiche Frames, Seed 42

Prompt an beide H3-Durchläufe gesendet

How the reference pictures align with the target video - Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 1) aligns with the 6.58-second mark of the target video.

integrated_multimodal_description: [Shot 1] Photoreal digital cinema, cinematic, a single continuous lateral dolly shot from left to right along the front of a miniature dollhouse, maintaining constant height, direction, and focal length. The scene opens exactly on <Picture 1>: a miniature trader in a warm brown and charcoal room, seated at a desk with three monitors displaying financial charts, a noodle cup in hand, black over-ear headphones resting on the desk to his left. He wears glasses, a black hoodie, dark trousers, and white trainers. Outside the rear window, a giant blond man in a light-blue T-shirt fills the frame, his face visible from forehead to chin, looking down at the trader. The room is lit by a warm pendant light above the desk; daylight streams through the rear windows. The camera begins a slow rightward dolly. At 00:01.400, the trader stops eating, sets the cup down, turns toward the rear window, and with a tight expression, says: <d>[English] Leave me alone!</d> His lips move in realistic sync; the observer’s lips remain closed. From 00:01.400 to 00:02.300, the trader snatches the headphones from the desk, swings his arm forward, and throws them hard at the rear window - a single connected motion. The headphones fly away from the camera, shrinking with distance, strike the inside of the glass directly in front of the observer’s face, and clatter to the floor below the window. The cup remains on the desk; the glass remains intact; the observer remains outside. SFX: hard plastic thud on glass, headphones clattering on floor. From 00:02.300 to 00:02.900, the observer blinks at the impact and softly says: <d>[English] Sorry!</d> His lips move in realistic sync; the trader’s mouth remains closed. From 00:02.900 to 00:05.500, the observer withdraws and moves right outside the building, the camera continuing its rightward dolly, passing the thick structural wall between rooms. The yoga room is revealed: cream walls, muted mauve accents, oak flooring, a woman in a mauve sports bra and leggings seated on a mat, focused on a tablet. The observer arrives at the yoga window and settles into the exact head position shown in <Picture 2>. The camera decelerates into the end-frame composition. From 00:05.500 to 00:06.000, the composition holds exactly as in <Picture 2>: floor height, window placement, wall thickness, the woman’s position, observer scale. The camera is settled. The final 0.5 seconds are silent.

overall_soundscape: The trader’s noodle cup clinks softly as he sets it down. A hard plastic thud and clatter as headphones strike the window and fall. The observer’s voice is clear and soft. Computer hum fades into quiet room ambience as the camera moves into the yoga room. The woman’s breath is shallow and steady; no other sounds.

non_diegetic_music: N/A

Erstellen Sie eine Einstellung wie diese selbst

AI Video Starter ist ein dreistündiger Live-Workshop: Idee, Einstellungsplan, Referenzbilder, Generierung, Korrektur misslungener Takes und Montage - an Ihrem eigenen Kurzvideo.

Zu den Workshops ↗