撞擊看起來可能很有說服力。但片刻之後,場景仍然合理嗎? 評估 Genie 4 的實用方法,是觀看一次清楚的碰撞,然後檢查接觸後發生什麼事。這份 Genie 4 碰撞與物件移動物理模擬指南,提供一套可重複的觀察流程,用於檢視物件身分、方向變化與場景連續性。它也說明生成視圖能顯示什麼,以及無法證明什麼關於質量、速度或物理定律。
重點摘要
- Genie 4 是互動式生成世界模型,不是會公開狀態變數的傳統物理引擎。
- 固定場景與單一碰撞,較容易判斷接觸、撞擊後運動與物件連續性。
- 有說服力的片段無法確立數值準確度,也不足以用於工程與安全測試。

Art Deco 門廳案例的參考影格:舞者、球、圓柱與拱門在動作開始前。
Genie 4 中的「物理模擬」是什麼意思
傳統物理引擎會根據設定的規則,更新位置、速度、碰撞幾何與材質屬性等明確量值。世界模型則是生成互動式場景,並回應使用者動作或指令。它的輸出或許能描繪因果關係,但可見結果不會自動揭露該描繪背後的數值或規則。
Google 將 Genie 4 描述為一種世界模型,能生成物件碰撞、水流動與樹木搖曳的環境。它也描述可透過自然語言進行即時變更,例如更換攝影機或新增物件。這使該模型很適合用來探索場景並追問接下來會發生什麼。但這本身不會讓每一次描繪出的碰撞都成為經量測的物理結果(Google DeepMind,Genie 4)。
對開發者而言,這項區別很實際。你可以使用互動式生成世界來探索某個機制、找出視覺方向,或提示後續事件。傳統引擎仍是調整恢復係數、摩擦、質量與碰撞形狀的正確場所,因為在傳統引擎中,這些都是明確且可檢查的輸入。除非系統公開這些數值與驗證方法,否則應把「逼真物理」視為視覺描述,而非基準測試宣稱。
如何設計 Genie 4 碰撞測試
從一個簡單到觀看者能判斷是否發生的事件開始。一顆球撞上靜止方塊,比繁忙場景中一堆移動物體更容易評估。保持攝影機穩定、兩個物件都可見,並讓地板保持單純。給每個物件獨特的顏色與形狀,這樣一來,身分變化就很容易看出。
這就是 Genie 4 碰撞與物件移動物理模擬測試的核心:一個簡單事件,以及可在多次執行間比較的後續發展。
Google 的 Project Genie 提示詞指南建議以直接語言描述環境,以及角色或物件動作。該指南是針對 Project Genie,並非已發布的 Genie 4 提示詞語法。它仍是清楚測試簡報的合理起點:指定場景、物件、初始狀態、事件,以及你想檢查的時間點(Google,Project Genie 提示詞技巧,2026 年 3 月)。
一個建議的測試提示詞可寫成:
plaintext1In a plain studio with a flat floor, show one red ball rolling from left to right toward one stationary blue block. Keep both objects fully visible from a fixed side view. The ball contacts the center of the block. Continue the same view after contact so the ball and block remain visible. Keep their colors and shapes consistent. No extra objects or scene changes.
這是一份測試規格,不是宣稱 Genie 4 接受這些確切控制項,或會產生特定結果。在可用工作階段中進行測試時,請儲存措辭、可見控制項與螢幕錄影。如果系統提供即時操控,請記錄每一項指令以及你輸入的時間。這些細節有助於區分模型行為與提示詞變更或攝影機移動。
每次執行都使用碰撞觀察卡:
| 觀察 | 記錄 |
|---|---|
| 起始狀態 | 表面、物件位置、初始運動與攝影機角度 |
| 接觸 | 接觸是否可見,以及撞擊時是否有任一物件被遮擋 |
| 立即反應 | 可見的方向變化、反彈、滾動、傾倒、滑動或停止 |
| 連續性 | 顏色、形狀、數量與相對位置是否保持可辨識 |
| 重複執行 | 在相同提示詞下,哪些保持相似、哪些改變 |
一次只改變一個條件。例如,先讓方塊保持靜止,再以方塊已在移動的狀態執行另一項測試。如果你同時改變了攝影機、表面或措辭,就不要把那些結果當成隔離單一變數來比較。當每個變化都有明確理由時,小型測試集會更有資訊量。
如何解讀撞擊後的物件移動
將接觸前一刻、接觸本身,以及接觸後的幾瞬間,分別作為獨立觀察來檢視。一顆球明顯碰到牆並改變方向,提供了場景描繪碰撞反應的證據。一顆球穿過牆、改變顏色、消失,或在別處重新出現,則引發連續性問題。先記錄螢幕上出現什麼,再解釋為什麼發生。
這張觀察卡刻意採質性描述。「球在接觸後向左移動」是可見的描述。「球守恒動量」則是關於數量的宣稱,而單靠算繪片段可能無法提供這些數量。若要支持該宣稱,你需要可靠的相關狀態量測、明確定義的單位與條件,以及可重複並分析事件的方法。透視畫面中的表面壓縮、反彈高度或行進距離,都可能造成誤導。
如果可取得權限,請將相同測試重複數次。多次執行間的一致性可告訴你,該提示詞是否傾向產生相似場景。它仍無法確立單次執行遵循物理定律:重複輸出可能共享視覺模式,卻未揭露底層計算。請分別記錄事件可辨識度、物件連續性與可重複性。這些是有用的原型觀察,不是合併成一個「物理分數」。
若要與引擎公平比較,請保持情境描述不變,並寫下每個系統實際公開什麼。引擎可能提供座標與速度;生成世界可能提供即時視覺互動。請依據你需要回答的問題來比較輸出,而不是把影片與數值模擬當成可互換的證據。
Genie 4 能證明與不能證明什麼
Genie 4 可作為建立與操控互動式場景的方式來探索。官方資料展示了生成環境,並描述直接產生複雜事件,同時也列出在一致性與互動方面的研究限制。那些示範確立了產品的預期能力,而非經量測的碰撞準確度基準。官方頁面描述當前狀態與限制;當存取權或能力對你的工作很重要時,請再次查看。
官方頁面也指出,複雜事件與物件持續存在仍是研究挑戰。這與碰撞測試直接相關:乾淨的撞擊畫面之後,物件仍可能改變形狀、消失或失去先前位置。在把連續性視為可靠行為之前,請檢查後續發展並重複執行。產品示範說明可能結果,但若要支持準確度宣稱,仍需獨立量測。
這道界線暗示了一種有用的分工:
- **早期遊戲或電影構思:**探索碰撞可能看起來如何,或互動式場景可能如何展開。
- **機制實作:**使用遊戲引擎的碰撞系統,並檢查其參數與狀態。
- **科學或工程分析:**使用適合該問題的經驗證模擬器或量測設定。
- **機器人安全或部署:**不要依賴生成的視覺場景作為唯一驗證方法。
世界模型或許能協助團隊發現有前景的情境或傳達構想。它不能取代必須計算或驗證行為的系統。與隊友分享時,請將生成素材標示為視覺原型,這樣就沒有人會把說明性結果誤認為經驗證的模擬結果。
一個視覺案例:Art Deco 門廳碰撞
視覺參考發生在空蕩的 Art Deco 劇院門廳,而不是花園。午後光線穿過幾何彩繪玻璃,落在玉色與奶油色水磨石上;靛藍絲絨窗簾與黃銅立柱框出清楚的地板路徑。一名舞者身穿短版酒紅絲絨外套、象牙色絲質襯衫、深李子色寬褲與奶油色皮革舞鞋,將一顆珊瑚色橡膠球掃向一個鈷藍色木製圓柱。圓柱輕推一個番紅花色木製拱門。她的表情從專注期待,轉為驚訝,再變成俏皮的放鬆,同時繞著移動中的物件走動。提供的六段十秒片段顯示了該鏈條的不同部分。它們是生成的影片分鏡腳本,不是 Genie 4 輸出,也不是完成的物理測試。請同時留意成功的運動與連續性失誤。
使用 Seedream v5.0 Pro Text-to-Image(bytedance/seedream-v5.0-pro/text-to-image)建立共用開頭影格。若要重複執行,請使用 Seedance 2.5 Image-to-Video(bytedance/seedance-2.5/image-to-video,16:9 第一影格)生成十秒片段。若要重複執行,請使用前一段片段中清楚的結尾影格作為下一個第一影格。將原始角色與地點參考放在旁邊。本文提供的六個已下載 MP4 已轉換為 GIF 並置於下方。其內嵌中介資料未標明生成模型或原始提示詞,因此此處的模型名稱與書面提示詞是可重現的建議,並非這些檔案經查證的來源。這些片段不是互動式 Genie 4 錄影,也不是經量測的模擬。
Seedream v5.0 Pro 的共用第一影格提示詞:
plaintext1Wide 16:9 cinematic photograph inside an empty 1930s Art Deco theater foyer, designed as a clear stage for a kinetic collision scene. Cream-and-jade terrazzo floor with long geometric inlay lines, two brass-clad columns, deep indigo velvet curtains, and tall amber-and-cobalt stained-glass windows. Late-afternoon sun enters from camera right, casting sharply defined colored window shapes across the floor. At frame left stands one young adult dancer with warm brown skin, expressive dark eyes, and hair in a secure sculpted bun. She wears a cropped burgundy velvet jacket over an ivory silk blouse, deep-plum pleated wide-leg trousers, and cream leather dance shoes. Her posture is coiled for movement, one heel lifted, gaze fixed on one matte coral-red rubber ball in the foreground. A single cobalt-painted wooden cylinder stands three meters ahead on the ball's path; a single lightweight saffron-painted wooden arch stands behind it. All three movable objects are fully visible and clearly separated. Low 35 mm camera, strong floor leading lines, one softly blurred brass column at the edge, tactile velvet, silk, wood, and terrazzo, controlled coral-cobalt-saffron accents against rich indigo and warm ivory. Elegant photographic realism, natural shadows and slight film grain. No garden, plants, crowd, stage performance, text, signage, logos, extra balls, extra dancers, floating objects, or UI.
若要進行另一次生成,請在全部六段片段中維持相同的舞者、服裝、物件數量、地板幾何、建築、陽光方向與調色。每段十秒的片段都需要有可見動作與變化的攝影機方向。提供的片段按原樣顯示;當接觸或轉場不清楚時,圖說會註明。
鏡頭 1:第一次觸碰
**要檢查什麼:**起始位置、腳與球的接觸,以及球的初始方向。
plaintext1Use the supplied Art Deco foyer still as the exact first frame. Ten seconds of continuous, expressive movement. From 0-3 seconds, track laterally at knee height as the dancer glides diagonally along the jade-and-cream terrazzo toward the single coral ball. Her gaze stays on the ball; her expression shifts from focused anticipation to a quick playful smile, while her burgundy velvet jacket and deep-plum pleated trousers move with natural inertia. From 3-6 seconds, she makes a precise side step, rotates her shoulders and hips, and sweeps one cream dance shoe into clear contact with the ball exactly once. The ball begins rolling toward the stationary cobalt cylinder. Pan smoothly from her foot to the rolling ball without hiding the touch. From 6-10 seconds, follow the ball in a low three-quarter view while the dancer completes a controlled turn, catches her balance, and raises her eyebrows as she watches its path. Keep the ball, cylinder, and saffron arch distinct and spatially stable. Amber-and-cobalt stained-glass light, indigo curtains, brass columns, crisp shadows, natural 35 mm motion, rich but controlled color. No dialogue, text, crowd, plants, extra props, duplicated limbs, teleportation, or zoom-only motion. Wide 16:9.

舞者接近,並似乎讓球開始移動。腳與球的觸碰很短暫,因此該片段是視覺參考,而非受控衝量的證明。
鏡頭 2:主要碰撞
**要檢查什麼:**球與圓柱的接觸、球改變的方向,以及圓柱的第一次反應。
plaintext1Begin from the chosen final frame of shot one in the identical empty Art Deco theater foyer. Preserve the dancer's warm brown skin, sculpted bun, burgundy velvet jacket, ivory blouse, deep-plum trousers, cream dance shoes, and the exact coral ball, cobalt cylinder, and saffron arch. Ten seconds. From 0-3 seconds, a low front three-quarter tracking view follows the ball across the terrazzo inlay lines toward the still cylinder; show a clean gap until contact. From 3-5 seconds, the camera arcs to a clear side angle while staying wide enough to show both complete objects. The coral ball strikes the cylinder once at its lower half. The ball deflects diagonally left; the cylinder tips and begins rotating right. Make the contact point visible and do not let either object pass through the other. From 5-10 seconds, track laterally to hold both resulting motions in one frame. In the background, the dancer reacts with widened eyes and an open-mouthed smile, then pivots sharply around the cylinder's path. Stained-glass sunlight moves across real terrazzo texture, brass surfaces remain fixed, indigo curtains stay in place. Smooth cinematic 35 mm motion, believable gravity, no hidden impact, dialogue, text, crowd, plants, duplicate objects, or slow push-in. Wide 16:9.

圓柱在球到達時傾倒。在宣稱是經量測的碰撞反應之前,請先檢查接觸影格。
鏡頭 3:圓柱的後續運動
**要檢查什麼:**藍色圓柱在第一次撞擊後是否保持相同形狀,並以連貫方式移動。這段額外提供的片段不在原始五鏡頭提示詞計畫中。確切的原始生成提示詞無法取得;以下是重複執行時的建議提示詞。
plaintext1Continue from the final frame of the ball-to-cylinder impact in the same empty Art Deco theater foyer. Ten seconds. Keep one coral rubber ball, one cobalt-painted wooden cylinder, one saffron wooden arch, and the same dancer in a burgundy velvet jacket, ivory blouse, deep-plum wide-leg trousers, and cream dance shoes. From 0-3 seconds, track low beside the ball as it rolls away from the cylinder on the cream-and-jade terrazzo. From 3-7 seconds, shift to a clear three-quarter side angle as the cylinder tips, rocks on its edge, and gradually falls to the floor, keeping its blue color, size, and solid shape. From 7-10 seconds, widen the view to show the ball, fallen cylinder, arch, and dancer together; she crosses the background in two quick dance steps, then turns her head toward the moving cylinder with a surprised expression. Preserve the stained-glass light pattern, brass columns, indigo curtains, gravity, and object count. Smooth camera movement, no hidden collision, extra objects, text, plants, teleportation, or zoom-only motion. Wide 16:9.

圓柱傾斜並穩定下來,同時球仍在畫面中。在把這兩段視為連續片段之前,請將其起始位置與片段 02 比較。
鏡頭 4:次要碰撞
**要檢查什麼:**運動從圓柱轉移到拱門,且所有物件都保持其身分。
plaintext1Use the selected final frame of shot two as the first frame in the same Art Deco foyer. Maintain the exact burgundy-and-plum outfit, dancer's face and bun, coral ball, cobalt cylinder, saffron arch, terrazzo pattern, brass columns, indigo curtains, and light direction. Ten seconds. The cylinder is already rotating toward the stationary arch. From 0-3 seconds, track at waist height beside the cylinder while the dancer crosses behind it in a quick, elegant chasse step; her expression tightens from delight to alert concentration. From 3-6 seconds, drop to a low three-quarter angle and show the cobalt cylinder making one clear contact with the arch's lower left side. The light wooden arch rocks back, while the cylinder rolls a short additional distance; neither changes shape or color. From 6-10 seconds, orbit gently to keep both objects and their slowing motion visible as the dancer makes a small airborne turn, lands softly beyond them, and steadies herself with one extended hand. The pleated trousers and velvet jacket settle naturally. Colored stained-glass light cuts across cream-and-jade terrazzo, warm brass glints remain restrained, realistic 35 mm film texture. No shattering, magical sparks, speech, captions, crowd, plants, teleports, duplicate objects, or zoom-only motion. Wide 16:9.

拱門在圓柱附近傾倒,但確切接觸點不清楚。這是預期次要碰撞的分鏡腳本,不是所描繪轉移在物理上正確的證明。
鏡頭 5:遮擋與恢復
**要檢查什麼:**同一顆球在經過固定立柱後方之後,是否會重新出現。
plaintext1Continue from the preceding shot's final frame in the same empty Art Deco foyer. The single coral ball is still rolling slowly; the cobalt cylinder has shifted, and the saffron arch is rocking toward rest. Ten seconds. From 0-3 seconds, follow the dancer from slightly behind as she takes two fast, fluid dance steps across the geometric terrazzo and spots the ball. Her brows draw together in concentration, then one corner of her mouth lifts. From 3-6 seconds, slide the camera behind a fixed brass-clad column so the ball becomes briefly hidden by the column itself. Keep the column, floor pattern, and dancer spatially consistent while the ball continues its established direction. From 6-10 seconds, emerge at a low three-quarter angle as the same single coral ball rolls out from the far side with unchanged size, color, and direction. The dancer turns her shoulders, reaches toward it, stops short, and lets it pass, breaking into a delighted grin. Keep the cylinder and arch in plausible background positions. Amber-and-cobalt stained-glass light, indigo velvet curtains, tactile silk and terrazzo, smooth 35 mm camera movement. No speech, titles, crowd, plants, duplicate ball, object morphing, location jump, or empty slow zoom. Wide 16:9.

球短暫被固定立柱遮住。請比較其在遮擋前後的顏色、大小與方向。
鏡頭 6:後續發展
**要檢查什麼:**最終相對位置、穩定下來的運動,以及完整場景的連續性。
plaintext1Use the final frame of shot four as the first frame, preserving the dancer's burgundy cropped velvet jacket, ivory silk blouse, deep-plum wide-leg trousers, cream dance shoes, the same coral ball, cobalt cylinder, saffron arch, brass columns, indigo curtains, stained-glass windows, and terrazzo floor. Ten seconds. From 0-3 seconds, begin low beside the coral ball as it decelerates along the floor inlay lines; the saffron arch completes a small believable rocking motion farther back. From 3-6 seconds, crane smoothly into a wide three-quarter view that reveals all three objects at once in their changed positions. Do not reset them to their starting places. The dancer travels around the settled cylinder with two flowing steps and a full controlled spin, her pleated trousers describing a clear arc. From 6-10 seconds, arc the camera to face her as she catches her balance, checks the ball and arch, exhales, then smiles with visible relief and playful pride. Hold the final composition long enough to read every object's position. Warm directional window light and colored glass shapes, cream-and-jade terrazzo, rich indigo and burgundy, restrained film grain, physically plausible settling. No speech, captions, crowd, plants, duplicate props, sudden camera teleport, or push-in-only movement. Wide 16:9.

寬鏡結尾顯示舞者的旋轉與最終場景。請將拱門位置與片段 04 比較;若看似重設,就是連續性失誤。
將這六個提供的結果作為一個案例來檢視。如果腳部接觸看不見、偏轉在不同鏡頭間改變、物件重複,或最終位置跳動,請直接記錄該失誤。分鏡腳本可以傳達構想,但無法驗證數值物理。若進行真正的 Genie 4 工作階段,請使用上方更簡單的固定攝影機流程,儲存實際互動,並用同一張觀察卡比較其可見後續發展。
存取與後續步驟
Google 的公告將 Genie 4 描述為研究預覽版,先從受信任測試人員開始,並有與 Project Genie 及特定合作夥伴相關的更廣泛開放計畫。這是持續變動的推出過程,因此請查閱目前的官方 Genie 頁面以了解存取方式,而不要假設每位讀者今天都能試用。Project Genie 先前的公開提示詞建議是有用的背景資訊,但不保證能存取 Genie 4,或取得相同控制項。
如果你能存取 Genie 4,請從一次固定攝影機碰撞開始,並儲存提示詞與錄影。在得出結論之前,先填寫觀察卡。如果你無法存取,仍可現在準備此流程,或使用文字轉影片模型製作清楚標示的影片參考。對於 Genie 4 碰撞與物件移動物理模擬,核心問題不是某一次撞擊看起來是否很有說服力。而是接觸後場景是否仍可解讀,以及證據是否足以支持你需要做出的決定。
常見問題
Genie 4 是物理引擎嗎?
Google 將 Genie 4 描述為用於生成模擬環境並與之互動的世界模型。傳統物理引擎會公開場景的設定屬性與計算狀態。不要假設生成世界會公開相同資料,或提供經驗證的數值模擬。
Genie 4 能顯示球反彈或兩個物件碰撞嗎?
官方 Genie 資料描述具有物件碰撞的環境。特定提示詞或工作階段可能有所不同,因此請評估實際執行結果,並記錄接觸與撞擊後運動是否仍可見。產品示範不保證每個要求的碰撞都能實現。
Genie 4 對遊戲或工程物理來說足夠準確嗎?
生成場景可支援視覺探索,但單獨一個算繪結果無法證明數值準確度。實作、工程分析與安全決策,請使用遊戲引擎或適當的經驗證模擬器。
AI 影片生成器能測試物件移動嗎?
它有助於為提議的運動建立視覺參考。生成片段不提供與物理引擎相同的互動或可量測狀態,而且除非它確實來自 Genie 4,否則不應呈現為 Genie 4 輸出。






