Gemini Omni Flash（含 1.1）｜提示詞教學（繁體中文）

1. 一句話定位

同時讀取文字、相片與影片並一併推理。最多 7 張參考圖＋1 段參考影片（有影片時最多 5 張圖），720p，4/6/8/10 秒。 適合品牌約束廣告、VFX 加元素、換角色但保留場景。2026-08 遵從度測試：Gemini Omni Flash 1.1 在風格化輸出最強。

官方定位（原文）：Fast multimodal reference-to-video. Use for prompt-guided video generation from image references and optionally one video reference, especially when the brief benefits from Google's Gemini/Veo-style understanding without making it the default over Seedance 2.5.
— 來源：skills-catalog

2. 規格速查

3. 提示詞結構公式

[Subject + action] → [Setting + constraints] → [Camera behavior]
Label each input explicitly: "@character_photo ...", "video_1 ..."

官方：「Subject and action first. Setting and constraints second. Camera behavior third.」並在提示中明確標註每個參考輸入（@character_photo），讓模型知道哪個輸入是角色、與動作有何關係。

來源：blog:how-to-use-gemini-omni-flash-multi-shot-video

4. 技巧與範例提示詞

4.1 結構化基本提示（標註參考輸入）

範例提示詞（英文原文，逐字引用）：

@character_photo standing at a rain-slicked market stall at dusk, looking directly at camera, slow dolly in, warm tungsten light from the stall, shallow depth of field.

中文解說：官方範例（譯：@character_photo 站在黃昏雨濕的市集攤檔前，直視鏡頭，緩慢推軌，攤檔的暖鎢絲燈，淺景深）。主體＋動作 → 場景 → 鏡頭 → 光線。

來源：blog:how-to-use-gemini-omni-flash-multi-shot-video

4.2 品牌約束：把限制直接寫明

範例提示詞（英文原文，逐字引用）：

The label color stays consistent across all cuts. The product shape does not distort.

中文解說：官方建議的品牌廣告約束寫法（原文引號內句子）。

來源：blog:how-to-use-gemini-omni-flash-multi-shot-video

4.3 真人＋3D CGI 混合

範例提示詞（英文原文，逐字引用）：

A cinematic mix of live-action and 3D CGI. A hand presses a key on a laptop, and a small, blue, furry cartoon character wearing pink glasses and a Hawaiian shirt jumps out of the screen. The character multiplies rapidly, flooding the desk and then a busy city intersection. A yellow taxi hits one of the small creatures, causing it to magically grow into a giant, Kaiju-sized monster. The giant blue furry creature stomps through the city streets, pressing its huge hands against a glass office building while people run away. A young Asian man in a purple hoodie stands on the street, pulls out a glowing pink and green box, and opens it. A magical pink light beams out, sucking the giant monster and all the small creatures into the box like a vacuum. Cinematic lighting, dynamic camera angles, blockbuster movie VFX.

中文解說：官方範例：手按鍵盤，藍色毛茸卡通角色從螢幕跳出。精確描述位置、光線、鏡頭與動作，模型一次過產出。

來源：blog:how-to-use-gemini-omni-flash-multi-shot-video

4.4 「保留一切，只換角色」編輯

範例提示詞（英文原文，逐字引用）：

Preserve everything about the scene exactly as it is: the static medium shot of a character sitting behind a white office desk, waist-up framing, in front of a large floor-to-ceiling window overlooking a dense sunny city skyline. Preserve the desk setup untouched, the black rotary telephone, the folded newspaper with a yellow pencil resting on it, the small black leather notebook, and the desk pad. Preserve the exact timing and actions: the character holds the phone receiver to their ear from the start, hangs it up at the same moment, then turns toward the window, exhales, and leans back with arms settling in a relaxed, satisfied posture. Preserve the camera, which stays completely static throughout, the bright natural daylight pouring in from the window, and the soft shadows across the office. Preserve the audio exactly the same spoken line, the same phone-hang-up sound, and the same exhale.

Replace the character entirely with a new one: swap the previous figure for a young man with dark curly hair and a chiseled jawline, wearing an oversized tan/beige retro 1980s business suit over a light blue button-up shirt, finished with a bold leopard-print silk tie. Give him the same confident, corporate energy, focused and slightly tense while on the call, then relieved and smirking as he leans back. Every other element of the shot, motion, timing, lighting, and sound must remain identical to the original.

中文解說：官方範例：先逐項列出要保留的（靜態中景、白色辦公桌、黑色旋轉電話…），再寫「Replace the character entirely with a new one」並詳述新角色。先鎖定、後替換。

來源：blog:best-ways-to-access-gemini-omni-flash-2026

4.5 VFX 加元素：SOURCE LOCK（V2V）

範例提示詞（英文原文，逐字引用）：

=== BIRD LEAVES THE SCREEN — V2V on video_1 (~10s, 16:9, 30fps) === SOURCE LOCK: This is an ADD-ELEMENT VFX pass on video_1, NOT a new generation. Keep 1:1 everything already in the plate—the black monitor, the desk setup, the hand and its motion/timing, and the exact CAMERA path (the initial hold, the pan to the right, and the slow sweep across the office space) and the EDIT. Do NOT re-frame, re-time, re-angle, re-cut, or change the hand's positioning. Only ADD the live bird and animate it. SUBJECT: One real Common Kingfisher—bright blue plumage on the back, vibrant orange-rufous underparts, long black bill, short red legs. It must 100% match the appearance and proportions of the kingfisher shown on the monitor screen. THE MECHANIC (mapped onto the plate's own beats): 0–1.5s: The bird stays perched on the branch inside the monitor image (screen image unchanged), then subtly comes alive—a quick head twitch, blink, and slight breathing movement. 1.5–2.5s: As the hand waits, the bird leans forward and flutters out of the display plane, crossing from the flat screen into real 3D space in front of the monitor with rapid, sharp wingbeats. Behind it, the image on the monitor screen seamlessly shows the same branch but empty. 2.5–3.5s: The bird alights on the open palm—real weight settles, feet grip the hand, wings fold tight, it glances around. 3.5–5.5s: As the camera begins to pan to the right, the bird stays perched on the hand, tracking with the hand's motion through the frame. 5.5–7s: As the hand remains extended and the camera continues scanning the office, the bird crouches and launches with a sharp, fast downstroke, taking off into the open office space. 7–10s: The camera completes its pan across the office desks, partitions, and curtains. The bird is seen flying dynamically through this background 3D space, darting between the workstations before exiting the frame or fading into the distance near the background curtains. LIGHT-MATCH / INTEGRATION (top priority): The real bird is lit by the room—the bright overhead LED panel light and ambient warm office lights. It must reflect the same exposure, color grade, and grain as the plate, not the outdoor lighting from the original screen image. Precise soft contact shadows of the bird and its feet must cast onto the palm. Feathers should catch the cool overhead glare as it moves. ANTI-SLOP: Real feather texture with high detail; lively eyes with a sharp catchlight; convincing flight physics with proper weight, acceleration, and air resistance. No robotic/CGI look, no floating, and no morphing shapes or extra limbs.FORBIDDEN: Changing the original camera pan, altering the hand's position, modifying the office background, or changing the timing of the camera movement.Diegetic SFX only: Rapid, high-pitched wingbeats, a sharp kingfisher whistle chirp, and the ambient office hum from the plate. No music. No on-screen text.

中文解說：官方 VFX 範例：開頭聲明「This is an ADD-ELEMENT VFX pass on video_1, NOT a new generation. Keep 1:1 every…」——明確告訴模型這是加元素，不是重新生成。

來源：blog:gemini-omni-flash-vfx-video-editing

5. Do / Don't 清單

來源：blog:how-to-use-gemini-omni-flash-multi-shot-video, skills-catalog

6. CLI 參數表（自動擷取自官方 MODELS.md）

自動解析 cli-models；以 higgsfield model get <id> 取得即時 schema。

gemini_omni — Gemini Omni Flash

• 限制：At most 1 video_references entry is allowed.
• 限制：When a video reference is provided, at most 5 image_references are allowed.
• 限制：At most 7 image_references are allowed.

higgsfield generate create gemini_omni --prompt "cinematic product reveal with smooth camera motion" --image ./product.png --duration 8 --resolution 720p --wait

來源

詳見同資料夾 sources.md。所有事實存取日期：2026-10-08。