10 Video Creation Tips for Google Flow: Helping Gemini Omni Understand Your Camera Directions

If you want to create a product ad with AI, you might already have product photos, character shots, or even a sketched storyboard. But going from these assets to a satisfying video requires communicating many intermediate details: how characters move, how the camera travels, when shots cut, and how to tweak a single detail after generation.

Google Flow is Google's AI creative workspace designed for creators to generate and edit images and videos. This official guide focuses on Gemini Omni Flash in Flow: you can describe what you want to capture using text, images, or reference videos, or continue requesting edits on existing footage.

A common pain point is that asking for something to look “more cinematic” fails to provide concrete specifications. You might only want to adjust the lighting, but the model also changes camera angles; you might want to capture an action in one continuous take, but it cuts in the middle. Once the creative vision is there, clearly communicating it to the model becomes the next challenge.

To address this, the Google Flow team compiled this creative prompting guide, outlining ten techniques alongside concrete examples: locking down subjects and framing with visual references, choreographing motion through camera and timing directives, and clearly specifying what to change and what to keep during edits. Below, we break down how to apply each technique step by step, complete with official video clips and bilingual prompts in Chinese and English.

The English prompts are preserved verbatim from the official original text, including all punctuation, asset placeholders, and timecodes; the Chinese lines serve as parallel references, with usage notes listed separately. A few unclosed quotation marks in the original text are also preserved as-is.

1. Lock Down the Visuals First: Text Constraints and Start & End Frames

Establish visual requirements before choreographing motion within the frame. Setting, wardrobe, and camera style can be defined through text; framing, subjects, and color grading can be guided more concretely with images.

Tip 1: Provide high-level constraints

Instead of micromanaging the sway of every single hair, provide Gemini Omni with high-level artistic and physical constraints. The official guide recommends explicitly instructing the model to focus on lighting consistency, set design, clothing textures, and movement details, allowing it to flesh out the scene within these boundaries.

Official prompt text and Chinese reference:

• Setting and wardrobe texture constraints

  中文： 细化场景环境。对角色应用服装设计原则，精确指定具体的材质纹理、道具与光影氛围。

  English: “Be extremely detailed in environments. Apply costume design principles to characters, specifying exact textures, props, and lighting.”

  Usage: Tells the model which environmental, wardrobe, and prop details to prioritize fleshing out.

• Micro-expression and natural dialogue constraints

  中文： 深入考量细节、面部表情以及节奏停顿，营造出自然的对话状态。

  English: “Consider micro-detail, facial expression, and timing to create a natural dialogue.”

  Usage: Incorporates facial expressions and conversational timing into the prompt rather than just supplying lines.

• Specialty camera lens constraint

  中文： 用手持鱼眼镜头拍摄房间里的角色。

  English: “Handheld fisheye camera shot of the character in the room.”

  Function: Directly invokes specific optical lens language to quickly establish perspective tension.

Tip 2: Use visual anchors

The biggest pitfall with long takes or complex motion is “drifting”—a camera sweep occurs, and suddenly the clothing changes color, or the background transforms into a different city altogether. Uploading or pre-generating a Start Frame and End Frame provides the model with clear anchor points for both the beginning and end of the shot. Because static frames establish framing, color palette, and subjects in advance, your text prompt can focus entirely on how the camera and characters move. The official guide uses this technique to reduce visual drift.

Official prompt text and Chinese reference:

• Drone pull-back and ascent

  中文： 基于给定的首尾帧，执行平滑连贯的高速无人机后撤镜头。摄像机快速向后倒退并俯倾，同时爬升数百英尺冲入高空，云雾散开，显露出壮丽雄伟的山脉全貌。

  English: “Smooth, continuous high-speed drone pull-back shot using this start and end frame. The camera rapidly retreats backwards and tilts down, ascending several hundred feet into the air. The mist parts to reveal the massive mountain range.”

• Day-to-night ambient lighting hyperlapse

  中文： 基于给定的首尾帧，使用锁定的静止机位进行延时摄影。转场无缝衔接，将环境氛围光从湿冷的暮色平滑过渡为温暖的晨间日光。

  English: “Hyperlapse with a locked-off, static camera using this start and end frame. The transition is seamless, shifting the ambient lighting from cold, wet twilight to warm morning sunlight.”

• Infinite seamless loop

  中文： 基于给定的首帧和尾帧，制作一个首尾相接、无缝循环的动态视频。

  English: “Create an endless loop using this start and end frame.”

2. Anchor Content with Assets, Shape Pacing with Directives

Images and storyboards can directly convey characters and framing relationships; pacing, however, requires further clarifying when actions occur and at what exact moments things speed up or slow down.

Tip 3: Leverage tagged ingredients

You don't have to rely exclusively on lengthy text descriptions. Inside the workspace, you can tag uploaded assets: use images to anchor a Character, Setting, or Style; use reference videos to guide motion pacing; or even input a complete multi-panel Storyboard. Tag your assets in the workspace first, then reference the corresponding asset tags in your prompt; placeholders below such as [@ Tagged Image] should be replaced with your actual assets.

Official prompt text and Chinese reference:

• Commercial product placement

  中文： 将 [@ Tagged Image]作为商业视频广告中的核心产品，画面中一位女性正在暖粉色影棚背景下拆封并试用该产品。

  English: “Use [@ Tagged Image] as the product in a video ad campaign, where the product is being opened and tested by a woman set against a warm pink studio.”

• Subject swap preserving full motion and audio reference

  中文： 以此参考视频作为动作与声音参考基准，将画面主体替换为 [@ Tagged Character]，其余一切保持不变。

  English: “Using this video as a motion and audio reference, change the subject to [@ Tagged Character], keep everything else the same.

• Animate multi-panel storyboard

  中文： 让这张 [@ Tagged Image]分镜序列动起来，按照自上而下、从左到右的关键节拍顺序生成动态画面。

  English: “Animate this [@ Tagged Image] storyboard image sequence, following key beats exactly in order from top to bottom, left to right.”

Tip 4: Refine pacing mid-scene

Generating the motion until satisfied and then adjusting its speed is one recommended approach in the guide. Rather than simply writing “slower,” specify which action slows down, when the deceleration begins, and what target speed you want to reach. The speed multipliers and second counts below are prompt targets; always inspect the actual output to verify pacing.

Official prompt text and Chinese reference:

• Localized speed ramp for interaction

  中文： 将两个主体之间互动的动作移动速度提升至 4 倍速。

  English: “4x the moving speed of the interaction between the two subjects.”

• Key-action-triggered ramp into slow motion

  中文： 当狗狗到达跳圈时，时间流速逐渐降低，进入极慢动作。

  English: “As the dog reaches the hoop, time ramps down into extreme slow motion.”

• Mid-shot speed ramp at a specified second

  中文： 角色快速行走；在第 3 秒，一朵从天而降的花朵靠近她的手掌，随后时间减速至 1% 的极度缓慢，记录花朵从指缝间滑落的过程。单镜头一镜到底，无镜头剪辑。

  English: “Fast-walking character; at sec 3, a sky-falling flower nears her hand, then time slows to 1% speed as it slips through her fingers. Single shot, no cuts.”

3. Transfer Reference Styles and Bring Text into the Frame

Beyond fundamental motion generation, Gemini Omni also has the ability to flexibly adapt external styles and natively render animated typography directly in the frame.

Tip 5: Transfer and mix styles

You can treat the model as a professional stylist and colorist for every shot. Provide a reference image or video, and clearly state which dimension you want to extract—whether it is pose, lighting and color grading, or surface texture—so it can apply that specific dimension to your target scene.

Official prompt text and Chinese reference:

• Transfer pose and motion

  中文： 将此参考视频中的姿势与动态轨迹，应用到参考图片中的角色身上。

  English: “Apply the pose and motion from this video to the character in this image reference.”

• Full material and visual style replacement

  中文： 将这张照片的整体视觉风格融入生成中，同时将所有物体的表面材质替换为参考视频中的质感外观。

  English: “Apply the visual look of this photo to the generation, replacing all materials with the appearance of the video reference.”

• Color Match

  中文： 将此参考图的色彩基调与调色应用到我的场景中。

  English: “Match the color grading of this image to my scene.”

Tip 6: Render animated text

The guide also introduces kinetic typography: having text appear and animate directly inside the video. You can directly specify font styles, screen placement, alignment rules, and text animation rhythms—ideal for tutorial callouts, stylish on-screen captions, and playful commentary.

Official prompt text and Chinese reference:

• Hand-drawn annotations and instructional arrows

  中文： 在这道烹饪视频中，为每一种食材添加手绘马克笔风格的文字注释与指向箭头。

  English: “Add hand-drawn marker-style text annotations and arrows for each ingredient in this cooking video.”

• Word-by-word animated sizzle typography

  中文： 屏幕上逐字浮现：“Did you know this model can do typography?”，每个单词都带有独特的动效样式，并与快节奏高能节拍同步。

  English: “Word-by-word on screen: ‘Did you know this model can do typography?’ Each word appears with a unique animated style, synced to a sizzle reel rhythm.”

• Internal monologue thought bubble

  中文： 在女性身旁浮现一个对话气泡，文字显示：“如果我现在睡着，我还能睡几个小时？”。

  English: “A speech bubble appears next to the woman reading, ‘If I fall asleep now, how many hours of sleep will I get?”

4. Keep the Shot Continuous or Cut Between Angles

The same action can be tracked in one continuous shot or broken into several cuts. Your prompt needs to specify which viewing experience you want and how shots should connect.

Tip 7: Enforce continuous scene rules

If what you need is a fluid, unbroken long take (a “oner”) or a side-scrolling tracking shot, you must explicitly rule out editing cuts in your prompt, as the guide notes the model will otherwise default to switching camera angles to add narrative depth.

Official prompt text and Chinese reference:

• Complex tracking oner across spaces

  中文： 摄像机跟随角色走下工作通道楼梯，穿过喧闹忙碌的后厨，最终步入拥挤热闹的夜总会大厅。

  English: “The camera follows the characters down service stairs, through a bustling kitchen, and into the crowded nightclub.”

• Flat side-scroller tracking shot

  中文： 拍摄为沿着田野平移的连续横版卷轴跟拍镜头。随着主角骑马进行障碍跳跃，摄像机如同架设在轨道上般水平侧向跟拍。

  English: “Shot as a continuous, flat side-scroller tracking shot along a field. The camera moves laterally along a rail as a single protagonist does show jumps on horseback.”

• Uncut in-frame and out-of-frame action sequence

  中文： 在没有任何场景剪切的情况下，以单一连续的手持运镜记录角色走向房门、拉开门走出去并在身后顺手关门的完整动作。

  English: “Without any scene cuts, capture a single continuous, handheld shot of the character walking up to the door, opening it, and stepping outside while closing it behind her.”

Tip 8: Introduce transitions and cuts

Conversely, when you do need a multi-shot narrative, do not leave cuts to chance. Instead, direct the sequence like a storyboard artist by clearly specifying the visual matching relationships and object associations (Match Cut) between shots.

Official prompt text and Chinese reference:

• Light-streak metamorphic transition

  中文： 黑暗城市高速公路上飞速掠过的车流光轨，无缝形变转场为深夜海滩上翻涌拍岸的生物发光荧光海浪。

  English: “Fast transit lights streaking across a dark city highway transforming into bioluminescent waves crashing on a night beach.”

• Geometric Match Cut sequence

  中文： 一组由“旋转的咖啡杯”、“巨大的摩天轮”以及“指针手表表盘特写”构成的几何相似体匹配剪辑序列。

  English: “A match-cut sequence of a spinning coffee cup, a massive Ferris wheel, and a close-up of an analog watch face.

• Progressive framing narrative sequence

  中文： 先是女子吹气球的近景，随后切到她在湖中泛舟的画面，最后以她在漂满气球的宽阔湖面上划船的宏大全景收尾。

  English: “A close-up shot of a woman blowing up a balloon, followed by a cut to her rowing on a lake, ending with a wide shot of the lake filled with balloons as she rows

5. Scope Your Edits and Direct Timing with Precision

Few things are more frustrating in video production than revisions: changing a single caption can reshuffle the entire video. Meanwhile, if a video is to have rhythm, the ability to align events second by second is indispensable.

Tip 9: Precision conversational editing

When requesting revisions on an existing video clip, keep your prompt strictly focused on what needs to change. Overloading the prompt with background environment descriptions may mislead the model into thinking you want to generate a completely new scene. The guide recommends a highly practical phrase: append “Keep everything else the same” to the end of your prompt.

Official prompt text and Chinese reference:

• Prop invisibility / object removal

  中文： 将她手中的手机完全隐藏隐形。其余一切保持不变。

  English: “Make the phone in her hand invisible. Keep everything else the same.”

• Precise sign text replacement

  中文： 将招牌指示牌上的文字修改为“Omni Flash”。其余一切保持不变。

  English: “Change the sign text to read ‘Omni Flash’. Keep everything else the same.”

• Localized relighting while preserving identity

  中文： 在保持角色面部与外貌特征的前提下，为主体打上一束极具戏剧感的轮廓边缘光。

  English: “Add a dramatic rim light on the subject while maintaining character features.”

Tip 10: Direct with timecodes

Gemini Omni features native audio-visual synchronization during generation. Creators can specify timing in seconds using natural language, or use bracketed timestamps such as [0-3s] to schedule action beats much like formatting a screenplay.

Official prompt text and Chinese reference:

• High-energy synth beat synchronization

  中文： 包含一段与画面转场对齐的高能合成器节拍音乐。

  English: “Include a high-energy synth beat synced to the visual transitions.”

• Bracketed timestamp screenplay (Multi-character staging)

  中文：
  美丽郊区人行道的固定机位镜头：
  [0-3 秒] 一只猫从画面左侧跑入，并从右侧跑出画框；
  [3-6 秒] 一名男子从左侧跑入追赶那只猫，并从右侧跑出画框；
  [6-10 秒] 一群狗狗从左侧跑入，追赶着前面的男子和猫，一起从右侧出画。
  
  

  English: A locked camera shot of a beautiful suburban sidewalk. [0-3s] A cat runs from left of frame and off frame right [3-6s] A man chases after the cat from left of frame and off frame right [6-10s] A group of dogs are chasing the man and the cat from left of frame and off frame right

• Natural language second-by-second action prompt

  中文： 角色在如附图所示的粉色冰场中滑冰：在第 2 秒时，角色从镜头左侧入场；在第 5 秒时，天空开始飘雪，角色抬头仰望。

  English: Character ice skating in a pink ice ring like the image attached: At 2 seconds, the character enters from the camera left. At 5 seconds, it starts snowing and the character looks up.

6. Supplement: Using Negative Constraints

When generating video, defining what you do not want is often just as crucial as defining what you do want. To avoid stray bystanders in the background, jarring voiceovers, or strange plastic sheen, consider appending clear, structured negative phrases to your prompts:

How to Combine These Techniques in a Single Workflow

In practice, you can sequence these techniques in the following workflow:

1. Anchor the visuals upfront: Take advantage of @ tags to introduce existing characters and storyboards, or specify Start and End Frames to stabilize your visual baseline.
2. Establish ground rules for camera motion and narrative flow: Add No scene cuts when you need a continuous take, or proactively structure transitions with Match Cut when you want edited pacing.
3. Bound your revision scope: Avoid rewriting the entire prompt when correcting minor flaws; isolate the edit target and append Keep everything else the same.
4. Schedule action beats with seconds: Use bracketed timestamps like [0-3s] to break down actions for the model, clarifying the sequence and timing targets of each event.

The overarching goal of these methods is to eliminate ambiguity in your instructions: assets define the subjects, camera directions define the perspective, timecodes define the sequence, and edit directives define the scope of changes. After each generation, review the output against these targets and iterate accordingly.