> For the complete documentation index, see [llms.txt](https://duoduo-1.gitbook.io/duoduo-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://duoduo-1.gitbook.io/duoduo-docs/how-to-create-a-veo-3.1-video-from-prompts-reference-images-and-keyframes.md). # How to Create a Veo 3.1 Video From Prompts, Reference Images, and Keyframes A practical workflow for planning a shot, choosing text, image, or keyframe guidance, generating a Veo 3.1 clip, and reviewing the result. AI video generation works best when you treat it like planning a single film shot, not writing a story outline. A clear subject, one readable action, a deliberate camera move, and a defined visual finish give the model fewer conflicting decisions to make. This tutorial shows how to build a short clip with [Voe Ai](https://voe31.com/), an independent AI video platform that provides workflows powered by Veo 3.1. It is not affiliated with or endorsed by Google. Depending on the project, you can begin with text alone, animate a reference image, or use keyframes to guide the beginning and end of the shot. ### What you need before generating Prepare a short shot brief before opening the generator. Write down the subject, the action, the setting, the camera behavior, the intended destination, and the aspect ratio. If you are using a reference image, choose a clean, high-quality frame with the subject clearly visible. If you are using keyframes, make sure the opening and closing images depict the same subject and a believable visual transition. Also decide where the video will be used. Landscape 16:9 is a natural choice for websites, YouTube, presentations, and cinematic scenes. Vertical 9:16 is better for Shorts, Reels, TikTok, and mobile-first ads. Making this decision early prevents important details from being cropped later. ### Step 1: Define one shot and one outcome Keep the first attempt focused on a single shot. For example: “A ceramic coffee cup sits on a wooden table while morning light moves across the steam; the camera slowly pushes forward.” This is easier to control than a prompt asking the subject to enter a room, speak, change clothes, pick up an object, and move to a second location. A useful shot brief answers six questions: * Who or what is the main subject? * What single action happens? * Where and when does it happen? * How should the camera move? * What lighting, color, and mood should the clip have? * What must remain stable or absent? If the result is for an advertisement or product page, add practical constraints such as “keep the label readable,” “do not change the product shape,” or “leave negative space on the left for a headline.” ### Step 2: Choose the right workflow **Text-to-video** is best when you want the model to invent the composition. Use it for atmosphere, locations, cinematic concepts, abstract visuals, or shots where exact character identity is not critical. **Image-to-video** is useful when the starting composition, product, person, or illustration already exists. The reference image establishes appearance and framing, while the prompt explains the motion. Describe what should move and what should stay fixed. **Keyframe-guided generation** is appropriate when the opening and ending states matter. For example, an opening frame may show a closed perfume bottle on a table, while the final frame shows the bottle surrounded by drifting flower petals. The keyframes define visual anchors; the prompt should describe a plausible transition between them. ### Step 3: Write a production-ready prompt Build the prompt in this order: subject, action, environment, camera, lighting, style, and constraints. Concrete visual language is more useful than a list of vague adjectives. For example: “A red electric bicycle parked beside a modern glass building at blue hour. The front wheel turns slightly in a light breeze while reflections move across the metal frame. Slow left-to-right dolly shot at waist height, shallow depth of field, realistic commercial lighting, clean product proportions, no text, no extra bicycles.” For a vertical social clip, you might write: “A pastry chef places a glossy fruit tart on a marble counter. The camera starts in a medium vertical frame and gently pushes toward the tart as warm window light reveals the fruit texture. Natural hand motion, premium food-commercial look, uncluttered background, leave clear space above the subject for a caption.” Avoid contradictory directions such as “locked camera” and “fast orbit” in the same prompt. Do not stack several camera moves unless the shot genuinely needs them. When identity or product accuracy matters, explicitly state which details must remain unchanged. ### Step 4: Prepare reference images and keyframes Use reference images with good exposure, clear silhouettes, and enough resolution to show the details you care about. Heavy compression, motion blur, tiny faces, and obstructed subjects give the model less reliable information. For image-to-video, describe motion rather than repeating every visible feature. A useful instruction is: “Preserve the subject’s clothing, face, and background composition. Add a subtle head turn and natural breathing while the camera remains stable.” For keyframes, keep the subject, lens perspective, lighting direction, and aspect ratio consistent between the first and last images. Large discontinuities can force the model to invent an unnatural transformation. If the two frames are very different, simplify the transition or create an intermediate shot instead. ### Step 5: Select the aspect ratio and check generation settings Choose 16:9 for landscape delivery or 9:16 for vertical delivery. Review any available duration, quality, privacy, audio, or generation options before starting. AI video uses credits, so confirm the displayed cost and current settings rather than assuming every generation has the same price. For an early test, prioritize composition and motion over a perfect final render. A lower-cost draft can reveal whether the prompt is understandable before you spend more credits on refinements. ### Step 6: Generate the first controlled pass Run the prompt without adding unnecessary variations. The first pass is a diagnostic: it shows how the model interprets the subject, action, and camera. Save the prompt and settings with the output so you can compare revisions. When a result is close, change only one or two variables. For example, keep the subject and lighting but reduce the camera speed. If you rewrite the entire prompt after every attempt, it becomes difficult to learn which instruction improved or damaged the result. ### Step 7: Review the video frame by frame Watch the clip at normal speed, then scrub through it slowly. Check: * Prompt adherence: does the intended action actually occur? * Identity and object consistency: do the face, hands, clothing, logos, and product geometry remain stable? * Physics: do contact, weight, shadows, reflections, and moving objects behave plausibly? * Camera control: is the direction, speed, and framing correct? * Beginning and ending: do the first and last moments feel intentional rather than abruptly cut? * Text and branding: are any labels, signs, or logos distorted? * Audio, when available: does it fit the action without unwanted speech or noise? Review the output in the final crop as well. A landscape clip that looks balanced may lose the subject when repurposed vertically. ### Common problems and practical fixes #### The clip tries to do too much Reduce the prompt to one action and one camera move. Turn a multi-event concept into several short shots that can be edited together. #### The subject changes appearance Use a stronger reference image, shorten the action, and state the features that must remain consistent. Avoid rotations or occlusions that hide the subject for long periods. #### The camera moves in the wrong direction Use direct language such as “camera slowly dollies forward” or “locked tripod shot.” Remove stylistic phrases that might imply a conflicting move. #### The ending does not match the desired composition Use a final keyframe and describe the transition, not just the destination. Ensure the opening and closing frames use compatible perspective and lighting. #### Hands, small objects, or fine details look unstable Make the action slower and simpler, keep important objects clearly visible, and avoid placing critical details at the edge of the frame. If necessary, use a closer shot with less background activity. ### A reusable prompt template Use this structure for future clips: > \[Main subject] performs \[one clear action] in \[setting and time]. The camera \[one specific camera behavior]. Lighting is \[direction and quality], with \[color and mood]. Preserve \[identity, product, or composition details]. Avoid \[unwanted objects, text, distortion, or extra motion]. Compose for \[16:9 or 9:16] with \[space or framing requirement]. Keep a simple generation log with the prompt, reference assets, aspect ratio, settings, result, and the next change. This turns trial and error into a repeatable production workflow. ### Rights, privacy, and disclosure Only upload images and create videos you have permission to use. Get consent before animating a recognizable person, and avoid deceptive impersonation or misleading endorsements. Review the platform’s current terms and privacy information for your intended commercial use. If a synthetic clip could reasonably be mistaken for real footage, label it appropriately. ### Conclusion A strong Veo 3.1 workflow starts with a small, testable shot. Select text, image, or keyframe guidance based on what must remain controlled; write the prompt in production terms; choose the final aspect ratio before generation; and review motion, identity, physics, camera, and branding one frame at a time. Iterating on one variable per pass will usually produce better and more predictable results than repeatedly rewriting the entire concept. --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://duoduo-1.gitbook.io/duoduo-docs/how-to-create-a-veo-3.1-video-from-prompts-reference-images-and-keyframes.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.