Google Flow · Make a 30-Second AI Video with Free Daily Credits

Quick Summary

• The workflow uses one reusable Google Flow character, three 10-second Gemini Omni Flash generations and Scenebuilder. Each new clip begins from the saved final frame of the previous clip, which creates a visual handoff without requiring Flow's Extend feature.
• As of October 21, 2026, Google's official credit table gives users 50 daily Flow credits. A 10-second Omni Flash generation at 720p costs 15 credits, so three single-output clips cost 45 credits. The same three clips at 360p cost 21 credits and are useful for drafts.
• Daily credits refresh after the first generation on the following day and unused daily credits do not roll over. Non-subscribers may also be unable to generate video during peak hours, so free access is real but not guaranteed at every moment.
• For continuity, select the character before every generation, explicitly choose and reuse one voice, lock the framing and lighting language, forbid captions and on-screen text, and end each clip with the character still and facing the camera.
• Google Flow's Scenebuilder can arrange, trim, preview and download the three clips as one sequence. Choose the target orientation before generation because reframing finished clips is less reliable than designing for vertical or landscape from the start.
• The companion Claude Code automation uses the community Flow MCP repository. It is a third-party workflow, not an official Google MCP. Review its code, authentication method and account permissions before connecting it to Flow.

The verified free-credit math

Google's current Flow pricing page lists 50 daily credits for users with or without a Google AI subscription. The creator uses Gemini Omni Flash at 720p, where one 10-second generation costs 15 credits. Three clips therefore cost 45 credits, leaving five daily credits unused.

The important word is generation. Google's help page says costs apply per generated result, not per request, and some requests can create two videos. Set the output count to one before approval or the same prompt can consume twice the expected budget.

There is a cheaper planning option. A 10-second Omni Flash clip at 360p costs seven credits, so three drafts cost 21 credits. Use 360p to test scripts, timing, continuity and camera behavior. Switch to 720p only after the prompt structure works. Free users cannot use 1080p upscaling, while supported paid plans currently include it.

Daily credits do not roll over. Google says the refresh is triggered by the first generation, with the next 50 credits available the following day. It also warns that non-subscribers may be unable to generate video during peak hours, roughly 2 PM to 5 PM UTC. Treat the free allocation as a daily creative budget, not a service-level guarantee.

Choose a fictional character or your verified avatar

The tutorial uses Mr. Smith, a fictional character created inside Flow. This is the broadly available route and avoids tying the demonstration to the creator's own biometric likeness. In the Characters area, describe the person, generate one or two reference images, name the character and assign a voice.

Google distinguishes a reusable character from an avatar of the account owner. Characters bundle visual and audio references for repeatable generations. A personal avatar is built from selfie and voice recordings, requires a mobile device to complete verification and can be called with the account's avatar reference.

The personal-avatar feature is currently unavailable in the European Economic Area, the United Kingdom and Switzerland. Flow itself is available in many of those regions, including Portugal, but the verified-avatar feature is not. A fictional character remains the practical route for creators in those locations.

Use only a real person's face or voice when you have clear authorization. For a fictional presenter, avoid designs that deliberately imitate a recognizable living person. The workflow works because the character is consistent, not because it resembles someone specific.

Step 1: build and lock the character

Create a new Flow project and open Characters. Start from a professional template or write a precise description covering approximate age, facial structure, hair, wardrobe, camera distance, background, lighting and intended tone. The tutorial uses Nano Banana Pro to generate the visual reference and keeps the image in landscape for a standard YouTube format.

A talking-head character needs a clean face and shoulders more than a full-body reference. Use a simple background, centered framing and even light. Extra props, text, hands near the face and complex scenery create more details for the video model to reinterpret from clip to clip.

Name the character immediately and select a voice. The creator left voice selection automatic and happened to receive a consistent voice, but he correctly identifies that as a risk. Google's current Flow documentation lets a character use a selected or custom voice, and the prompt can reference a chosen voice directly.

Keep a small continuity card beside the project: character name, voice, wardrobe, background, lens feel, camera height, key-light direction, color temperature and speaking style. Repeat the important details in all three video prompts even when the character reference is attached.

Step 2: write all three clips before generating

Plan a three-beat structure before spending credits. Clip one delivers the hook and first useful point. Clip two begins with a natural bridge and adds the second point. Clip three resolves the idea and gives the call to action. Each spoken segment must fit comfortably inside ten seconds.

Ask ChatGPT or another writing model to return all three prompts together. This exposes repeated ideas, weak transitions and scripts that are too long before Flow charges anything. Read every line aloud with a timer. Ten seconds of generation does not mean ten seconds of comfortable speech if the character also needs a pause at the beginning or end.

Every prompt should restate the character reference, selected voice, camera lock, lighting lock and the instruction to avoid captions, subtitles, logos and all other text. Clips one and two should end with the character still, mouth closed, facing the camera and holding the same neutral pose for the final half second.

Step 3: generate the first 10-second clip

Turn on Flow's Agent mode, set the model to Gemini Omni Flash and confirm one output. Add the character to the prompt through the reference control before pasting the script. The creator's first failed attempt omitted Mr. Smith, so Flow used his personal avatar and consumed 15 credits on the wrong subject.

Review the cost preview before approval. For the demonstrated 10-second 720p generation, the expected cost is 15 credits. If the interface shows a different value, stop and inspect the selected model, duration, resolution and output count because Google explicitly says costs can change.

Inspect the result before moving on. Check character age, facial identity, clothing, background, voice, camera position, lighting and unwanted writing. The first result in the tutorial looked younger than the reference and included burnt-in captions. Those failures should become explicit negative instructions in the next prompt or trigger a regenerated first clip if identity is too far off.

Prevent unwanted captions and generated text

Video models sometimes interpret a presenter scene as social content and add captions even when none were requested. Put the restriction in every prompt: no captions, no subtitles, no lower thirds, no logos, no labels, no watermarks and no text anywhere in the frame.

This instruction reduces the risk but cannot guarantee a clean frame. Review the first seconds, the lower third and the background. If text appears, regenerate while the prompt and reference are still fresh. Removing burnt-in text later can damage the face, clothing or background and may cost more time than a controlled retry.

Do not ask the generation model to create final typography for a branded short. Add accurate captions, titles and calls to action in an editor after the picture is locked. That keeps spelling, timing, accessibility and brand styling under direct control.

Step 4: connect clips with the last-frame technique

The tutorial uses a manual continuation method because Gemini Omni Flash did not support Extend in the creator's tested workflow. Google's current model documentation still lists Extend as coming soon for Omni Flash, while first-frame and first-plus-last-frame controls are supported on the web.

Open clip one in All Media, move to the final frame and choose Save frame. The saved image returns to the project as a reusable asset. Turn off Agent mode, set that image as the starting frame, attach the same character and voice, paste the second prompt and generate one new 10-second clip.

Repeat the process with clip two to create clip three. The saved image preserves the immediate pose and composition, but it does not lock everything. The tutorial's second clip became brighter. Repeat the lighting direction, exposure, white balance, background and camera wording in the continuation prompt to reduce that drift.

A clean handoff starts before the saved frame. Direct the character to stop gesturing, close their mouth naturally, face the camera and hold still at the end of clips one and two. A frame captured mid-blink, mid-word or during a hand movement forces the next generation to begin from an unstable pose.

Step 5: assemble and trim the 30-second sequence

Use Flow's Scenebuilder to add the three clips in order. Google's current documentation confirms that Scenebuilder can arrange and rearrange clips, trim clip boundaries, preview the complete sequence and download it as one scene.

Rename the clips before assembly, for example 01-hook, 02-explanation and 03-close. The creator accidentally inserted clip one twice, a small mistake that clear naming prevents. Trim dead air at each join but preserve enough breathing room for speech to feel natural.

Watch the sequence without sound to find lighting, pose and eye-line jumps. Then listen without looking to hear changes in voice, room tone or pacing. A subtle cutaway, screen recording, graphic or deliberate camera crop can hide a visible join while adding useful information.

Download the finished scene as one file. Google notes that Scenebuilder sequences must be downloaded before uploading through YouTube Studio because direct upload currently applies only to individual clips.

Design vertical shorts before generation

Flow supports both landscape and vertical generations. Choose 9:16 before creating the character shots when the destination is YouTube Shorts, Instagram Reels or TikTok. Keep the face, hands and any planned graphic space inside a safe central column.

Do not solve vertical delivery by cropping a finished landscape talking head unless the composition was designed for it. A later crop can remove gestures, make the face too large and leave no room for captions. Generate the presenter in the intended orientation and add platform-specific typography after export.

For a reusable series, lock the same safe areas, caption height, intro timing and end-card duration. The free daily credit is more valuable when each finished video drops into an established editing template instead of requiring a new layout every day.

Fix the problems shown in the tutorial

Wrong subject: attach or mention the correct character before approval and verify the reference chip is visible. Wrong voice: assign one voice to the character and reference it in each generation. Burnt-in captions: repeat the no-text instruction in every clip prompt.

Lighting jump: repeat the light direction, exposure and color temperature, then match exposure in the final editor. Hard cut: end on a neutral pose, trim carefully and cover the join with relevant B-roll or a designed transition. Character age drift: strengthen the age and facial identity language and use the same clean reference image.

Speech feels rushed: shorten the line rather than asking for faster delivery. Visual continuity is strong but the story is weak: rewrite all three beats together. Credits disappeared too quickly: confirm one output, test at 360p and read the approval cost before every generation.

Flow-generated video includes an invisible SynthID watermark according to Google's project documentation. Do not attempt to remove it. Add transparent disclosure when the synthetic presenter or voice could reasonably mislead viewers about who is speaking or how the footage was made.

Optional: automate the workflow with Claude Code and Flow MCP

Guney's companion video demonstrates a Claude Code workflow around the same process. It uses Flow Studio, also linked as Flow MCP, to let an agent prepare and run Flow tasks without manually repeating every browser step.

The linked repository is a community project under the GTAI-1 GitHub account. It is not presented by Google as an official Flow MCP. Review the repository, installation commands, authentication flow, permissions and update history before running it against a valuable account. Keep credit approval and final publication under human control.

Automation is most useful after the manual workflow is stable. First prove one character card, one three-clip prompt format, one continuity method and one review rubric. Then automate the repeatable parts while retaining approval before any credit-spending generation, download or publication.

A repeatable daily checklist

Before generating: check today's credit balance, current cost, model, resolution, orientation and one-output setting. Confirm the character, voice, three timed scripts, no-text instruction and still ending pose.

Between clips: save the true final frame, use it as the next start frame, restate continuity details and inspect identity, voice and light before continuing. Do not spend the third generation when clip two already needs to be replaced.

Before download: order and rename the clips, trim pauses, check the joins with and without sound, add any B-roll or exposure correction, verify captions in the final editor and obtain human approval before publishing.