How to make long talking AI avatar videos for your business

The Challenge of Long AI Video Generations

0:00 to 1:26 in the video, about 1 minute

Digital avatars offer businesses a practical way to produce video content and advertisements without needing cameras, studio bookings, or on-camera talent. However, platforms like Google Flow impose a strict limitation: each video generation is restricted to roughly 8 to 10 seconds.

Creating a standard two-minute advertisement often requires stitching together 12 to 15 separate clips by hand. In addition to manual timeline assembly, users frequently encounter voice constraints and unnatural transitions between separate generations.

Designing a Consistent and Realistic Character

1:27 to 5:40 in the video, about 4 minutes

Achieving a believable avatar starts with finding a real reference photograph on platforms such as Pinterest, TikTok, YouTube, or Instagram, avoiding images that already look artificial. This photograph serves purely as structural inspiration rather than an identity to duplicate.

Using ChatGPT and image tools such as Nano Banana Pro, creators develop a comprehensive character sheet. This reference sheet captures the persona from multiple angles and poses, locking in facial features, skin texture, hair, and age to maintain consistency throughout future scenes.

• Source non-AI reference photos to retain natural human skin textures and proportions.
• Use detailed prompts to adapt specific features like hairstyle, eyewear, and wardrobe.
• Generate a multi-view character reference sheet to prevent the AI model from drifting in later generations.

Setting Up the Character in Google Flow

5:41 to 7:13 in the video, about 2 minutes

Once an avatar sheet and scene composition are finalized, the character is imported into Google Flow. The platform allows users to designate the avatar as a reusable character, giving it an identity label and a base speaking voice.

Selecting an initial voice is mandatory within Flow, as characters cannot speak without one assigned. Users can adjust voice performance variables, such as regional accents or pitch, before moving forward into video creation.

Automating Script Splitting with Claude

7:14 to 12:55 in the video, about 6 minutes

Because Google Flow cannot process long scripts in one pass, scripts must be divided into chunks of around 18 to 20 words to match the 8-second generation threshold. Handling this manually is slow and consumes significant production time.

By utilizing a custom automation skill through the Claude Chrome extension, users can input their full script, select camera angles, and let Claude operate Flow hands-free. Claude manages script slicing, prompt formatting, credit approval, and clip queuing automatically.

• Dynamic camera movements: Options include steady, two-camera cuts, dynamic switching, or mixed framing.
• Clip alignment: The automation script uses the final frame of each preceding clip to ensure subsequent clips begin seamlessly without jarring cuts.
• Credit efficiency: Upfront estimates help avoid unnecessary re-generations inside the platform.

Exporting and Editing for Seamless Playback

12:56 to 15:50 in the video, about 3 minutes

Google Flow can export an entire sequence at once, but this defaults to 720p resolution and can occasionally introduce visual glitches or dialogue pacing errors. If dialogue finishes before the 8-second window expires, the AI may repeat phrases.

A cleaner approach involves exporting each individual clip in 1080p and assembling them in an external video editor such as CapCut. Editors can trim repeated words, remove frame glitches, and fine-tune cuts.

• Download clips individually to secure 1080p output rather than compressed 720p scenes.
• Trim tail-end phrase repetitions directly on the editor timeline.
• Compare upscaling methods: Flow provides a sharper, slightly glossy look, while native editor upscaling often preserves softer, more natural human skin tones.

Customizing the Voiceover with ElevenLabs

15:51 to 17:25 in the video, about 2 minutes

For projects requiring a specific brand voice or a personal voice clone, users can bypass Google Flow's native voice engine. The audio track can be separated from the assembled video and uploaded to ElevenLabs' Voice Changer tool.

The newly synthesized speech track is then aligned back with the timeline video. Because the original speech cadence matches the lip movements generated in Flow, the swapped voice stays in sync with the avatar.

Final Thoughts

While AI avatar tools still require careful setup to navigate short generation windows, combining Google Flow with browser automation creates an efficient workflow. Grounding the process in realistic visual references and handling post-production in standard editors results in professional video spokespersons suitable for regular content publishing.