AI Brand Video:
How We Made Meet Pip, and the Six Things That Broke

Meet Pip is a 2 minute 48 second AI brand video we made on 30 and 31 October 2026 for Mailbox MCP, the service I built this autumn that connects your existing mailbox to an AI assistant. Every moving picture of Pip in it came out of an AI video model. Every word on screen, every cut and every level in the mix was decided by a person.

I set one rule before we started: utterly brilliant or not at all. HD from start to finish, smooth transitions, no sudden cuts. If the tools couldn't get there, the film wouldn't go out. They got there, but not on their own, and six things broke along the way. Here's the film first, then what it really took, for anyone weighing up an AI explainer video for their own business.

2:48

Pip's backstory, and what the product does, is in our post on where Mailbox MCP stands two months in. This one is about the making of the film.

How we made an AI brand video, step by step

The order matters more than the tools. We locked the words and the voice first, because the voice sets the length of every shot. A beautiful clip that doesn't fit the line it sits under is just a beautiful clip.

1. Script. Pip's lines, written to be spoken, each with a caption version.
2. Voice-over, timed word by word. Those timings gave every shot its length.
3. A first frame for each shot, made with Gemini 3 Pro Image using our Pip master image as the reference, so he looks like himself in all 17 shots.
4. Each first frame animated into an 8-second clip with Google's gemini-omni-1.1-flash video model.
5. On-screen graphics and captions built as a web page in HTML with GSAP, then rendered frame by frame with Playwright into transparent PNGs.
6. A lo-fi music bed from Google Lyria 3.5, ducked 8 dB whenever Pip speaks, plus foley for what happens on screen.
7. A frame-exact assembly in ffmpeg, with a dissolve between every pair of shots.
8. Loudness at -14 LUFS, and en-GB subtitles written out as an .srt file from the same cut.

Step 5 is the decision I'd make again first. AI video models still can't be trusted to spell, so nothing a viewer needs to read came from one. The chat window, the calendar, the prices and the captions are a web page we animated ourselves. That means the £34.99 on screen is exactly £34.99, and if the price ever changes it's one line of HTML and a re-render, not a fresh roll of the dice.

The quality bar for our AI brand video: HD, smooth dissolves and no hard cuts

"Utterly brilliant or not at all" sounds like a slogan. In practice it became three rules we could check: HD throughout, smooth transitions, and no sudden cuts. A hard cut between two AI clips is exactly where a viewer notices that the light has shifted or Pip's arm is somewhere new, so every join in Meet Pip is a dissolve of 12 frames, half a second at 24 frames a second. Wyzowl's 2026 survey found that 89% of consumers say video quality affects how much they trust a brand, and a brand film that jolts is an odd way to ask for trust.

This is the line in our assembly script that does it. The dissolve borrows clean footage from just past the outgoing clip's out-point and just before the incoming clip's in-point, so each cut stays exactly where the voice needs it:

DISSOLVE = 12   # frames: half a second at 24 fps
chain.append(f"[{cur}][p{i}]xfade=transition=fade:duration={d / FPS:.6f}:offset={off:.6f}[{nxt}]")

The other half of the quality bar is choosing. We generated 60 takes for 17 shots, about 482 seconds of footage, and roughly 96 seconds of it made the film. One second in five. The rest of the 2:48 is graphics we built ourselves.

We're not unusual. For Coca-Cola's 2025 Christmas campaign, TheWrap reported that "a mere team of five AI specialists was hired to sort through more than 70,000 clips in a 30-day span". The ratio moves with the budget. The selecting doesn't go away, and it's the part a demo reel never shows you.

Six things that broke while making the AI video

None of these were in a tutorial I'd read. All of them cost time.

1. Google shut down Veo 3.1 partway through the project

We'd planned the clips around Veo 3.1; our clip script is still called veo.py. Then Google's documentation changed underneath us: the Veo 3.1 Preview models "in the Gemini API was shut down on October 22, 2026", with gemini-omni-1.1-flash named as the replacement. So the clips moved to Omni.

The lesson isn't about Google. Any AI video model is a rented tool with a shelf life, and a project that runs longer than that shelf life needs its decisions written down. Every shot prompt lived in one file, every first frame and every take was kept, so changing model meant re-running the clips rather than starting again.

2. Asking for a "talking" robot produced gibberish

Pip is a presenter, so the obvious prompt was a robot talking to camera. The model took it literally. Mouth shapes appeared on a face that's meant to be a glowing line of light, and the clip's own audio came back with an invented voice babbling in no language at all.

The fix was to stop asking for speech. Pip's real voice comes from the voice-over, so the clips only needed actions: a wave, a shrug, a reach for the cape. This is part of the motion instruction that went into every clip after that:

Pip's face is ONLY two glowing blue eyes and a glowing blue LED smile line on his
black visor: the light line moves and changes shape as his expression changes, he
never has a mouth opening, teeth, nose or eyebrows. [...] No speech, no dialogue,
no voice, no singing. Sound: only light, cheerful sound effects of what happens on
screen.

Describe what the character does, never what it says. The voice is a separate job for a separate model.

3. The first AI voice-over was wooden

Our first voice was OpenAI's text-to-speech (the gpt-4o-mini-tts model, its "fable" voice), with careful direction about warmth and comic timing. My verdict on hearing the cut was two words: "very wooden". It said the lines. It didn't perform them.

So we re-voiced the whole film with ElevenLabs on its eleven_v4 model, which ElevenLabs describes as "our most emotive, high quality speech synthesis model". I heard a sample first. My reply was "It's great!!", two exclamation marks and all, and the film was rebuilt around the new voice: new word timings, new shot lengths, re-timed captions and chapters. That rebuild is why the finished film runs 2:48.

If you're making your own, this is the cheapest place to be fussy. A viewer will forgive a slightly odd frame long before they forgive a flat voice they have to sit through for three minutes.

4. An AI reviewer saw extra teeth and a third hand

We asked Gemini to watch the cut and critique it, which is quicker than a human pass for a first look. It came back with alarming notes: extra teeth in one shot, a third hand in another. Pip doesn't have teeth. We checked both shots frame by frame and neither fault was there.

Those are exactly the faults people now hunt for in AI video. Lindsay Brillson, Head of Brand and Content at the AI video company Pika, described them in an interview with Tom's Guide on 1 October 2026:

"There are the explicit tells that we're all familiar with, like hands and faces morphing, lip-sync misaligning with audio, and text that's almost a word, but not quite. But these issues are becoming less and less prevalent with model advancement. That leaves more subtle cues that are more an indictment of the people using the technology than the technology itself. It shows that no one is putting attention and craft into the work."

Lindsay Brillson, Head of Brand and Content at Pika, Tom's Guide, 1 October 2026

I agree with her, and our reviewer proved her point from the other side. It had learned the list of tells so thoroughly that it found them where they weren't. Brillson's "indictment of the people" line is the one that stuck with me: the craft she's talking about is a person's job, and handing the checking to another model just moves the problem.

So we kept the AI review as a lead and trusted two dull, objective scans for the verdict: the brightness of every frame, which finds black frames, and the difference between each frame and the next, which finds jumps. Those scans found real problems that the AI reviewer had missed, which brings us to number five. The subtler cues she describes, light that drifts between shots and camera moves with no reason, no scan catches. A person watching does.

5. Black flashes came from mixing RGB and RGBA frames

The brightness scan found single black frames flashing past, too quick to see consciously but enough to feel like a stutter. The cause was in the graphics layer. Chromium writes a fully opaque screenshot as an RGB PNG and anything with transparency as RGBA, so our frame sequence switched pixel format partway through, and ffmpeg's overlay dropped those frames. The fix was one line in the renderer: force every frame to RGBA.

# always RGBA: Chromium writes fully opaque frames as RGB, and a sequence that
# switches pixel format mid-stream makes ffmpeg's overlay drop those frames
Image.open(io.BytesIO(shot(pg, f / FPS, layer))).convert("RGBA").save(dest)

No AI reviewer flagged it. It took a script that looked at numbers.

6. YouTube played the first upload at 240p

Our first upload, a 1080p file that still carried the original OpenAI voice, played at a blurry 240p by default. YouTube's own help says that "when you upload a video, it will initially be processed in low quality", with higher qualities added once processing finishes. So when we re-voiced the film, we re-uploaded it from a 1440p master instead, and the new upload, the one at the top of this page, played in HD. The first one is now private. My 365i colleagues have the full story of why a YouTube video plays at 240p and what to upload instead.

One master, three cuts: the brand video for YouTube, X and LinkedIn

A brand video that only lives on YouTube is half used. From the same master we cut three versions:

Captions are burned into the social versions because those feeds are mostly scrolled with the sound off. On YouTube they're a separate subtitle file, which viewers can switch off and machines can read. That last part matters more than it used to: to an AI crawler, an embedded video is an empty iframe with nothing in it, so the words have to exist somewhere it can read them. How AI answer engines read a video is its own subject, written up on AI Visibility as how to make a video legible to AI search.

What AI video still can't do reliably

I'm proud of Meet Pip. I still wouldn't hand an AI video model a brief and walk away, and Google's own model card for the model we used says why:

"While Gemini Omni Flash demonstrates strong progress, maintaining complete consistency throughout edits, generating scenes with complex motion, or rendering perfectly accurate text remains a challenge."

Google DeepMind, Gemini Omni Flash model card

Three limits in one sentence, and we met all three. Consistency: we only got a Pip who looks the same in every shot by starting each one from a still made with the same reference image. Complex motion: the cape shot took five takes to get one where he actually takes the cape. Text: we didn't let the model write a single word. Google is being straight with its users there, and I'd read that sentence before believing any demo reel.

The rest of the honest list, from our own two days:

• Clips are short. Omni makes 3 to 10 seconds at a time, so a 2:48 film is an edit of many generations, never one prompt.
• Speech inside the clip isn't usable. Ask for talking and you get babble. We wouldn't rely on lip sync to a separate voice either, which is one reason Pip has a light for a mouth.
• Realistic people are a different job. We chose an animated robot on purpose. YouTube says "we require creators to disclose AI-generated or -altered content that seems realistic"; a cartoon robot doesn't need the label, but we'd rather say how the film was made, which is what this post does.
• Taste isn't automated. No model told us the first voice was wooden or that a hard cut felt cheap. Somebody has to sit through the whole thing and mind.
• You still own the result. Whoever publishes the film answers for what's in it, as we argued in AI made your content, you're still liable.

AI brand video vs an agency explainer video

Neither route is always right. This is how they compared for a film like ours, a character explaining a software product.

The price guide is Gilkes Media's 2026 UK figures. If your film needs your actual team, your shop or your product on a real table, film it. An AI version of a real place would need a disclosure label and would still look slightly off to the people who know it best: your customers.

Should a small business make an AI explainer video?

Video isn't optional any more. Wyzowl's 2026 figures say 91% of businesses use it for marketing, 63% of video marketers have now used AI to create or edit video (up from 51%), and 96% of people have watched an explainer video to learn about a product or service. The real question is whether the AI route suits yours. If you're working out how to make an AI video for your own business, here's what we'd tell a client after making Meet Pip:

• Pick a story that suits a character or an illustrated world, and somebody who can write the script and judge the result.
• Lock the script and the voice before you generate a single clip.
• Keep every word on screen out of the AI model.
• Budget for takes. We used about one generated second in five.
• Choose a voice you'd happily listen to for three minutes, then listen for three minutes.
• Join shots with dissolves, and scan the finished file for black frames before anyone else sees it.
• Check the upload the way a stranger would, in a private window.

Then give the film a page worth landing on. This one shows a still and loads YouTube's player only when you press play, so the film costs the page nothing until someone asks for it; that's the kind of detail our speed optimisation work is built on, and the page around a film matters as much as the film, which is where our web design comes in. Google cares whether content helps people, not which tool made it, which is the line we drew in where Google draws the line on AI content. And if you want to see what Pip actually does for a mailbox, 365i explains it in keep your mailbox and add an AI assistant.

Two days, 60 takes, one wooden voice and a lot of frame-by-frame staring. Pip got his film. I'd do it again, and I'd still check every frame.

Frequently asked questions about AI brand videos