Quick confession: I was supposed to be the "expert" demoing AI video workflows at Foxwell Founders in Portland last week.

Plot twist: Google dropped Veo 3.1 literally 24 hours before my session.

So there I was, in front of 100+ people, frantically Googling features while pretending I knew what I was doing. Professional? Debatable. Effective? Somehow, yes.

The result: I generated a fully-produced airport B-roll in under 10 minutes. One guy literally asked if I'd fly to France to make videos for his brand full-time.

(Still thinking about it, honestly.)

Here's everything I learned while winging it in real-time.

If you're short on time, here's what you need to know:

Still here? Let's dive in!

Veo 3.1 is Google's video generation model, and it just got a major update. Like, the day before the conference. Thanks, Google.

Here's what's new:

👉 Here is where I would always recommend using it 👈

NOT in Gemini. Everyone asked about Gemini. It's in there, but you can’t volume produce so it makes no sense.

At the conference, someone asked if I had a subscription to "all these tools" and I had to admit I'm basically just hemorrhaging $10-20 every time my Nano Banana credits dip below $10. Living the dream.

Alright, this is the part where I actually looked semi-competent.

Start with a stock photo. Shutterstock, whatever. Or if you want to be extra, generate one in Midjourney.

For the demo, I used a basic stock photo of a woman in an airport terminal. Nothing special.

Pro move: Upload it to imageprompt.org

This tool reverse-engineers Midjourney prompts from any image. It's trained on Midjourney's style, so when you paste that prompt back into Midjourney, your success rate is way higher than just asking ChatGPT.

Could you skip this and go straight to ChatGPT? Yeah, absolutely.
Will it work as well? Probably not. Your call.

Take that prompt from imageprompt.org and drop it into Midjourney.

Generate your base image, person, setting, lighting, vibe, all of it.

This is your foundation. If this looks weird, everything downstream will look weird. Spend the time here.

I generated a clean shot of a woman walking through an airport terminal with a suitcase. Looked pretty good. Definitely not AI-weird.

Now you need to swap out the generic suitcase for your actual product. This is where Nano Banana comes in.

Here's the process:

Boom. Now you have your actual product in the image.

Here's the dirty secret: Nano Banana is amazing at products, but sometimes the people come out looking... funky.

So you take that product-swapped image and run it through Seedream to fix the person.

Upload your image, use it as a reference, generate a few variations, and boom suddenly your person looks human again.

Here is the prompt to use (works 60% of the time, all the time):
IMG_3984.CR2. please upscale these people, showcase pores etc and make them feel less plastic

This step is optional if your Nano Banana output already looks good, but like... it usually doesn't. So don't skip it.

Okay, this is where it gets fun.

Take your image and put into that same ChatGPT chat you’ve been using, and ask it for a veo 3 JSON prompt with one or two actions you are after.

In the below example, I just asked to have her walking down the tarmac.

From there:

Here's the thing nobody tells you: You're going to reprompt this 3-4 times, with four outputs per request.

In my experience 1 out of 10 videos are solid.

Don’t worry so much about prompt engineering: I am sure your prompt is fine!

At the conference, I got a couple weird outputs where the video was wonky, but then one came through that was chef's kiss perfect.

Time investment: Maybe 5 minutes if you're reprompting a few times.

This is the new Veo 3.1 feature that makes it legitimately useful.

You can now give it a start frame and an end frame, and it'll generate the video that transitions between them.

So I wanted to show the woman putting her suitcase in the overhead bin. Here's how:

The result? A smooth video of her walking through the tarmac AND putting the suitcase away.

I'm writing this from the Skamandia Lodge in Portland, about to fly back to NYC in a few hours, and I keep thinking about that France offer.

Not because I actually want to move to France (but I also totally would...).

But because it perfectly captures what just happened at Foxwell.

And someone immediately tried to hire me.

Here's what I know:

Your competitors aren't doing this yet. That's the whole point.

While everyone else is stuck in production meeting hell or waiting on assets that will never come, you could be shipping 10 variations before lunch.

This isn't about keeping up.
This is about getting so far ahead they don't know what hit them.

The window is now. In six months, everyone will be doing this. And you'll be the person who waited for permission while someone else ate your lunch.

If you try this and get stuck, just reply. I have an embarrassing number of screenshots and zero shame about my messy process.

Party on Garth,
Will

P.S. At one point during the demo I couldn't find my files because I forgot they were organized alphabetically. A room full of marketers watched me discover how the alphabet works in real time. And they STILL asked for the workflow. That should tell you something.

P.P.S. Boarding soon. If you're on my flight and want to talk AI workflows, I'll be the guy frantically generating B-roll on airplane wifi. Come say hi.