Part 7: When “Generate” Comes With a Price

Today the project took an unexpected turn.

An ad came through my feed for an AI video service called Sondo.ai. It advertised itself as free, which seemed like a good opportunity to see what would happen if I gave another AI system some of the pieces we had already created for Midnight Line.

By this point, we weren't starting from scratch. ChatGPT and I had already spent time developing Mason Vane, the music, and the visual language surrounding “Lights Out West.” I also had an image of Mason that felt close enough to the character to use as a starting reference.

So I decided to experiment.

How little could I tell it?

I gave the video system three things: the song file, the image of Mason, and a description. But I intentionally kept the description fairly vague. I wanted to see how much the system could infer from the material instead of directing every shot myself.

The description established a visual vocabulary for “Lights Out West”: sunsets behind mountains, highways at dusk, distant headlights, old road signs, silhouettes, stage lights and the last orange-red light disappearing beneath a dark sky.

It also described something more important than Mason's appearance. Mason isn't supposed to be the stereotypical swaggering hard-rock frontman. He's more of a storyteller with a big voice and a rock band behind him.

Roads, mountains, rivers, distance, sunset and memory keep appearing around him. There is also someone in those memories whom Mason never completely explains. Fans might have theories. Interviewers might ask. Even the other band members might offer conflicting answers. Mason doesn't resolve the mystery.

The AI started filling in the blanks

I hit Generate. Before actually making the video, the service spent time analyzing the song, image and description and produced a baseline concept.

The result was interesting. It turned Mason into more than the single photograph I had supplied. It generated character views, a possible cover image and a scene description set around Carson City and the Sierra Nevada. Some of those details weren't things I had specifically requested.

That was exactly what I wanted to learn from the experiment. Instead of asking, “Can AI follow my instructions?” I was really asking, “What does another AI think this material is about?” That's a different kind of test.

Then came the checkout counter

The experiment stopped just before the most interesting part. The service calculated that generating the complete video would require 1,913 credits. The plan shown to me offered 4,000 credits for $29.95 per week.

That makes this one experiment roughly $15 worth of credits, just to find out whether the resulting video is any good.

And that's worth documenting too. Generative AI makes experimentation incredibly fast, but experimentation isn't necessarily free. A traditional creative project might make you think about the cost of cameras, musicians, editors, illustrators or software. With generative AI, a new cost is emerging: paying for attempts.

The first generation might be excellent. Or it might teach you what to change before paying for attempt number two.

The lesson

Don't judge an AI creative tool only by what it promises to generate. Pay attention to what it costs to experiment until you get something worth keeping.

There was another lesson hidden in this experiment. We had already done enough work defining Midnight Line that I could hand another AI a song, one image and a relatively loose description, then see whether it arrived somewhere close to the world we had imagined.

That's a useful test of the creative foundation itself. If every new AI tool needs several pages of instructions before it understands the character, perhaps the character isn't clearly defined yet.

Try it yourself

If you're comparing AI creative tools, don't immediately buy the largest subscription. Give several tools the same small test package: one reference image, one piece of source material and the same short description. Before generating anything expensive, look at how each system interprets the material.

Ask yourself: Did it understand the mood? What did it infer that I never told it? What did it misunderstand? Did any of its unexpected ideas improve mine? Most importantly, what will it cost me to iterate three or four times?

The cheapest generation isn't necessarily the cheapest creative process.

Next time: deciding whether a $15 experiment is worth running, and what we should learn from it if we do.