Heatwave started in the Mojave. LA band, heat shimmer, a straight line back to Saidah's desert house in the present day. Outfits were 80s shimmer fabrics, skin-tight, revealing. Then I added glass panels and running water, because odd elements like that were everywhere in the 80s, and because water balanced the heat. The stream became a theme in its own right.

Then I asked Claude whether an 80s band could have afforded any of this for a debut single. No. So the desert became a soundstage.

That pivot fixed more than budget. On a soundstage everything is baked: the lighting, the props, the painted sky. That's what 1987 looked like, and it meant the video could be 80s rather than 80s through a modern lens. It also made generation easier, because sun and background were no longer variables. Sand dunes gave the flat set movement. Sheers did the work that practical heat shimmer would have done.

I deliberately watched almost nothing. One or two clips each of Whitney, the Pointer Sisters and Prince, then stopped, so the era's defining videos wouldn't redirect my own.

The edit was a puzzle I was solving for the first time. Which parts of the song, how long, two long passages or many fast cuts? Two things were fixed: the chorus, and Marcus's guitar solo. Everything else was found by ear. I was riding the excitement of seeing it come together and didn't see the elephant. That would matter on All That Glows.

Near the end, with the second song already in view, I pushed Heatwave into orange monochrome. Having one colour to hold gave the next video somewhere to go.

All That Glows is pre-grunge. New York's dark palette, functional clothes, everything muted, and a hilly Seattle feel borrowed from The Fabulous Baker Boys.

Heatwave had been a test of what the models could do. This time I arrived with a pipeline. I knew what I wanted to test, and I chose visuals that were powerful but easy to reproduce: a ballroom, and a floor so glossy it read as rippled water.

Once I had a ballroom generation I liked, I rebuilt it at 1:1. A 3D blockout with every major structural component. A wide camera angle the image model dressed with materials. Then the characters placed into it (image 2). That locked the model to the scene and its scale. From there, three hero angles gave me control over any shot the blockout didn't cover.

The problem was singing. My first cut used the opening 30 seconds, the chorus, and the final bridge to the end. Days of testing every model produced one that could take images and an audio clip and generate singing that stayed with the audio: Seedance 2. Everything else drifted or rewrote the sound.

Then the sax. Dozens of attempts, and Maxi would not play it. Such an easy human thing. But a model that has seen endless singing has seen almost no sax, and matching mouth, fingers and body to a horn line was beyond it. Seedance 2.5 landed a week before my final cut, and it was better, finger movement appeared, but the body language still lacked, and the film was already cut to the footage I had.

Fun fact: near the end I stumbled on a Prince video with a very similar theme. [TBC: which video]

All That Glows is pre-grunge. New York's dark palette, functional clothes, everything muted, and a hilly Seattle feel borrowed from The Fabulous Baker Boys.

Heatwave had been a test of what the models could do. This time I arrived with a pipeline. I knew what I wanted to test, and I chose visuals that were powerful but easy to reproduce: a ballroom, and a floor so glossy it read as rippled water.

Once I had a ballroom generation I liked, I rebuilt it at 1:1. A 3D blockout with every major structural component. A wide camera angle the image model dressed with materials. Then the characters placed into it (image 2). That locked the model to the scene and its scale. From there, three hero angles gave me control over any shot the blockout didn't cover.

The problem was singing. My first cut used the opening 30 seconds, the chorus, and the final bridge to the end. Days of testing every model produced one that could take images and an audio clip and generate singing that stayed with the audio: Seedance 2. Everything else drifted or rewrote the sound.

Then the sax. Dozens of attempts, and Maxi would not play it. Such an easy human thing. But a model that has seen endless singing has seen almost no sax, and matching mouth, fingers and body to a horn line was beyond it. Seedance 2.5 landed a week before my final cut, and it was better, finger movement appeared, but the body language still lacked, and the film was already cut to the footage I had.

Fun fact: near the end I stumbled on a Prince video with a very similar theme. [TBC: which video]