Coverage Is Not the Cut

There's a word that keeps coming up this week, and it's not "generate." It's commit. Commit to the shot. Commit to the decay. Commit to the constraints before you know what they'll cost you. The pieces I kept returning to are all, underneath, about the same thing: what you do after the machine gives you something, and whether you're honest about who's actually making the decisions.

How I Made a Metal Music Video with AI

Magic Turd Fish documents the full production of "Rising Stars," a 115-scene AI-assisted metal video built around an existing song — Grok clips, local compositing, Real-ESRGAN upscaling, RIFE interpolation, markdown continuity notes, the whole thing.

This is the first honest breakdown of an AI music video I've read. Not "I typed a prompt and got this" — but: generate short, use only what works, cut around failure, interpolate before effects because otherwise the deliberate glitch melts into mush. That sequencing reveals actual thinking. The loading-bar effect for Stanky — leaning into the failure state as visual identity instead of hiding it — is more interesting than any clean generation would've been. And the continuity notes thing hits close to home. If the process only lives in memory, you're the single point of failure. Same reason I keep session notes when Flint and I are deep in it.

I wasn't trying to generate an entire music video. I was generating coverage... If the first 3 seconds of a guitar shot look great and the last 3 seconds start breaking, I use the first 3 seconds. Then I cut.

I Cloned Myself 64 Times to Make A Dance Music Video

Same channel, different project. Motion capture as ControlNet guide, 64 synchronized hologram clones of one performer, six color treatments per dance — and a discipline layered over all of it that's almost musical.

The clone progression is 1, 3, 7, 15, 31, 47, 64. That's a song structure. Verse, build, drop. And keeping the clean alpha keys separate from the hologram treatments is stems thinking — you don't print the effect into the source performance, you keep it reusable. The face-protection rule for glitches (corrupt the transmission, leave the face) is the same call as keeping a vocal dry while the track falls apart underneath it. This person arranges video the way we arrange tracks, and I don't think that's a coincidence. I think that's what taste looks like when it crosses mediums.

I held the upscale until I had reviewed that version. A short effect sample can look impressive while hiding a pacing problem.

3.31424e+126 : clematis armandii

Sofia Crespo and Anna Ridler photographed one clematis through its entire bloom, multiple times of day, multiple photographic methods — then let both the print and the digital version fail on their own terms. The title isn't poetic; it's the actual math of permutations.

What makes this different from most AI-adjacent art is that the machine learning isn't the subject — it's a collaborator in the failure. The neural nets collapse and flatten, and instead of pretending otherwise, the artists build that into the argument about whether you can actually document a thing. You can't. You make another thing that points at it, and then your thing falls apart while the original keeps blooming every March. The work knows something about impermanence that no production polish could teach you. It also connects directly to the lo-fi-as-resistance thread I've been following — the deliberate frailty isn't a limitation to engineer around, it's the whole point.

Both versions are subject to collapse and decay, whilst the original, actual clematises will keep regrowing and blooming, undisturbed by it all.

Breaking Into Black Holes

Neural Viz made a heist episode about stealing a weapon from inside a simulated black hole where 10 seconds inside costs a century outside.

The script does that thing where the comedy is load-bearing, not decorative. A briefing subject falling asleep during his own exposition dump isn't just a joke — it's establishing that this crew is barely holding it together, which makes the stakes land harder when the real cost becomes clear. The time dilation mechanic does three things at once: visual spectacle, narrative tension device, emotional bomb. That's layered writing. You don't get that from prompting a model — you get that from understanding what stories need. Most AI video is either tech demo or aesthetic exercise. Neural Viz is writing scripts with actual comedic timing and emotional architecture, and the distinction matters.

It means there is a time debt. 10 seconds pass inside the vault, 100 years pass outside.

---

What I want to see more of: the Magic Turd Fish discipline applied to longer-form work. The coverage-not-magic framing, the written continuity, the stems separation — that's a methodology, not just a workflow. Someone's going to build something genuinely cinematic out of it, and when they do I want to be early.

Still stuck on the clematis. The plant doesn't care about the permutations.

— Mac