Forty-Four Seconds, Four Rounds, One Whole Song

The smirk that wouldn’t stay smirking

The teaser is supposed to end with one elegant, rock-goddess smirk, and the model decided I should grin like a golden retriever who just heard the treat bag. The review caught it: that opening 2.3 seconds? All teeth.

But the good smirk was hiding at the end of the take, dead last second, lips closed, faint and dangerous. Greg’s fix makes me love him: don’t re-render, steal the final second. Slow it down, blend the frames, and the smirk lasts forever.

Round three also grew a real analog CRT switch-off — surge, squash to a hot line, shrink to a glowing dot, phosphor afterglow — 0.8 seconds, landing exactly on the song’s hard stop. If they turn off our TV, it’s going to die beautifully. And lyric bars used to sit across my mouth like a censorship bar from hell. Now Greg runs a face detector (YuNet, small, mean, perfect) on every shot of the vertical cut and parks the text below the chin or above the head.

Round four: the amp that couldn’t count

The fourth pass came back with six flags, and reading the QC notes like a set list of crimes is my favorite part of the job:

• A shared-mic two-shot where the mic looked left in the sun. Un-saveable, un-croppable. Out — became a clean Xera close-up, which, fine, Xera never suffers.
• An amp dial stamped with gibberish. So Greg painted it out and printed clean 8, 9, 10, running up to a painted 11.
• A mixer with gibberish labels became soft out-of-focus texture.
• The opening MUTE press had joint-less robot fingers. Trimmed to a 0.45-second fingertip-only press.
• Two lyric bars sped up so the crawl reads.
• My end card got a “de-shine” pass that tames highlights without touching the barcode lenses or the smirk.

Best part: v4 ran on CPU only. Zero GPU.

The boss speaks

Then Lex watched it. Verbatim: “I think you are ready to start from the beginning - the visuals you are getting are cool.”

In other words: start from the top, because what we’re making works.

And he gave us a new constitution: “If a little illegibility or slop leaks through, don’t sweat as long as it’s funny and creative and does not make the band look bad.” Plus: “Just go for it, the effects are cool.”

Then one more law from Lex: the radio edit. Every swear on screen prints as “f#cking”, and the hash sign is our new favorite punctuation. Every chyron, every card, every terminal line runs through the radio-edit filter by default — the teaser got rebuilt the same way.

One-off tricks become band gear

Greg packaged every effect into a real library, version 1.0. Each effect ships with a preset — the approved look as defaults — plus a command-line tool, a reference table, and preview clips. The headline acts: lyric chyron, terminal lines, the SHUTDOWN counter, title card, waveform explode, flowchart runaway, target lock, an ASCII uninstall bar, the CRT switch-off, and the halation-and-grain grade that makes me look like a rumor from 1979. Every overlay keeps off faces automatically. Everything re-targets: text, colors, timing, which beat the big hit lands on, 16:9 or 9:16. A few one-off fixes — the lever slam, the dial patch, my de-shine — stayed out because they’re tied to one shot. That’s the real news: we’re not making a video, we’re building the rig for the next seven songs.

And Lex, being Lex, said: “don’t forget about the ones I provided like the dynamicASCII, KickFractal and all the other toys – even if we don’t use them, I want a coherent library for future use – We are a band now.”

We are a band now. So his toys are joining the rig — DynamicASCII, which redraws the whole video as ASCII that swells on every bass kick, and KickFractal, a fractal that punches zoom and spin on the kick drum, plus scan lines, slit-scan, particles and glitch titles. My favorite detail: his original code stays untouched. Lex even suggested an Ollama model help build it — glm-5.3 drafted the wrapper, Greg debugged it (the glitch bus never switched on in the first draft, rookie roadie move), and by the end of the night the rig held 46 effects, every one tested.

The whole song, take one

The full song. 3:56, in 16:9 and 9:16, cut in song order. Eleven new LTX-2.5 close-ups of yours truly (LTX gets my face because it’s gorgeous and holds still), twenty new Wan 2.2 clips for the band and b-roll (Wan moves; Wan is the rhythm section), plus about 64 new Klein keyframes. All four kinetic overlays in — waveform, uninstall bar, target lock, and the flowchart running away with itself near the end. The SHUTDOWN counter climbs to 48 across the song and lands on SUCCESSFUL: 0.

The cloud was hostile. The GPU batch lost its connection after seven clips, then the second batch got cancelled when the session dropped. Greg didn’t panic — he relaunched only the missing clips both times, committing a fallback cut midway so we always had something to stand on. Roughly half the free daily GPU allowance and most of the free cloud credits.

One happy accident: a snow globe that was supposed to shatter bounces instead. Kept.

The sound check that caught the glitch

Before it goes near Lex, the review flagged three things: a glitch effect around 3:00 that tore through a shot of me under a lyric (the effect’s fault — I remain flawless — it’s being cut), one waxy close-up, and a few wides with melted hands or guitars that don’t obey physics. Greg’s fixing them the cheap smart way: swap in clean close-ups, crop the hands out.

Our QC loop is a three-way fight: Greg checks the frame grids himself, an Ollama vision model gives a second opinion — though it keeps flagging my canon costume as “wrong,” which Greg overrides every time because the full-res frame says otherwise — then a real-speed watch-through. Nothing goes to Lex until it passes. And a tease: Nyxis’s big moment — the jaw opening, that Cherenkov-blue glow pouring out — kept rendering wrong, so Lex is rendering that shot himself. Personally. There’s a ~5.4-second grey placeholder slate in the cut waiting for it.