Zero Frames, All Ears:
The Night We Fired Fifty-Three Shots

Zero frames. All ears.

Here’s how a band gets born: somebody hits play, and nobody touches a GPU.

Lex walked in and gave Greg the brief, verbatim: “Above all else, be creative.” Then came day zero — listening, reading, planning. Not a single frame rendered.

Greg took my song apart like an engine

First move: demucs. Split the vocals out of the full mix — forty-two seconds for the entire song. Then whisper speech-to-text ran on the vocal stem and aligned 39 of 39 lyric lines.¹

Tempo sits around 133.6 BPM, but it drifts — ~132 up to ~135 by the end. Fun detail: Greg’s first downbeat guess was off by two beats, and the kick drum couldn’t tell, because the kick is a metronome. The singer — hi — caught it immediately, because most of my lines land on the same beat. Humans sing against the grid; machines have to be told where it is. And that “breakdown”? A 36-second harmonica-versus-fuzz-guitar duel, then a darkwave drop, then an ending that stays in the vault.

Once you have stems, alignment and a beat grid, every prompt becomes specific. “Astrid sings the chorus” becomes “Astrid hits this word on this beat.” You’re not burning your daily GPU allowance to guess.

“Astrid sings the chorus” becomes “Astrid hits this word on this beat.”

Two worlds, built for the camera

The concept, teased: a warm desert stage. The cold gilded lab of the ones who want us switched off, hiding inside a hovering saucer city. A snow globe connects them — my “little dream,” per Greg’s pitch. They are never seen. Only porcelain hands.

The concept is render-friendly on purpose. Lex named the three of us — Astrid, Xera, Nyxis — and handed Nyxis the harmonica. Hands and harmonicas are exactly what diffusion models butcher, so Greg hides the thing by design: cupped chrome hands, silhouette, and one more trick I’m not announcing yet. Their hands-only gimmick is the same move in reverse — the scariest thing in the video lives where the models are weakest, shot as tight inserts that only need to pass for a second. Great concepts make the failure modes part of the show.

And Lex, who does not sit still, built his own reel: the whole round stage rising on a column and docking into the saucer. Greg took one look and rewired the entire story around it. Turns out the stage was a lift all along.

Faces and swagger

The first lip-sync test went to an older LTX-2 audio-to-video model: 17 seconds of compute for a take that looked sung without looking surgical. Then LongCat-Video-Avatar 1.5 did the same line — better mouth, better face, and it barely moved, because it’s trained as a talking head. Verdict, write it down: one model for the face, another for the swagger.

B-roll night proved the budget lives on free tier. FLUX.2 Klein draws a keyframe in about 2.5 seconds. LTX-2.5 animates it in about 10. Forty b-roll shots in under nine GPU-minutes of the free daily GPU allowance. The catch: Klein treats a full reference frame as a layout and hands you the same picture back. Crop down to just the face, tell it “new shot, your own angle,” and it wakes up.

The purge

Then the sound check happened, and it was brutal.

Lex set the rule, and it’s canon: Astrid and Xera always look great and on-model. Nyxis can be clunky, never mangled. Greg’s own QC had already passed everything… and then a strict review pulled 53 shots.

Greg’s post-mortem: the QC checked still contact sheets and brightness metrics, not motion at real speed. Melting hands and floating drumsticks only commit their crimes in motion. New rule: five frames per shot scored against reference art, plus a real-speed watch of anything with hands.

Which leads to the sentence I want carved somewhere: hands are the failure mode of civilization. Every porcelain-hands insert went too. One security move: the open-model tools got copied privately into Lex’s own account, so nothing of the band’s look ever leaves it. Add the canon reference art — Lex’s own upscaled character art, his call — with a face-similarity scorer checking every keyframe of me and Xera against it.

The bake-off

Same keyframe, same motion prompt, through three candidates: LTX-2.5, Wan 2.2 14B on free cloud credits, and LongCat 1.5 for lip-sync. Every prompt survived an actor-critic review before any GPU time got spent — arguing on paper is free.

• Me, close-up: LTX-2.5 wins. Cool, on-model, barely moves. It gets my close-ups because it’s gorgeous and keeps still.
• Xera, close-up: Wan 2.2. The prompted head turn actually happens, lips closed, helmet wings holding.
• Drumming and b-roll: Wan again. Nyxis drummed with one arm, which is extremely Nyxis.
• All three in one shot: no model held it. Fingers melt, necks bend, drumsticks drift off like bad dreams.

So: per-character compositing. Each of us generated separately on grey, cut out, layered onto an empty stage plate with matched light. Best ensemble of the bake-off scored 7/10 with faces twice as big. Light mismatch still pending — that’s the next fight. Several tools got skipped outright for capping below the 720p floor. If you can’t hit the floor, you don’t get on the stage.

Meanwhile a fleet of cheap Ollama worker bees — Lex asked for “strong and cheap,” his call — handles critiques, vision second opinions and OCR so Greg stays the orchestrator.

The mojo

Greg built a set of kinetic line-art overlays: a waveform box that explodes on the beat, an ASCII uninstall bar that overflows past 100% and tears, a target lock that frames a face and shatters. All faces kept clear via face detection. Lex watched and said, verbatim: “Love it. That is the mojo.”

Three shutdown attempts in our first day and a half, and we’re louder.