§01 / What It Can Do
6 capabilities · one per stageWhat it can do.
Six pipeline stages, one entry each. Click one and the entry states what the stage does and the production rules it runs by — limits included.
We built the reconciler because client storyboards routinely contradict themselves: the scene text names one colorway, the sock photo pasted next to it shows another. It walks the Word doc in order, pulls each scene's text and its embedded image, and catches the disagreement.
It detects the SKU candidates, cross-checks the product catalog, and applies one hard rule we set — an explicit text SKU beats the embedded image, which beats the family description — to suggest the pick, while leaving the genuinely ambiguous calls to you. One rule it never breaks: colorways are matched by what the photo shows, never by filename, because numbering drifts between Drive and the local mirror.
Every later stage builds on this reconciliation.
— from the product notes
We keep this stage read-only — it is the product brain of the pipeline. It serves six style presets — Wellness by default, then Lifestyle, Travel, Active, Premium/Quiet Luxury, and a fourteen-second Color-Bomb pattern-interrupt — each with its own music vibe and BPM range.
It holds the per-SKU anatomy parsed from the spec — silhouette, cuff, toe construction, where the brand-mark sits, the colorways — and surfaces the candidate lifestyle anchor photos for any SKU, with dimensions, orientation, and a flag when a file is over 10MB.
We loaded the presets with real production settings: Premium runs 80-100 BPM cinematic piano; Color-Bomb is trap and hard cuts, reserved for pattern-heavy launches only.
— from the product notes
It makes the frames before anything moves — start frames and finished hero stills, generated in parallel, each anchored on the real product photo, via fal's nano-banana. The prompts carry recipes that took us a hundred bad renders to learn: use the canonical full-res reference, not a thumbnail, or the colorway drifts; call a panel 'glowing' so the model doesn't quietly drop it; and use training-data color names — 'military green' and 'sky blue' render, 'olive' and 'ecru' don't.
When fidelity has to be exact, it pastes a transparent-bg product PNG onto the generated plate so the pattern, colorway, and wordmark stay bit-identical. Catching a problem at this stage is far cheaper than re-rendering a finished clip.
— from the product notes
It turns each frame into motion on Kling, running in a background thread keyed to the spot. One rule sounds backwards but we earned it: never name the wordmark in the positive prompt, not even to 'preserve' it — name it and Kling fabricates fake text onto the cuff. We forbid text-on-sock in the negative prompt instead.
It auto-compresses anchors under 10MB, keeps prompts under Kling's 2500-char limit, and persists every beat's fal request id. Restart mid-render and it reattaches the orphaned jobs by re-polling — no re-submit, no double charge.
One more constraint we state flat: a subject that walks off frame needs both a start and an end frame, or the clip reverses.
— from the product notes
It finds the track and lands it on the cut. It searches Pixabay for the preset's vibe, downloads the candidates, and uses librosa to score how well each one's beats line up with the cut targets — the beat boundaries, say [0, 4, 14, 24, 34, 44, 54] for a seven-beat sixty-second spot.
Back come ranked candidates with tempo, duration, the best start-offset, and an alignment score, and you lock the one you want. We let this stage run slow and synchronous — sixty to a hundred and twenty seconds — scraping via curl to dodge the 403 that plain requests hits, with the analysis off the event loop so nothing else blocks.
— from the product notes
This is the final cut. Once every beat is rendered and a track is locked, it runs the whole ffmpeg pass: letterbox and trim each beat to its target length, concat them, burn in the fade-in and fade-out captions, build the audio bed from the track's best offset, and mux it to a finished 9:16 MP4.
It pulls labeled keyframes per beat and hands back a signed, browser-fetchable download URL. We strip apostrophes from captions and default them to 70px.
The blocking job runs off the event loop, and when it finishes the spot flips to 'completed' with the final path saved — the storyboard that came in at stage one leaves as a finished spot.
— from the product notes
At the counter
Heard the intro? Book a session — a trial on your own machine, a read-only Phase-0 assessment, or the full build. Nothing runs until you approve it.
§02 / Also on the Shelf
Previous · NextNext on the shelf.
Seven products in the shop. These are the two either side of this one.