Zero to Director #10 — Turbo Boxes: LoRA for Video
From Last Time
Post #9 opened the reference desk; every arm made rolls richer, and richer rolls are slower rolls. The one thing motion always charges you is time — in memory, in steps, in waiting.
Post #7 taught the toolbox on a stills machine: a small box of specialist deltas, clipped onto a carver you trust. This post clips boxes onto the video machine — with one job above all the others: giving steps back.
Why Boxes on a Video Machine
A 40-stroke video roll is a meditative experience: minutes of latent carving per roll, every one of them multiplied by frames. Now the recall from post #7: a LoRA is a box of differences, trained onto one machine, that changes how it carves without changing what it is.
Train a box whose whole job is "make 8 strokes look like 40" and you have a turbo box — distillation as a toolbox. Clip it on, drop the step dial, and the same carver finishes rolls at a fraction of the patience.
Two Turbo Boxes
Two are on the bench, both married to H3 — and both bring the iron law with them, sharpened:
minimax_h3_turbo_v4_step600_ema_pruned— the general turbo box, trained against the H3 line. Plain text-to-video speedups.minimax_h3_ref2v_turbo_8step_v1.0_768p— a specialist, and the name is a contract:ref2v,8step,768p. It expects the reference arms of post #9, eight strokes, 768-pixel rolls.
A video box does not just marry the machine — it marries the step count and the resolution it was distilled at. Clip the 768p specialist onto a 1080p roll and the delta speaks the wrong dialect at the wrong size; post #7's mush, now in motion.
Steps in Turbo Land
The step economy, honestly:
- 40 steps, no box — the patient way. The quality ceiling, and the patience floor.
- 8 steps, turbo box — the promise: same neighborhood of coherence, one-fifth of the wait.
- The dial still turns — 4, 6, 10 — but the box was distilled at one number; treat the label as home base and orbit it, not wander.
Runtime is the metric this post records like a lab notebook: same roll, both configs, the clock does not lie.
Examples
The boost
10-turbo-boost.json is the same roll twice: 40 strokes bare, 8 strokes with the turbo box — and the runtimes go in the table.
The trade, timed and measured: {OBSERVE-BOOST}
The step dial
10-step-dial.json orbits home base: 4, 6, 8, 10 strokes, box clipped, one seed.
Where turbo land breaks: {OBSERVE-STEPS}
Turbo with references
10-turbo-with-refs.json is the specialist's contract exercised: post #9's reference arms at 768p, eight strokes, the ref2v box earning its name.
Speed and arms, coexisting: {OBSERVE-TURBOREFS}
Style over time
10-style-over-time.json clips a style box beside the turbo box — {STYLE-LORA-NAME} — the series-consistency play from post #7, now holding a look across a whole roll.
Style that must not fight motion: {OBSERVE-STYLE}
Pitfalls
- The law has three clauses now. Machine, step count, resolution — a video box marries all three. Read the filename like a contract; it is one.
- Eight-step eyes. A turbo roll is not a 40-step roll in hiding — it is a distillation with its own softness. Do not zoom crops to compare; watch it as a video, which is how it will be seen.
- Boxes compound in time. Turbo at 1.0 plus style at 0.8 is 1.8 boxes of opinion per frame, times every frame. Stack low, watch, raise.
- Speed invites scope creep. The first thing everyone does with 5× speed is 5× resolution, and rediscovers the memory ceiling from post #8. Grow one dial at a time.
Your Move
Run the boost experiment with your own blueprint and write your own runtime table — your machine's clock is the only one that matters for your patience budget. Then keep the box on for the rest of the series: waiting less is a creative feature.
Next post the series stops teaching and starts reading: a real 54-node production workflow, group by group, with everything you now know.