NodeValid

Published video-model weights, licences and hardware notes

Cold start, the cost before generation

A cold start is everything before the first denoising step: reading the weights from storage, moving them onto the card, and initialising the pipeline. It happens on every run and no memory figure contains it. As of 2026-09-22.

Where the time before generation goesStorage read, transfer onto the card and pipeline initialisation all precede the work a published timing usually measures.ReadTens of gigabytes offstorageinto host memoryTransferHost memory onto thecardweights residentGenerateThe work a timingmeasuresMeasured from the wrong endA published timing usually starts at the third box.
Fig. 1 Three stages that happen before the stage anybody publishes a number for.
What a cold start costs that a memory figure does not show. Recorded 2026-09-22.
CostWhere it lands
Storage read timeWall clock, every run
Transfer timeWall clock, every run
Fast storageCapital, not memory
Idle card timeUtilisation, not capacity

Inclusion rule. Costs incurred before generation begins. A cost that is capital rather than time is included, because it is the one that gets left out of comparisons. Order. From the recurring time costs to the one-off capital cost.

1It dominates short clips

For a release that produces five seconds, loading thirty gigabytes of weights can take longer than the generation. A timing published for the generation step alone therefore describes a fraction of the elapsed time.

Keeping a model resident avoids it and costs a card that sits idle between jobs, which is a utilisation problem rather than a capacity one.

2It belongs in any local comparison

Anyone weighing local running against a per-second rate has to put storage, load time and failed runs on the same page as the rate. Otherwise the comparison is hardware against a price.

None of it appears in this register, because no repository publishes it. The entries record what vendors state, and this is a cost they leave to the reader.

3Why no cell here holds it

No release in the register publishes a load time, a storage requirement or a throughput figure that includes either. The hardware column holds memory statements and, in two cases, a timing for the generation step.

So this cost is real, recurring and entirely outside the register. Anyone comparing local running against a per-second rate has to add it themselves, along with the runs that have to be repeated.

A reading note, not an entry: no vendor value appears on this page. Where the register records this term for a particular release, it is on the hardware column and on that family's own page. Nearby terms: Throughput, Storage footprint.