NodeValid

Published video-model weights, licences and hardware notes

Pixel count, and the memory it moves

Pixel count is width times height for one frame. It drives the activation memory rather than the weights, which is why a memory figure at one resolution does not transfer to another and why the two columns stay joined. As of 2026-09-22.

Where pixels land in a memory figureWeights are fixed once a model is loaded and activations scale with the frame, so the share of a memory figure that moves with resolution depends on the model size.Scales with pixelsFixed once loadedActivationsYes, per frame—Autoencoder workYes, at full resolution—Weights—Set by the count and the formatShare of the totalLarger on small modelsLarger on big onesThe moving part of the figure
Fig. 1 On a large model the pixels barely move the total; on a small one they dominate it.
What the pixel share of a memory figure implies, by model size. Recorded 2026-09-22.
Model sizeEffect of doubling pixels
Small, with low memoryA large relative change
Large, with high memoryA small relative change
Long clipsCompounded by the frame count
Short clipsVisible, but bounded

Inclusion rule. Cases where a resolution change has a different effect on the memory figure. Rows are regimes rather than releases, so no vendor value appears. Order. From the case where pixels matter most to the case where they matter least.

1Two rows in this register show both regimes

One repository moves from 45GB to 60GB when the pixel count roughly doubles, which is a third rather than a doubling. Another barely moves at all across three configurations, because thirty billion parameters dominate the total.

So how much resolution matters is a property of the model rather than a general rule, and it can only be read off releases that publish more than one figure.

2Pixels and frames multiply

Activation memory scales with pixels per frame and with the number of frames held at once, so a long clip at a high resolution compounds both. A figure that states one and not the other is half a configuration.

This is why the register keeps resolution and length in separate columns and keeps both attached to every memory value it records.

3Two regimes visible in the register

One family's figures move from 45GB to 60GB when the pixel count roughly doubles, which is a third rather than a doubling. Another barely moves across three configurations, because thirty billion parameters dominate the total.

So how much resolution matters is a property of the release, and it can only be read off vendors that publish more than one figure. Eight families do; seven do not.

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 resolution column and on that family's own page. Nearby terms: Context frame, Latent space.