NodeValid

Published video-model weights, licences and hardware notes

Latent space, and why pixels are not the unit

Most current video models generate in a compressed representation rather than in pixels. The autoencoder compresses frames into it and reconstructs afterwards, which is why doubling the frame size does not double the work. As of 2026-09-22.

Where the generation actually happensCompression before, reconstruction after, and the expensive part in between operating on something much smaller than a frame.CompressFrames reduced to acompact representationa small tensorGenerateDenoising steps in thatcompact spacea generatedtensorReconstructExpanded back to framesat full sizePixels at the ends, not the middleThis is why a published memory figure does not scale with the pixel count in a straightline.
Fig. 1 The unit the model works in is not the unit a reader measures the output in.
Which stage each resource cost belongs to. Recorded 2026-09-22.
CostStage
Denoising computeThe compressed stage
Peak activation memoryOften the reconstruction
Frame-size sensitivityThe two ends
Step-count sensitivityThe middle

Inclusion rule. Costs a reader might attribute to resolution. Each is assigned to the stage it actually belongs to, with no vendor figure attached. Order. From the cost in the compressed stage outward.

1Compression ratios are not published

How much the representation compresses a frame is an architectural property that no release in this register states. Without it, a reader cannot predict how a memory figure moves with resolution.

That is one more reason the register quotes figures at their stated configuration rather than interpolating between them.

2The reconstruction is often the peak

Expanding back to full-resolution frames touches more data than the denoising steps do, so the highest memory moment in a run can be at the end rather than in the middle.

It also explains why a small autoencoder can account for a large share of the memory, and why its parameter count is a poor guide to its cost.

3Why the register does not interpolate

No release here publishes a compression ratio, so nobody can predict from these pages how a memory figure moves between two resolutions. The figures are quoted at their stated configurations and joined to nothing.

Two families publish this stage's parameter count separately, which at least names the component. Its share of the weights is small and its share of the memory is not, and no cell here records the second figure.

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: Upscaling, Frame count.