The Shape of the Latent

The latent video is a four-dimensional tensor: frames, channels, height, and width. The frame axis is the temporal dimension — how many latent frames the clip contains, which after decoding becomes the number of visible frames. The channel axis holds the learned feature dimensions the VAE chose to represent the video. Height and width are the spatial dimensions of each latent frame, smaller than the pixel resolution by the VAE's compression factor. Every denoising step operates on this whole tensor, so the denoiser sees all frames and all spatial positions together. When the loop finishes, the VAE decoder expands the tensor back into pixel frames.