The VAE: a learned codec between pixels and latents

A variational autoencoder is a codec that was trained rather than designed. Like ProRes or H.264 it comes in two halves: an encoder that takes pixels and emits a much smaller representation, and a decoder that expands that representation back into pixels. Unlike ProRes, nobody hand-wrote its transform. It was learned, using the encoder and decoder formulation from Auto-Encoding Variational Bayes, in which the encoder produces a distribution over the compressed code rather than one fixed code. In a text-to-image or text-to-video pipeline the diffusion model never touches pixels. The encoder runs during training, turning source footage into latents. At generation time only the decoder runs, once, at the end, on the finished latent. That division of labour explains a whole class of artifact. If a shot is soft in sensible places but fine texture buzzes, if gradients carry a faint regular grid, if skin has a plasticky sheen, the problem is decode rather than denoise. Upsampling layers are known to produce checkerboard patterning, documented in detail by Odena, Dumoulin and Olah. Rerolling the prompt will not remove a decoder artifact, because the same decoder produces it every time. Video codecs of this kind also compress the third axis. CogVideoX built a 3D VAE that compresses along time as well as space, replacing the older practice of applying an image VAE frame by frame, which leaves motion between consecutive frames unsmooth. CV-VAE makes a related argument and adds a compatibility constraint so a video codec can share an existing image model's latent space. LTX-Video goes further and hands its decoder the final denoising step, producing clean pixels directly rather than paying for a separate upsampling module. So the decoder is not a passive last stop. It partly decides how detail, grain and motion land.

What people get wrong

Each of these is a belief worth dropping, followed by what is actually happening.

• the VAE is part of the diffusion model. It is a separate network, trained separately and sometimes swapped independently, so two pipelines can share a diffusion model and differ in their decoder.
• every artifact comes from the diffusion model, so another seed will clear it. Decoder artifacts are deterministic given the latent, which is why grid patterning, texture buzz and waxy skin reappear across every seed.
• a video VAE is just an image VAE applied frame by frame. Early pipelines did that and it leaves motion between consecutive frames unsmooth. Current video models compress along time too, which is what a 3D VAE means.
• more latent compression is free efficiency. It buys speed and duration and costs fine detail, which is why LTX-Video states the limitation directly and gives extra work to the decoder to compensate.
• the decoder simply enlarges the latent. It reconstructs pixels from a learned description, and in some designs it also performs the last denoising step, so it is making decisions about texture rather than scaling anything.

Terms this introduces

encoder

The half of the autoencoder that turns pixels into a latent, used during training and for image-to-video conditioning.

decoder

The half that turns a finished latent back into pixels, run once at the end of generation.

3D VAE

A video autoencoder that compresses along time as well as height and width.

reconstruction artifact

A defect introduced by decoding rather than by the diffusion process, so it repeats across seeds.

checkerboard artifact

Regular grid patterning produced by certain upsampling layers in a decoder.