Verified65% confidenceFactExact time
The VAE (Variational Autoencoder) is responsible for encoding and decoding between latent space and pixel space, and is one of the significant VRAM consumption points in video/image generation pipelines.
3
Sources
65%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn image-to-video generation, the VAE (Variational Autoencoder) acts as an image compressor, with a typical compression ratio of 8×, meaning a 720p image is encoded into a 90×160 feature map.73% similarUnverifiedVAE decoding may account for 30% to 60% of the overall video generation time.71% similarUnverifiedVAE decoding optimizations for images and video deliver a 1.7× to 6.3× decoding speedup.70% similarUnverifiedVideo latents are compact encoded representations after the raw video pixels are compressed by a pretrained autoencoder into a low-dimensional, low-resolution latent space.68% similarUnverifiedThe latent space is the core working area of diffusion models, where images or videos first complete denoising and sampling in a compressed low-dimensional space before being restored to pixel images through a VAE decoder.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/931906API
curl https://kongchang.com/api/v1/knowledge/claims/931906MCP
get_claim(id=931906)