Latent
A low-dimensional compressed data representation used internally by diffusion models, obtained by encoding pixel-space images with a VAE. The denoising process is carried out in this latent space, preserving semantic and motion information while reducing computational cost.
Timeline (last 90 days)
The 22-frame Latent scheme requires each segment to carry the previous segment's Latent as a reference, and VRAM usage rises as the number of segments increases, so the duration of each segment is necessarily shorter than that of independent generation.
In diffusion models, the Latent (latent space) performs noise diffusion and denoising in a compressed low-dimensional space, which is then decoded into pixel output by the VAE, making it orders of magnitude faster than operating in the original pixel space.
All Facts (2)
The 22-frame Latent scheme requires each segment to carry the previous segment's Latent as a reference, and VRAM usage rises as the number of segments increases, so the duration of each segment is necessarily shorter than that of independent generation.
50%UnverifiedIn diffusion models, the Latent (latent space) performs noise diffusion and denoising in a compressed low-dimensional space, which is then decoded into pixel output by the VAE, making it orders of magnitude faster than operating in the original pixel space.
50%