sCMs

MallocSimenonsOctober 31, 2024Less than 1 minute

sCMs

发布于 24.10

Features

更高的压缩率：可以将压缩率提高到 $128 \times$ ，同时保持 reconstruction 的质量；提高后的压缩率可以用于加速 high-resolution diffusion models 的生成过程。（latent/high-res diffusion models？）
核心技术：Residual Autoencoding、Decoupled High Resolution Adaptation。

Methods

Figure 1: High spatial-compression autoencoders are more difficult to optimize. Even with the same latent shape and stronger learning capacity, it still cannot match the f8 autoencoder’s rFID — Figure 1: High spatial-compression autoencoders are more difficult to optimize. Even with the same latent shape and stronger learning capacity, it still cannot match the f8 autoencoder’s rFID