Researcher: diffusion autoencoders are lossy, training data info already lost
Kalomaze argues that the autoencoders in SOTA latent diffusion models are lossy tokenizers that discard information at the bottleneck, noting that generated videos lack low-frequency effects like Doppler shifts.
2026-11-04 ~ 2026-11-04 · 2 related posts
- kalomaze: crude diffusion autoencoders may lose low-frequency info at the latent stage — kalomaze · 2026-11-04
- kalomaze: Diffusion autoencoders are lossy tokenizers — data is gone before training — kalomaze · 2026-11-04