Open-Sora
Open-Sora copied to clipboard
Trade off between VQ-VAEs and VAEs
Dear authors,
Thanks for the great work!
While OpenAI's Sora itself and sota 3D VAEs (MAGVIT) uses the Vector Quantized (VQ-VAE) variant, open-sora adopts the KL-VAE. What's the rationale behind it and what could be the potential trade-off? Do you have any insight regarding this issue?
Thanks in advance!