At what quality would you be interested in a new vae for sd class models?
A developer is seeking community input on the potential release of a new VAE (Variational Autoencoder) for Stable Diffusion models. The developer has achieved an LPIPS score of 0.490 with a 32-channel variant, aiming for a theoretical limit of 0.40, and is gauging interest from tinkerers and fine-tuners. The decision to continue development or pivot to retraining for current SD models will be influenced by the community's response. AI
IMPACT This is a niche discussion among Stable Diffusion users about a potential VAE improvement, with minimal direct industry impact.