Experimental Custom VAE (miniflux R&D)
Custom Autoencoder built from scratch as part of the miniflux exploratory pipeline. It incorporates combined LPIPS perceptual loss, GAN adversarial loss (PatchGAN), and KL divergence regularization, optimized for latent space dynamics.
Dataset & Training
- Dataset: 10,000 images from the UltraEdit dataset.
- Objective: Evaluate latent space spatial compression and reconstruction fidelity prior to Flux.2 VAE integration.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support