Inconsistency due to changes in the Variational Autoencoder (VAE) for the Kandinsky diffusion model and with EfficientNet v2small model.
| Testing/Training | Kandinsky 2.2 | Autoencoder with KL Loss (AKL) 1.5 | Autoencoder with KL Loss (AKL) 2 | Consistency Decoder | VAE from Pixart-α | VAE from SD 2.1 | VAE from SD Turbo | οVAE from SD XL Turbo | SD1.4 EMA fine-tuned VAE | SD1.4 MSE fine-tuned VAE | SDXL fine-tuned VAE | Tiny AutoEncoder for SD |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Kandinsky 2.2 | 1.26% | 1.41% | 1.40% | 1.33% | 1.39% | 1.42% | 1.37% | 1.37% | 1.40% | 1.33% | 1.43% | 1.44% |
| Testing/Training | Kandinsky 2.2 | Autoencoder with KL Loss (AKL) 1.5 | Autoencoder with KL Loss (AKL) 2 | Consistency Decoder | VAE from Pixart- | VAE from SD 2.1 | VAE from SD Turbo | οVAE from SD XL Turbo | SD1.4 EMA fine-tuned VAE | SD1.4 MSE fine-tuned VAE | SDXL fine-tuned VAE | Tiny AutoEncoder for SD |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Kandinsky 2.2 | 1.26% | 1.41% | 1.40% | 1.33% | 1.39% | 1.42% | 1.37% | 1.37% | 1.40% | 1.33% | 1.43% | 1.44% |
Sharing content requires targeting cookies to be enabled. Please update your cookie preferences to use this feature.