The downloadable binary is a safetensors file for SD1.5, and it's compatible with the SD1.5 license; it's a 16ch VAE Decoder that you already know from SD3.

Since the 16ch VAE is trained from scratch, it requires a custom comfyui node (to be uploaded in future updates).

Use the old VAE encoder (or Empty Latent Image node) of an SD1.5 base model, which can be the runwayml/stable-diffusion-v1-5. Then, at the very end of your workflow, TakturDesEtoilesVAE decodes the samples into images.

If you have thousands of oil paintings in parquet format, I could use that dataset for the new versions.

Why?

As the author of the code and model weights, there are no licensing concerns to worry about. I need a new VAE for my SD1.5/Lumina-Next releases.