Mage-VAE

But with different latent channels.

Credits to MinhNH232331M for the unpatchifier and Chenyang Zhu for the original pull request.

As the channel size increases, the image's structure and edges become clearer.

If you're hoping for an SDXL VAE replacement, nope, the 4ch version is unattainable.

These files are intended for the pre-training of a custom UNet model when the 32ch input is too wide for a quick training session.

The first three channels loosely follow the CIELAB values by design in the unpatched latent.

References

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support