Z-Image Anime VAE

Introduction

A Decoder finetune(That means it can be used as-is) of the Z-Image (Flux) AE on a dedicated anime illustration dataset.
It allows to reduce high frequency compression artifacts and oversharpening, that are inherent to realism data.

Additionally helps with higher resolution, when multiple re-encodes are involved.

As of now, it is not extremely useful, as we don't have anime base yet, but i've seen quite a few anime gens done, so why not. And Tongyi team plans to do something with anime data, so im patiently waiting.

Visual Examples

1-2-4k hires:

1-2k hires:

Full comparison: https://slow.pics/c/kmCwRG4l

Usage

Load as VAE instead of ae.safetensors

Reprinted from Hugging Face