SDXL_anime_vae
This repository provides VAEs (Variational Autoencoders) optimized for SDXL anime illustration models.
SDXL_anime_natural_vae
Test Workflow
Sample Image
Difference Map
SDXL_anime_natural_vae | SDXL_Base_0.9_vae (Original)
| SDXL_anime_natural_vae | SDXL_0.9_vae | |
|---|---|---|
| MAE_similarity | 99.4 % | 99.0 % |
| SSIM_similarity | 99.9 % | 99.6 % |
| Red | -0.2 % | 0.3 % |
| Green | -0.1 % | 0.5 % |
| Blue | 0.0 % | 0.0 % |
SDXL_anime_natural_vae is a VAE fine-tuned based on SDXL_base_0.9_vae. Compared to the base model, it significantly minimizes fluctuations in MAE and SSIM, prevents green color casting, and produces images that are closer to the original with minimal degradation.
While this model was specifically tuned for anime illustrations, it is also expected to be effective for realistic photo-style illustrations.
Usage
Simply select this model as a replacement for your original VAE.
Precautions
- Recommended Formats: It is highly recommended to use FP32 or BF16.
- FP16 Issues: Because this model inherits the characteristics of the original SDXL_base_0.9_vae, running it in FP16 may result in "black images" (NaN errors).
- Alternative: If you specifically need to perform VAE processing in FP16, we recommend using the sdxl-vae-fp16-fix model instead of this one.
SDXL_anime_clear_vae
Comparison with SDXL_Base_0.9_vae (Original)
SDXL_clear_vae | SDXL_Base_0.9_vae (Original)
It features clearer rendering compared to the original and is characterized by deeper color expression, primarily in the red spectrum.
However, in areas of high complexity, there is a slight tendency for extremely fine details to be lost or simplified.