VAEs for Image Generation

This repository hosts a curated collection of VAE checkpoints used by diffusion and transformer-based image generation pipelines.

Available VAEs

AutoencoderKL

Model	Source	Latent Channels
SD21-VAE	Stable Diffusion 2.1	4
SDXL-VAE	Stable Diffusion XL	4
SD35-VAE	Stable Diffusion 3.5	16
FLUX1-VAE	FLUX.1	16
FLUX2-VAE	FLUX.2	32
SANA-VAE	SANA (DC-AE)	32
Qwen-VAE	Qwen-Image	16

VQModel

Model	Source	latent_channels	num_vq_embeddings	vq_embed_dim	sample_size
VQDIFFUSION-VQVAE	VQ-Diffusion (microsoft/vq-diffusion-ithq)	256	4096	128	32
IBQ-VQVAE-1024	IBQ (TencentARC/SEED)	256	1024	256	32
IBQ-VQVAE-8192	IBQ (TencentARC/SEED)	256	8192	256	32
IBQ-VQVAE-16384	IBQ (TencentARC/SEED)	256	16384	256	32
IBQ-VQVAE-262144	IBQ (TencentARC/SEED)	256	262144	256	32
MOVQGAN-67M	MOVQGAN	4	16384	4	256
MOVQGAN-102M	MOVQGAN	4	16384	4	256
MOVQGAN-270M	MOVQGAN	4	16384	4	256

Diffusers usage

AutoencoderKL (SD, FLUX, SANA, Qwen, etc.):

from diffusers import AutoencoderKL

vae = AutoencoderKL.from_pretrained(
    "BiliSakura/VAEs",
    subfolder="SDXL-VAE",
)

VQModel (VQ-Diffusion, IBQ, MOVQGAN):

from diffusers import VQModel

vae = VQModel.from_pretrained(
    "BiliSakura/VAEs",
    subfolder="VQDIFFUSION-VQVAE",
)

Notes

• All models are VAE checkpoints intended for inference use in their corresponding pipelines.
• Latent channel count is listed to help match with the correct backbone.