OpenVAE

OpenVAE is a medical-image VAE family for CT/MRI. It provides pretrained latent backbones for diffusion models, with better anatomical fidelity than general-image VAEs.

Contribution to Generative AI

OpenVAE brings domain-specific latent modeling to medical generative AI, making medical diffusion pipelines more reliable, reproducible, and easier to build.

Quick Start

import torch
from diffusers import AutoencoderKL

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load OpenVAE
vae = AutoencoderKL.from_pretrained("SMILE-project/OpenVAE", subfolder="vae").to(device)
vae.requires_grad_(False)
vae.eval()

img = torch.randn(1, 3, 512, 512, device=device)

with torch.no_grad():
    # Encode to latent space
    latent = vae.encode(img).latent_dist.sample()

    # Decode to image space
    reconstruction = vae.decode(latent).sample

Models

Name	VAE Type	# Patients
stable-diffusion-v1-5	KL-VAE	0
stable-diffusion-3.5-large	KL-VAE	0
OpenVAE-2D-4x-20K	KL-VAE	20K
OpenVAE-2D-4x-100K	KL-VAE	100K
OpenVAE-2D-4x-300K	KL-VAE	300K
OpenVAE-2D-4x-PCCT_Enhanced	KL-VAE	300K
OpenVAE-3D-4x-20K	KL-VAE	20K
OpenVAE-3D-4x-100K	KL-VAE	100K
OpenVAE-3D-4x-1M	KL-VAE	1M
OpenVAE-3D-4x-100K-VQ	VQ-VAE	100K
OpenVAE-3D-8x-100K-VQ	VQ-VAE	100K

Benchmarks

Name	LPIPS	SSIM	PSNR	DSC
stable-diffusion-v1-5	-	-	-	-
stable-diffusion-3.5-large	-	-	-	-
OpenVAE-2D-4x-20K	-	-	-	-
OpenVAE-2D-4x-100K	-	-	-	-
OpenVAE-2D-4x-300K	-	-	-	-
OpenVAE-2D-4x-PCCT_Enhanced	-	-	-	-
OpenVAE-3D-4x-20K	-	-	-	-
OpenVAE-3D-4x-100K	-	-	-	-
OpenVAE-3D-4x-1M	-	-	-	-
OpenVAE-3D-4x-100K-VQ	-	-	-	-
OpenVAE-3D-8x-100K-VQ	-	-	-	-

Citation

@article{liu2025see,
  title={See More, Change Less: Anatomy-Aware Diffusion for Contrast Enhancement},
  author={Liu, Junqi and Wu, Zejun and Bassi, Pedro RAS and Zhou, Xinze and Li, Wenxuan and Hamamci, Ibrahim E and Er, Sezgin and Lin, Tianyu and Luo, Yi and Płotka, Szymon and others},
  journal={arXiv preprint arXiv:https://www.arxiv.org/abs/2512.07251},
  year={2025},
  url={https://github.com/MrGiovanni/SMILE}
}