language: en
license: mit
pipeline_tag: image-to-image
tags:
- diffusion
- autoencoder
- feature-space
- svg
SVG: Latent Diffusion Model without Variational Autoencoder
SVG is a novel latent diffusion model framework that replaces the traditional Variational Autoencoder (VAE) latent space with semantically structured features from self-supervised vision models (e.g., DINOv3). This design improves generative capability and downstream transferability while maintaining efficiency comparable to standard VAE-based models.
Resources
- Paper: Latent Diffusion Model without Variational Autoencoder
- Project Page: https://howlin-wang.github.io/svg/
- GitHub Repository: https://github.com/shiml20/SVG
Model Description
SVG constructs a feature space with clear semantic discriminability by leveraging frozen DINO features, while a lightweight residual branch captures fine-grained details for high-fidelity reconstruction. Diffusion models are trained directly on this semantically structured latent space to facilitate more efficient learning.
Key features:
- Replaces low-dimensional VAE latent space with high-dimensional semantic feature space.
- Includes a lightweight residual encoder for refining fine-grained details.
- Enables accelerated diffusion training and supports few-step sampling.
- Improves generative quality while preserving semantic and discriminative capabilities.
Usage
For full instructions on training and evaluation, please refer to the official GitHub repository.
Installation
conda create -n svg python=3.10 -y
conda activate svg
pip install -r requirements.txt
Generation
To generate images using a trained model:
# Update ckpt_path in sample_svg.py with your checkpoint
python sample_svg.py
Citation
If you find this work useful for your research, please cite:
@misc{shi2025latentdiffusionmodelvariational,
title={Latent Diffusion Model without Variational Autoencoder},
author={Minglei Shi and Haolin Wang and Wenzhao Zheng and Ziyang Yuan and Xiaoshi Wu and Xintao Wang and Pengfei Wan and Jie Zhou and Jiwen Lu},
year={2025},
eprint={2510.15301},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2510.15301},
}