hustvl/LightningDiT
LightningDiT is a latent diffusion system that achieves FID=1.35 on ImageNet-256 with 21.8× faster training than standard DiT.
Not currently ranked — collecting fresh signals.
star history
This repository presents VA-VAE and LightningDiT, addressing the optimization dilemma in latent diffusion models where reconstruction and generation objectives conflict. The work introduces a variational approach to balance reconstruction quality with generative capability, and proposes architectural optimizations for faster diffusion transformer training. It achieves competitive FID scores on ImageNet-256×256 image generation with significantly reduced computational requirements.
Frequently asked
- What is hustvl/LightningDiT?
- LightningDiT is a latent diffusion system that achieves FID=1.35 on ImageNet-256 with 21.8× faster training than standard DiT.
- Is LightningDiT open source?
- Yes — hustvl/LightningDiT is open source, released under the MIT license.
- What language is LightningDiT written in?
- hustvl/LightningDiT is primarily written in Python.
- How popular is LightningDiT?
- hustvl/LightningDiT has 1.5k stars on GitHub.
- Where can I find LightningDiT?
- hustvl/LightningDiT is on GitHub at https://github.com/hustvl/LightningDiT.