sihyun-yu/REPA
A method that aligns noisy diffusion transformer states with pretrained visual encoder representations to improve training efficiency and generation quality.
Not currently ranked — collecting fresh signals.
star history
REPA (Representation Alignment for Generation) aligns noisy input states in diffusion models with representations from pretrained visual encoders. The method significantly improves training efficiency, speeding up SiT diffusion transformer training by 17.5x while achieving state-of-the-art image generation quality on ImageNet 256x256 benchmarks. This research targets the core training methodology for generative diffusion models.
Frequently asked
- What is sihyun-yu/REPA?
- A method that aligns noisy diffusion transformer states with pretrained visual encoder representations to improve training efficiency and generation quality.
- Is REPA open source?
- Yes — sihyun-yu/REPA is open source, released under the MIT license.
- What language is REPA written in?
- sihyun-yu/REPA is primarily written in Python.
- How popular is REPA?
- sihyun-yu/REPA has 1.7k stars on GitHub.
- Where can I find REPA?
- sihyun-yu/REPA is on GitHub at https://github.com/sihyun-yu/REPA.