Pretrained Reversible Generation as Unsupervised Visual Representation Learning

🎉 What I cannot create, I do not understand.

🔥 News

• We are working on separating the codebase from GenerativeRL into its own standalone repository. We will first release the version that depends on the GenerativeRL library.
• 2025.07.01: We update PRG camera ready version for ICCV 2025 and release project website
• 2024.11.29: We release PRG on arxiv.

What is PRG

PRG proposes to turn a pretrained continuous time flow diffusion generator upside down running the model backward produces multi level features that after light fine tuning serve as an unsupervised representation extractor.

Requirements

• Python 3.10 or later
• Conda (optional, but recommended)
• GenerativeRL (temporarily required)

Installation

1. Prepare the environment

conda create -n prg-env python=3.10  
conda activate prg-env

2. Install GenerativeRL

git clone https://github.com/opendilab/GenerativeRL.git
cd GenerativeRL
pip install -e .

3. Install PRG

git clone https://github.com/opendilab/PRG.git
cd PRG
pip install -e .

Acknowledgement

The Pretrained Reversible Generation codebase is adapted from the following repositories:

• Swin-Transformer
• GenerativeRL
• flow_matching
• SiT
• RectifiedFlow

A huge thanks to the authors of these projects for their outstanding contributions! 🎉