CVPR2026 Code Release

Installation

1. Create Python 3.11 Virtual Environment

conda create -n cvpr2026 python=3.11
conda activate cvpr2026

2. Install Dependencies

pip install -r requirements.txt

3. Install Modified diffusers Library

Download diffusers v0.34.0 and replace the sana_transformer.py file:

# Clone diffusers repository
git clone https://github.com/huggingface/diffusers.git
cd diffusers
git checkout v0.34.0

# Replace sana_transformer.py
# Location: src/diffusers/models/transformers/sana_transformer.py
cp ../sana_transformer.py src/diffusers/models/transformers/sana_transformer.py

# Install modified diffusers
pip install -e .
cd ..

4. Install Project

pip install -e .

Usage

Training

Modify the model paths and parameters in scripts/sft.sh, then run:

bash scripts/sft.sh

Inference

Modify the model paths and parameters in inference.py, then run:

python inference.py

Acknowledgements

This code is built upon BLIP3o and MeanFlow. We sincerely thank the authors for their excellent work.