This repository is the official implementation of FlowAlign, an inversion & training free image editing algorithm.

💡 Recent inversion-free, flow-based editors leverage models like Stable Diffusion 3 to enable text-driven image editing via ODE integration.

🤔 However, skipping latent inversion often leads to unstable trajectories and poor source consistency.

🚀 FlowAlign addresses this by introducing a flow-matching loss—a simple yet effective regularizer that ensures smooth, semantically aligned, and structurally consistent edits.

🌟 Thanks to its ODE-based formulation, FlowAlign naturally supports reverse editing, highlighting its reversible and robust transformation capability.

Clone this repo:

git clone https://github.com/FlowAlign/FlowAlign.git
cd FlowAlign

To install requirements:

conda create -n flowalign python==3.11
conda activate flowalign
pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

For the text-based image editing, run:

Examples 1

python run_edit.py \
  --img_path "samples/bicycle.jpg" \
  --src_prompt "a slanted mountain bicycle on the road in front of a building" \
  --tgt_prompt "a slanted rusty mountain bicycle on the road in front of a building"

The expected result:

Example 2

python run_edit.py \
  --img_path "samples/cat.jpg" \
  --src_prompt "a opened eyes cat sitting on wooden floor" \
  --tgt_prompt "a closed eyes cat sitting on wooden floor"

The expected result:

You can freely change the editing method using arguments:

• method : dual / sdedit / flowedit / flowalign

If you use --efficient_memory, text encoder will pre-compute text embeddings and is removed from the GPU.

This allows us to run image editing with a single GPU with VRAM 24GB.

All edited images were generated on a single NVIDIA RTX 3090 GPU, using a fixed random seed of 123 and a Classifier-Free Guidance (CFG) scale of 13.5.