AlphaFlow

This repo will contain the official pytorch code for "AlphaFlow: Understanding and Improving MeanFlow models". Our proposed method for understanding and improving MeanFlow models.

Uncurated samples (seeds 1-8) from the DiT-XL model for MeanFlow [1] and $\alpha$-Flow (our proposed method) produced with 1 (upper) and 2 (lower) sampling steps for ImageNet-1K $256^2$.

Installation

Step 1: Clone and set up environment

Clone the repo, install the conda environment and the packages in environment.yml:

git clone git@github.com:snap-research/alphaflow.git
cd alphaflow
conda env create -f environment.yml -p ./env
conda activate ./env

Step 2: Configure user settings

Copy an example of user settings:

cp configs/env/user-example.yaml configs/env/user.yaml

Now, edit configs/env/user.yaml to set the correct paths to the data, the environment and Wandb config (if required). You need to at least define project_path and conda_init_bin_path in this file.

Inference

Download checkpoints pretrained on ImageNet 256

Model	FID-NFE-1	FDD-NFE-1	FID-NFE-2	FDD-NFE-2
MeanFlow-B/2	43.1	819.2	38.5	787.6
α-Flow-B/2	40.2	781.0	37.1	775.0
MeanFlow-B/2-cfg	6.04	312.3	5.17	232.1
α-Flow-B/2-cfg	5.40	287.1	5.01	231.8
MeanFlow-XL/2-cfg	3.47	185.8	2.46	108.7
α-Flow-XL/2-cfg	2.95	164.6	2.34	105.7
α-Flow-XL/2+-cfg	2.58	148.4	2.15	96.8

Download the checkpoints and VAE encoder checkpoint sd_vae_ft_ema.pt from our Hugging Face repository into the ./snapshots folder.

mkdir snapshots

Generating samples

Use the following commands to generate samples.

# AlphaFlow-B/2 NFE-1
torchrun --max-restarts=0 --standalone --nproc-per-node=8 scripts/generate.py ckpt.snapshot_path=./snapshots/alphaflow-B-2.pt output_dir=data/generate/alphaflow-B-2-1-NFE/ seeds=0-49999 batch_size=32 sampling=recflow sampling.sigma_noise=1 sampling.num_steps=1 sampling.enable_trajectory_sampling=true

# AlphaFlow-XL-2-plus-cfg NFE-2
torchrun --max-restarts=0 --standalone --nproc-per-node=8 scripts/generate.py ckpt.snapshot_path=./snapshots/alphaflow-XL-2-plus-cfg.pt output_dir=data/generate/alphaflow-XL-2-plus-cfg-2-NFE/ seeds=0-49999 batch_size=32 sampling=recflow sampling.sigma_noise=1 sampling.num_steps=2 sampling.enable_trajectory_sampling=true sampling.enable_consistency_sampling=true

We provide a script ./scripts/evaluate.sh to generate samples and compute metrics for all released models.

Computing metrics

To evaluate a specific directory of generated samples (for FID, FDD, FCD):

# AlphaFlow-B/2 NFE-1
torchrun --max-restarts=0 --standalone --nproc-per-node=8 scripts/evaluate.py dataset.resolution=\[1,256,256\] gen_dataset.src=data/generate/alphaflow-B-2-1-NFE/ gen_dataset.resolution=\[1,256,256\] gen_dataset.label_shape=\[\]

Training

Preparing dataset

Download and reorganize the ImageNet dataset using the provided script. This script prepares the data structure required by our custom dataloader.

python scripts/data_scripts/prepare_imagenet.py --target_directory ./data/imagenet

Create an experiment config

We use hydra for experiment configuration. All primary experiment settings are stored in ./infra/experiments/experiments-alphaflow.yaml.

Launching an experiment

You can launch experiments in two ways:

Run the experiment directly without cloning the repository internally.

torchrun --max-restarts=0 --standalone --nproc-per-node=8 infra/launch.py alphaflow-sigmoid-latentspace-B-2 direct_launch=true wandb.enabled=false

Clone the repository into an ./experiments directory first, then run the experiment from the cloned version.

mkdir experiments
python infra/launch.py alphaflow-sigmoid-latentspace-B-2 wandb.enabled=false

A full list of possible arguments for the launch script can be found in ./configs/infra.yaml.

Citation

bibtex

@article{zhang2025alphaflow,
  title={AlphaFlow: Understanding and Improving MeanFlow Models},
  author={Zhang, Huijie and Siarohin, Aliaksandr and Menapace, Willi and Vasilkovsky, Michael and Tulyakov, Sergey and Qu, Qing and Skorokhodov, Ivan},
  journal={arXiv preprint arXiv:2510.20771},
  year={2025}
}