Self-Flow ImageNet 256×256

Self-Flow is a self-supervised training method for diffusion transformers that combines flow matching with a self-supervised feature reconstruction objective. This checkpoint is trained on ImageNet 256×256.

Key Features

• Architecture: SiT-XL/2 with per-token timestep conditioning
• Training: Flow matching + self-supervised feature reconstruction
• Resolution: 256×256 pixels
• Parameters: ~675M

Evaluation Results

Metric	Value
FID ↓	5.7
IS ↑	151.40
sFID ↓	4.97
Precision	0.72
Recall	0.67

Results computed on 50,000 generated samples vs ImageNet validation set.

Usage

Download Checkpoint

python -c "
from huggingface_hub import hf_hub_download
hf_hub_download(
    repo_id='Hila/Self-Flow',
    filename='selfflow_imagenet256.pt',
    local_dir='./checkpoints'
)
print('Downloaded!')
"

and follow the instructions in our repository: https://github.com/black-forest-labs/Self-Flow

License

This model is released under the Apache 2.0 License.