DA-Flow: Degradation-aware optical flow model
1 min readupdated
DA-Flow provides pretrained weights for optical flow estimation using image restoration diffusion features. Based on stabilityai/stable-diffusion-3.5-medium, the model integrates a DiT4SR transformer and a dedicated network structure for flow processing. It enables more accurate motion estimation on damaged, noisy, or blurry frames by leveraging the capabilities of the diffusion model. This tool is designed for researchers and developers working in computer vision and video processing.
Key features
- Uses diffusion restoration features from Stable Diffusion 3.5 Medium
- Includes a dedicated DiT4SR transformer for the initial stage
- Built for handling complex image degradation conditions
- Based on open research and the official code repository
What you can do
- Run optical flow estimation for noisy or blurry frame sequences
- Experiment with motion visualization using diffusion features
- Integrate the model into video processing pipelines alongside SD3.5 Medium components
Example
Model usage is documented in the official GitHub repository, which loads the transformer/ and daflow_flow.pth weights along with the VAE and scheduler from Stable Diffusion 3.5 Medium.
Read more
DetailsBase model: stabilityai/stable-diffusion-3.5-medium
Parameters: 2.94B
Library: diffusers
Components: transformer/ and daflow_flow.pth










Comments