FlowFeat

FlowFeat is a pixel-dense visual representation learned from optical flow, released with the NeurIPS 2025 Spotlight paper FlowFeat: Pixel-Dense Embedding of Motion Profiles.

• Paper page: https://hf.co/papers/2511.07696
• Project code: https://github.com/tum-vision/flowfeat
• Interactive demo: https://huggingface.co/spaces/neek-ans/flowfeat-demo

This model repo is focused on easy access to pretrained checkpoints and a clean inference path. It does not re-host training data.

Included variants

Variant	Backbone	Checkpoint file
dinov2_vits14_yt	DINOv2 ViT-S/14 trained on YouTube-VOS	dino2_s14_flowfeat_yt.pth
dinov2_vitb14_yt	DINOv2 ViT-B/14 trained on YouTube-VOS	dino2_b14_flowfeat_yt_v2.pth
dinov2_vitb14_kt	DINOv2 ViT-B/14 trained on Kinetics	dino2_b14_flowfeat_kt_v2.pth

Quickstart

Install directly from this repo:

pip install git+https://huggingface.co/neek-ans/flowfeat

Load the default checkpoint:

from flowfeat_hf import flowfeat

model = flowfeat(name="dinov2_vits14_yt", pretrained=True)
model.eval()

Run inference:

import torch

x = torch.randn(1, 3, 224, 224)
with torch.no_grad():
    y_enc, y_dec = model(x)

print(y_enc.shape)
print(y_dec.shape)

Notes

• The repo packages only the inference-time code needed to load FlowFeat and its decoder.
• Backbone weights are still resolved through the upstream DINO / DINOv2 loading path used by the original research code.
• Checkpoint files are hosted in this repo under checkpoints/.

Citation

@inproceedings{Araslanov:2025:FlowFeat,
  author = {Araslanov, Nikita and Sonnweber, Anna and Cremers, Daniel},
  title = {{FlowFeat}: Pixel-Dense Embedding of Motion Profiles},
  booktitle = {NeurIPS},
  year = {2025},
}