Flow-Anything: Learning Real-World Optical Flow Estimation from Large-Scale Single-view Images

demo.mp4

📢 Project Status

We are actively organizing and cleaning up the full codebase (training, inference, evaluation).
The repository will be continuously updated in the coming weeks — please stay tuned for:

🛠️ Full training scripts
📦 Data preparation tools
📈 Evaluation pipelines
🔖 Additional pre-trained models for different datasets

🚀 Pre-trained Checkpoints

You can find the released pre-trained checkpoints here:
👉 Flow-Anything Checkpoints

Quick Start

installation

conda create --name SEA-RAFT python=3.10.13
conda activate SEA-RAFT
pip install -r requirements.txt

inference

python infer.py \
    --input [path to images] \
    --out [path to save] \
    --cfg config/eval/sintel-M.json \
    --model [path to ckpt]

✅ TODO

Data generation and pre-training code
(including dataset preprocessing, augmentation, and full training pipeline)
Inference and evaluation code on Point Tracking tasks
(standardized pipelines for Point Tracking benchmarks and visualization)
Clean inference-optimized code
(lightweight, modular implementation specifically designed for fast deployment & real-time inference)

Thank you for your interest and support! ⭐️
Feel free to open an issue if you have any questions or suggestions.