DiReCT: Disentangled Regularization of Contrastive Trajectories for Physics-Refined Video Generation

1. 1Identify the limitations of existing flow-matching video generators in capturing physically consistent dynamics
2. 2Develop a contrastive flow matching approach to distinguish between physically consistent and impossible dynamics
3. 3Implement disentangled regularization to push apart velocity-field trajectories of differing conditions
4. 4Evaluate the effectiveness of DiReCT in generating physics-refined videos

ML researchers and engineers working on video generation tasks can benefit from this approach to produce more realistic and physically consistent outputs, and it can be applied by ai-engineers and ml-researchers

💡 Disentangled regularization of contrastive trajectories can help produce more realistic and physically consistent video outputs

💡 DiReCT: Disentangled regularization for contrastive trajectories improves physics-refined video generation