Learning Conditional Source Distribution via Flow Reversal for Temporal Flow Matching
On the map: Diffusion models → Diffusion and flow policies for robotics
Abstract (arXiv)
We introduce CNP-Flow, a flow matching framework for temporal generation that learns conditional source distributions through flow reversal. Whereas standard conditional flow matching (FM) incorporates conditioning through the vector field and draws source samples from a standard Gaussian, CNP-Flow uses a conditional noise predictor (CNP) to produce an isotropic Gaussian source for each temporal condition. The CNP is supervised by source samples obtained through flow reversal, which maps observed targets backward through a pretrained FM model. A three-stage pipeline pretrains the FM model, trains the CNP, and fine-tunes the FM model using the learned source distribution, while preserving the FM backbone architecture. Across video prediction, video interpolation, and 7-DoF Franka robot motion planning, CNP-Flow consistently improves generation quality. It also matches baseline performance with fewer function evaluations. Project page: https://embodiedai-ntu.github.io/cnpflow
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Diffusion and flow policies for robotics, one of 14 directions in Diffusion models. Builds on 46 of its 53 references that are in the map. First posted on arXiv 4 Oct 2026.
Builds on
46 of 53 references.
Closest ancestors · 4
Same direction, newest first
- 20253D FlowMatch Actor: Unified 3D Policy for Single- and Dual-Arm Manipulation14 Aug
- 2025Steering Your Diffusion Policy with Latent Space Reinforcement Learning18 Jun
- 2024Motion Planning Diffusion: Learning and Adapting Robot Motion Planning with Diffusion Models27 Dec
- 2023Motion Planning Diffusion: Learning and Planning of Robot Motions with Diffusion Models3 Aug
Other references · 42
- 2017Attention Is All You Needcited by 60,361
- 2020Denoising Diffusion Probabilistic Modelscited by 21,076
- 2020Score-Based Generative Modeling through Stochastic Differential Equationscited by 9,122
- 2021Diffusion Models Beat GANs on Image Synthesiscited by 8,255
- 2022Scalable Diffusion Models with Transformerscited by 6,054
- 2022Flow Matching for Generative Modelingcited by 5,702
- 2024Scaling Rectified Flow Transformers for High-Resolution Image Synthesiscited by 4,181
- 2022Video Diffusion Modelscited by 2,305
- 2021SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equationscited by 1,937
- 2022Planning with Diffusion for Flexible Behavior Synthesiscited by 1,061
- 2023Improving and generalizing flow-based generative models with minibatch optimal transportcited by 879
- 2024Latte: Latent Diffusion Transformer for Video Generationcited by 546
30 more
- 2019CLEVRER: CoLlision Events for Video REpresentation and Reasoningcited by 484
- 2022MAGVIT: Masked Generative Video Transformercited by 346
- 2023VideoFusion: Decomposed Diffusion Models for High-Quality Video Generationcited by 331
- 2022MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolationcited by 309
- 2023Preserve Your Own Correlation: A Noise Prior for Video Diffusion Modelscited by 285
- 2023Multisample Flow Matching: Straightening Flows with Minibatch Couplingscited by 278
- 2018Learning Blind Video Temporal Consistencycited by 253
- 2022Diffusion Probabilistic Modeling for Video Generationcited by 238
- 2022FILM: Frame Interpolation for Large Motioncited by 153
- 2024iVideoGPT: Interactive VideoGPTs are Scalable World Modelscited by 146
- 2017Self-Supervised Visual Planning with Temporal Skip Connectionscited by 139
- 2022Diffusion Models for Video Prediction and Infillingcited by 135
- 2023CADS: Unleashing the Diversity of Diffusion Models through Condition-Annealed Samplingcited by 129
- 2024ReNO: Enhancing One-step Text-to-Image Models through Reward-based Noise Optimizationcited by 119
- 2025Go-with-the-Flow: Motion-Controllable Video Diffusion Models Using Real-Time Warped Noisecited by 117
- 2020Stochastic Latent Residual Video Predictioncited by 108
- 2021FitVid: Overfitting in Pixel-Level Video Predictioncited by 87
- 2023Generating images of rare concepts using pre-trained diffusion modelscited by 79
- 2021CCVS: Context-aware Controllable Video Synthesiscited by 69
- 2024Probabilistic Forecasting with Stochastic Interpolants and F"ollmer Processescited by 66
- 2022Efficient Video Prediction via Sparsely Conditioned Flow Matchingcited by 55
- 2025How I Warped Your Noise: a Temporally-Correlated Noise Prior for Diffusion Modelscited by 32
- 2024TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Modelscited by 31
- 2023The Lottery Ticket Hypothesis in Denoising: Towards Semantic-Driven Initializationcited by 23
- 2026ACWM-Phys: Investigating Generalized Physical Interaction in Action-Conditioned Video World Modelscited by 10
- 2026Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matchingcited by 9
- 2025CAR-Flow: Condition-Aware Reparameterization Aligns Source and Target for Better Flow Matchingcited by 7
- 2025Is There a Better Source Distribution than Gaussian? Exploring Source Distributions for Image Flow Matchingcited by 5
- 2026COT-FM: Cluster-wise Optimal Transport Flow Matchingcited by 3
- 2024Efficient Continuous Video Flow Model for Video Predictioncited by 1
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2025FlowMP: Learning Motion Fields for Robot Planning with Conditional Flow Matching
- 2026ConFlow: Constraints-Guided Learning with Flow Matching for Motion Generation
- 2024Reflected Flow Matching
- 2026Difficulty-Calibrated Interpolation Paths for Conditional Flow Matching
- 2026Coreset-Induced Conditional Velocity Flow Matching
- 2026Flowing With Purpose: Latent Action Guided Flow Matching Policies For Robotic Manipulation
- 2026Velocity Scheduled Flow Matching
- 2024Flow Matching Posterior Sampling: A Training-free Conditional Generation for Flow Matching