Latent Dynamics · 2026-10-06
Learning Conditional Source Distribution via Flow Reversal for Temporal Flow Matching
Kuan-Hsun Tu, Hsuan-Chi Liu, Jia-Wei Liao, Chien-Sheng Chiang, Tsung-Wei Ke
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
CNP-Flow is a flow matching (FM) framework for temporal generation that learns condition-dependent source distributions. It uses a conditional noise predictor (CNP) to output an isotropic Gaussian per temporal condition, supervised by flow reversal of a pretrained FM model. A three-stage pipeline pretrains FM, trains CNP via likelihood/KL losses, and fine-tunes FM with the learned source prior. The authors report improved generation quality on video prediction/interpolation and 7-DoF Franka motion planning.
Why it matters
Standard conditional FM conditions only the vector field and uses a fixed standard Gaussian source. CNP-Flow instead conditions the source distribution itself, trained from flow-reversal-induced pseudo sources. The authors report better distribution-level alignment (e.g., lower FVD), improved motion-planning outcomes (more valid trajectories, lower errors/collisions), and efficiency via fewer function evaluations.
Method
- Pretrain a conditional FM using a standard Gaussian source; then use flow reversal through the pretrained FM to retrieve per-condition surrogate source samples.
- Train a Conditional Noise Predictor (CNP) to parameterize an isotropic Gaussian source distribution for each temporal condition, using NLL to match flow-reversal-induced sources plus KL regularization toward N(0,I).
- Fine-tune the FM by replacing the standard Gaussian source with CNP-sampled conditional sources, without changing the FM backbone architecture.
Abstract (from 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
Related papers
- $\Delta$ Flow Matching for Robot Foundation Models
- PRIME W0: Physical Priors from World Models for Embodied Manipulation
- Residual Modeling Closes the Regression and Generative Policy Gap in Robot Learning
- WorldFlow: Joint Video and 3D Motion Prediction via Large-Scale Pretraining
- Video2STL: Grounding VLM-Generated Temporal Specifications for Robot Learning
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-04 · Indexed by Pub-AI: 2026-10-06 UTC
Learning Conditional Source Distribution via Flow Reversal for Temporal Flow Matching
Preprint posted: 2026-10-04 · Original record ↗
- arxiv
- 2610.05349
Explicit source versions (1)
- v2 · 2026-10-06
- Preprint available
Source date: 2026-10-04 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
- Explicit source version detected · v2
Source date: 2026-10-06 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
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
