Flow-matching-style generative models learn time-dependent vector fields, but standard objectives supervise each timestep independently, leaving shared-path temporal structure unused. We introduce Temporal Pair Consistency (TPC), a simple training-time objective that couples velocity predictions at paired timesteps along the same probability path. TPC leaves the architecture, probability path, sampler, and inference-time computation unchanged. Under matched backbones, training budgets, solvers, and numbers of function evaluations, TPC improves sample quality across flow matching, OT-CFM, rectified flow, DiT-style backbones, and MeanFlow. On CIFAR-10, TPC improves FM from 6.35 to 3.19 FID and OT-CFM from 3.58 to 2.90 FID at the same NFE. Mechanism diagnostics show lower measured gradient variance, higher paired-gradient correlation, and reduced temporal roughness, while random and local pairing controls are substantially weaker. Higher-resolution ImageNet experiments further show consistent gains for DiT-XL/2 at 256 and 512 resolution and for MeanFlow at 256 resolution, all at matched inference cost.
Figures & tables
Figure 1: Temporal Pair Consistency (TPC) at a glance. (a) TPC couples velocity predictions at two timesteps on the same conditional probability path, adding a gated consistency penalty to the standard flow-matching objective. (b) TPC improves FID across five representative settings at matched inference NFE within each comparison. The architecture and sampler remain unchanged; CIFAR-10 TPC training requires approximately 1.68–1.70 × baseline wall time.
Figure 2: Temporal Pair Consistency (TPC) training pipeline. Two timesteps t and t′=ψ(t) are evaluated along the same conditional probability path using shared endpoints and network parameters. The primary prediction retains standard flow-matching supervision, while a Bernoulli-gated auxiliary evaluation contributes the TPC consistency loss. Pairing can be fixed or learned, with optional monotonicity regularization for learned pairing only. Inference uses the original architecture, probability path, sampler, solver, and NFE.
Figure 3: TPC improves sample quality at matched inference cost. Each row compares a baseline model with its TPC-augmented counterpart using the same architecture, probability path, sampler, and number of function evaluations. Endpoint labels show FID values; exact NFEs are reported in the corresponding tables. TPC consistently shifts FID lower across standard FM, OT-CFM, rectified flow, noise-augmented ImageNet flow matching, DiT backbones, and MeanFlow without changing inference-time computation.
Model (IN 128 × 128)
NLL ↓
FID ↓
FM w/ OT Lipman et al. (2023)
2.90
20.9
TPC-FM (Ours)
2.90
18.6
Table 1: Unconditional diffusion and flow-matching results on CIFAR-10, ImageNet 32×32 , ImageNet 64×64 , and ImageNet 128×128 . Blank entries indicate methods that do not report the corresponding unconditional setting. The ImageNet- 128×128 panel compares FM w/ OT and TPC-FM. Table 3 reports the separate noise-augmented conditional ImageNet evaluations, with ImageNet- 128×128 in panel (a) and ImageNet- 64×64 in panel (b).
Method
Relative steps
Relative wall time
Peak memory
FID ↓
FM
1.00×
1.00×
1.00×
6.35±0.11
FM, longer training
1.70×
1.70×
1.00×
5.48±0.09
FM + second-timestep supervision, no consistency
1.00×
1.68×
1.04×
5.72±0.10
FM + random pairing
1.00×
1.68×
1.04×
5.95±0.12
FM + local pairing
1.00×
1.68×
1.04×
5.20±0.09
Fixed TPC
1.00×
1.68×
1.04×
3.59±0.06
Table 5: Equal-compute CIFAR-10 controls. Measurements use matched hardware, precision, batch size, and data pipelines. Relative steps, wall time, and peak memory are normalized to FM. FID values are shown as reported.
Family
Paired quantity
Same-path randomness
Inference change?
FM
velocity vθ(xt,t)
same endpoint pair (x0,x1)
none
OT-CFM
velocity vθ(xt,t)
same minibatch-OT pair / endpoint assignment
none
Rectified Flow
rectified-flow velocity
same rectified path endpoints
none
DiT native velocity model
native velocity prediction
same noise/data pair and timestep schedule
none
DiT ϵ / x0 model
converted velocity-equivalent prediction
same noise/data pair and timestep schedule
none
Noise-augmented FM
velocity under fixed noise protocol
same endpoints and fixed noise augmentation
none
Table 6: Applying TPC across model families. TPC is always computed between two predictions expressed in the same velocity or velocity-equivalent space and evaluated along the same probability path/randomness. The sampler, solver, architecture, and NFE are unchanged for every family.
Method
NFE ↓
Relative training wall time
FID ↓
Consistency-FM
2
1.00×
5.34±0.10
Consistency-FM, longer training
2
1.70×
5.08±0.09
Consistency-FM + fixed TPC
2
1.68×
4.70±0.08
Consistency-FM + learned TPC
2
1.70×
4.29±0.07
Table 7: Same-codebase CIFAR-10 Consistency-FM composition comparison. All variants use the same backbone, dataset, optimizer, evaluation protocol, and 2-NFE sampler. Relative training wall time is normalized to the Consistency-FM reproduction. FID values are mean ± standard deviation over five independent training seeds; the reproduced baseline mean coincides with the published FID of 5.34 .
Variant
FID ↓
Mean ∣ψ(t)−t∣↑
Near-id. frac. ↓
Range of ψ(t)↑
Identity control ψ(t)=t
6.35
0.00
1.00
1.00
Fixed symmetric ψ(t)=1−t
3.59
0.50
0.05
1.00
Learned TPC, best
3.19
0.24
0.04
0.83
Learned TPC, moderate
3.88
0.20
0.12
0.74
Learned TPC, overcoupled
6.22
0.06
0.75
0.45
Table 8: Learned-pairing non-collapse diagnostic. The best learned TPC model maintains substantial separation from identity while achieving the lowest FID. In contrast, the overcoupled setting is much closer to identity and has poor FID, indicating that near-collapse is associated with a failed regime rather than the best-performing learned model.
Method
NFE ( ↓ )
IS ( ↑ )
FID ( ↓ )
Recall ( ↑ )
One-Step Generation + Distil (Euler solver, N=1)
1-Rectified Flow
1
9.08
6.18
0.45
2-Rectified Flow
1
9.01
4.85
0.50
3-Rectified Flow
1
8.79
5.21
0.51
TPC-1RF (ours)
1
9.21
5.86
0.47
TPC-2RF (ours)
1
9.14
4.55
0.53
Table 9: Full rectified-flow and probability-flow comparisons. The main paper reports the adaptive-RK45 full-simulation subset; this table includes both one-step/distillation and full-simulation regimes.
Training phase
TPC-FM / FM grad var ↓
Grad corr ↑
Mid-time var ratio ([0.3,0.7])↓
Early
0.84±0.03
0.34±0.03
0.79±0.03
Mid
0.71±0.02
0.52±0.03
0.66±0.02
Late
0.66±0.02
0.61±0.03
0.60±0.02
Table 10: Training-phase diagnostics for TPC-FM relative to FM on CIFAR-10. The variance and correlation effects strengthen over training.
Figure 4: Training diagnostic behavior. TPC-FM shows lower measured gradient variance than the matched FM baseline throughout training.
Figure 5: Sample quality at matched inference cost . Fréchet Inception Distance (FID ↓ ) versus number of function evaluations (NFE, log scale) on CIFAR-10. Temporal Pair Consistency (TPC) is applied only during training and leaves the inference-time architecture, probability path, sampler, and NFE unchanged. Under the same sampling settings as the corresponding baselines, TPC variants achieve lower FID.
Figure 6: Qualitative samples. Unconditional generations on CIFAR-10, ImageNet-32, and ImageNet-64.
Figure 7: ImageNet-128 qualitative samples. Unconditional generations on ImageNet-128.
Figure 9: Class-conditional ImageNet-512 qualitative samples. Multi-page appendix figure with class-conditional ImageNet-512 comparisons; DiT samples are shown on the left and TPC-DiT samples are shown on the right.
Flow matching (FM) trains a time-dependent vector field that transports samples from a simple prior to a complex data distribution. However, for high-dimensional images, each training sample supervises only a single trajectory and intermediate point, yielding an extremely sparse and high-variance training signal. This under-constrained supervision can cause flow collapse, where the learned dynamics memorize specific source-target pairings, mapping diverse inputs to overly similar outputs, failing to generalize. We introduce Posterior-Augmented Flow Matching (PAFM), a theoretically grounded generalization of FM that replaces single-target supervision with an expectation over an approximate posterior of valid target completions for a given intermediate state and condition. PAFM factorizes this intractable posterior into (i) the likelihood of the intermediate under a hypothesized endpoint and (ii) the prior probability of that endpoint under the condition, and uses an importance sampling scheme to construct a mixture over multiple candidate targets. We prove that PAFM yields an unbiased estimator of the original FM objective while substantially reducing gradient variance during training by aggregating information from many plausible continuation trajectories per intermediate. Finally, we show that PAFM improves over FM by up to 3.4 FID50K across different model scales (SiT-B/2 and SiT-XL/2), different architectures (SiT and MMDiT), and in both class and text conditioned benchmarks (ImageNet and CC12M), with a negligible increase in the compute overhead. Code: https://github.com/gstoica27/PAFM.git.
George Stoica, Sayak Paul, Matthew Wallingford +6
Georgia Tech · University of Washington · Hugging Face +2
Timestep sampling p(t) is a central design choice in Flow Matching models, yet common practice increasingly favors static middle-biased distributions (e.g., Logit-Normal). We show that this choice induces a speed--quality trade-off: middle-biased sampling accelerates early convergence but yields worse asymptotic fidelity than Uniform sampling. By analyzing per-timestep training losses, we identify a U-shaped difficulty profile with persistent errors near the boundary regimes, implying that under-sampling the endpoints leaves fine details unresolved. Guided by this insight, we propose Curriculum Sampling, a two-phase schedule that begins with middle-biased sampling for rapid structure learning and then switches to Uniform sampling for boundary refinement. On CIFAR-10, Curriculum Sampling improves the best FID from 3.85 (Uniform) to 3.22 while reaching peak performance at 100k rather than 150k training steps. Our results highlight that timestep sampling should be treated as an evolving curriculum rather than a fixed hyperparameter.
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