Temporal Flow Matching (TFM)

• Temporal Flow Matching (TFM) is a framework that integrates temporal dynamics into flow matching to generate predictions for evolving trajectories.
• It enables prediction of complex data types such as 4D medical images, clinical time series, and event sequences through adaptive source distributions and coupling methods.
• Empirical studies show that TFM-based models consistently outperform traditional spatio-temporal baselines, achieving state-of-the-art results in diverse applications.

Temporal Flow Matching (TFM) denotes a class of temporally structured flow-matching constructions in which the generative transport is tied to sequential state evolution, past context, multiple time variables, or temporally informed source distributions. In (Disch et al., 29 Aug 2025), TFM is introduced for 4D longitudinal medical image prediction as a unified generative trajectory method that aims to learn the underlying temporal distribution, can fall back to predicting the last context image (LCI) as a special case, and supports $3D$ volumes, multiple prior scans, and irregular sampling; the abstract further states that it consistently surpasses spatio-temporal methods from natural imaging on three public longitudinal datasets (Disch et al., 29 Aug 2025). Across adjacent literature, however, the same name or closely related formulations are used for simulation-free Neural SDE training on clinical time series, causal latent forecasting for robotic manipulation, one-step transition operators for generative modeling, and non-autoregressive marked temporal point processes (Zhang et al., 2024, Wang et al., 20 Apr 2026, Ma, 16 Mar 2026, Shou, 6 Aug 2025). This suggests that TFM is not yet a single universally standardized algorithm, but rather a recurring design pattern in which flow matching is adapted to temporal structure.

1. Terminological scope and historical placement

The most direct use of the name appears in "Temporal Flow Matching for Learning Spatio-Temporal Trajectories in 4D Longitudinal Medical Imaging" (Disch et al., 29 Aug 2025). That abstract positions TFM as a generative trajectory method for longitudinal imaging, with three explicit design claims: it aims to learn the underlying temporal distribution, it can degenerate to a nearest-image predictor by outputting the last context image, and it supports multiple prior scans, irregular sampling, and volumetric $3D$ data. Within the supplied material, however, no substantive method section, theory, ablation, or architectural description from that paper is available. Accordingly, only these abstract-level claims can be stated precisely.

Outside that medical-imaging usage, the same acronym or an immediately adjacent usage appears in several distinct settings. "Trajectory Flow Matching with Applications to Clinical Time Series Modeling" names TFM as a simulation-free training method for Neural SDEs on stochastic, irregularly sampled trajectories (Zhang et al., 2024). "Transition Flow Matching" uses TFM for a two-time transition operator that maps a state at time $t$ directly to a future state at time $r$ (Ma, 16 Mar 2026). OFlow describes its foresight module as "temporal flow matching" for causal prediction of future semantic latents in a shared representation space (Wang et al., 20 Apr 2026). "Unified Flow Matching for Long Horizon Event Forecasting" presents a marked temporal point process model that is explicitly summarized in the supplied material as Temporal Flow Matching for long-horizon event forecasting (Shou, 6 Aug 2025).

A recurrent misconception is therefore that TFM already denotes one canonical formalism. The current literature does not support that reading. The phrase is applied to multiple, technically non-equivalent objects: conditional velocity regression over future latent frames, bridge-like trajectory matching for time series, direct finite-time transition prediction, and joint continuous/discrete sequence transport. The commonality is temporalization of flow matching, not a single shared parameterization.

2. Core mathematical pattern

The common baseline is standard flow matching: a model learns a vector field on an interpolant between a source variable and a target variable. In OFlow, the conditional future-latent objective is written as

$$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$

with the future latent $z^i$ predicted from Gaussian-noised interpolation conditioned on the history $z^{<i}$ (Wang et al., 20 Apr 2026). WarmPrior keeps the same standard flow-matching loss and linear interpolant,

$a_t = (1-t)a_0 + t a_1,\qquad \dot a_t = a_1-a_0,$

but changes the source distribution so that generation begins from a temporally grounded prior rather than $\mathcal N(0,I)$ (Kang et al., 13 May 2026). This makes temporal information enter through the coupling rather than through a modified loss.

In clinical time series, Trajectory Flow Matching constructs conditional path distributions over observed trajectories and trains drift and diffusion terms directly, bypassing backpropagation through SDE dynamics. Its matching loss is

$$\mathcal{L}_\text{match}(\theta) = \mathbb{E}_{t, q(z), p_t(x \mid z)} \left[ \left\| f_\theta(t, x_t) - v_t(x_t \mid z) \right\|^2 + \lambda_t^2 \left\| g_\theta(t, x_t) - \sigma_t(x_t \mid z) \right\|^2 \right],$$

and, for evenly sampled trajectories, it uses a Gaussian bridge between adjacent observations together with the conditional flow

$3D$0

A central theoretical point is that ordinary flow matching does not automatically preserve trajectory couplings when trajectories cross in state space; the paper gives sufficient conditions, including history-based disambiguation, under which the learned coupling matches the data coupling (Zhang et al., 2024).

Transition Flow Matching departs more radically from local vector-field learning. Instead of learning only $3D$1, it learns a global transition map

$3D$2

and grounds this in the Transition Flow Identity

$3D$3

The practical consequence is that the model can generate in one step through $3D$4 or on an arbitrary time grid without explicit ODE integration of a local velocity field (Ma, 16 Mar 2026).

In marked temporal point processes, UFM-TPP temporalizes flow matching jointly over continuous inter-event times and discrete event types. The continuous path is

$3D$5

with $3D$6, while the discrete path for marks is

$3D$7

with a uniform categorical source. The final objective is

$3D$8

so temporal flow matching here is explicitly hybrid: continuous for timestamps and discrete for marks (Shou, 6 Aug 2025).

3. Principal formulations across domains

The term is most useful when disambiguated by state space and temporal mechanism rather than by acronym alone.

Work	State space	Temporal mechanism
(Disch et al., 29 Aug 2025)	$3D$9 longitudinal medical images	Unified generative trajectory method; LCI fallback; multiple prior scans; irregular sampling
(Zhang et al., 2024)	Clinical time series	Conditional trajectory bridges; simulation-free Neural SDE training; next-time and uncertainty prediction
(Wang et al., 20 Apr 2026)	Semantic video latents for VLA control	Causal autoregressive future-latent forecasting with conditional flow matching
(Shou, 6 Aug 2025)	Marked temporal point processes	Joint continuous/discrete non-autoregressive future-event generation
(Ma, 16 Mar 2026)	Continuous generative trajectories	Direct finite-time transition map for arbitrary-step or one-step generation

In the medical-imaging formulation, the abstract alone indicates a trajectory model over longitudinal scans, with the ability to reduce to last-context-image prediction as a special case (Disch et al., 29 Aug 2025). In the clinical time-series formulation, TFM is a training principle for stochastic continuous-time dynamics, explicitly adapted to irregular sampling and uncertainty prediction (Zhang et al., 2024). In OFlow, temporal flow matching is not about raw pixels but about framewise semantic latent states, with the next latent frame predicted causally from previous ones (Wang et al., 20 Apr 2026). In UFM-TPP, the temporal structure is long-horizon event forecasting without sequential decoding, enabled by a conditional independence assumption over future events given the context (Shou, 6 Aug 2025). In Transition Flow Matching, time becomes a pair $t$0 and the learned object is the finite-time transport itself (Ma, 16 Mar 2026).

A plausible implication is that TFM has become a reusable abstraction for temporally structured generative transport: the state may be a volume, a latent frame, a clinical trajectory, or a marked event horizon, but the modeling move is to make temporal structure explicit inside the transport map or its conditioning variables.

4. Temporal conditioning, architectures, and training schemes

How temporal information enters the model differs sharply across implementations. OFlow uses a Diffusion Transformer backbone with a block-causal attention mask over time and dense attention within each frame. Video is encoded into DINOv2 latent features, future latents are forecast autoregressively, and the predicted framewise embeddings are factorized into object-aware prototypes by $t$1-means. The appendix specifies cluster counts $t$2 with up to 8 iterations. Those object-aware future representations are then injected into a pretrained VLA action expert based on Eagle-2.5/GR00T-N1.5 through a ControlNet-style cross-attention mechanism. Training is explicitly two-stage: first train the foresight model on latent future prediction, then freeze the VLM and foresight model and finetune the action head. The foresight backbone is a 12-layer transformer trained on video tubes of length 10 to predict 4 future frames, and inference uses 4 denoising steps together with a 16-action chunk (Wang et al., 20 Apr 2026).

Trajectory Flow Matching introduces temporal structure through trajectory-conditioned conditional paths and through auxiliary heads. For irregular sampling it adds a time predictor

$t$3

trained with

$t$4

It also introduces a target-prediction reparameterization, in which the model predicts the next observation $t$5 rather than the velocity directly, using the identity

$t$6

A separate uncertainty head $t$7 is trained from prediction error. In this formulation, temporal flow matching is as much about preserving trajectory identity under crossing paths as it is about fitting marginal distributions (Zhang et al., 2024).

UFM-TPP uses a shared context encoder rather than autoregressive rollout. Historical events are embedded as

$t$8

and an RNN encoder produces a context vector $t$9 that conditions both the continuous-time vector field for inter-event times and the discrete classifier for marks. Inference is fully batchwise and non-autoregressive: inter-event times are updated by a second-order midpoint method, marks are updated in the simplex and resampled from the resulting categorical distribution, and the reported generation cost is fewer than 10 flow steps (Shou, 6 Aug 2025).

WarmPrior shows a different temporalization route. It replaces the standard Gaussian source with a temporally grounded Gaussian prior centered on recent action history. The paper instantiates WP-Past, which centers the source on the previous executed chunk, and WP-Preview, which centers it on the model’s previous forecast of the current chunk. The loss is unchanged; the temporal prior changes the coupling seen during training and can be interpreted as an amortized OT-like warm start (Kang et al., 13 May 2026).

A closely related latent-sequence design appears in HO-Flow, where a masked auto-regressive transformer provides temporal context and a flow-matching head generates continuous latent motion tokens for hand-object interaction sequences. The paper explicitly attributes improved temporal coherence to the combination of auto-regressive context accumulation and continuous latent generation rather than discrete token classification (Chen et al., 12 Apr 2026). Although not named TFM, it exemplifies the same temporalized flow-matching pattern.

5. Empirical record

The available evidence is application-specific rather than unified. For longitudinal medical imaging, the abstract of (Disch et al., 29 Aug 2025) reports that TFM consistently surpasses spatio-temporal methods from natural imaging across three public longitudinal datasets and establishes a new state of the art, but the supplied material contains no detailed metrics, datasets, or ablations beyond that statement.

In clinical time series, Trajectory Flow Matching and its deterministic variant TFM-ODE outperform NeuralODE, NeuralSDE, LatentODE, and an FM baseline on multiple clinical datasets. On ICU Sepsis, the reported MSEs $r$0 are $r$1 for NeuralODE, $r$2 for NeuralSDE, $r$3 for TFM-ODE, and $r$4 for TFM. On ICU GIB, the corresponding values are $r$5, $r$6, $r$7, and $r$8. The paper also reports stronger uncertainty prediction for TFM than for TFM-ODE and improved RBF-MMD, for example $r$9 for TFM versus $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$0 for NeuralSDE on ICU GIB (Zhang et al., 2024).

OFlow attributes much of its robustness under distribution shift to temporal flow matching in semantic latent space. On LIBERO-Plus it reports the best overall success rate of $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$1, exceeding GR00T-N1.5 and $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$2-family baselines. In the summarized perturbation comparison it reports gains of $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$3, $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$4, $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$5, and $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$6 over GR00T-N1.5 on camera, robot, language, and layout-like shifts. In the LIBERO-Plus ablation, the full model with DINO, object-awareness, and foresight reaches $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$7 average success on the evaluated subset, and the appendix reports that under real-world background, object-instance, and human perturbations on Apple-Pot, OFlow drops only to around $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$8 whereas baselines fall below $$\mathcal{L} = \mathbb{E}\!\left[ \left\| \, v_{\theta}\!\left( z^{i}_{t}, t \,\big|\, z^{<i} \right) - (z^i - \epsilon) \right\|_{2}^{2} \right],$$9 (Wang et al., 20 Apr 2026).

UFM-TPP reports best Optimal Transport Distance on 4 of 6 datasets and best RMSE$z^i$0 on 5 of 6 datasets. It highlights, for example, OTD $z^i$1 versus $z^i$2 on StackOverflow and $z^i$3 versus $z^i$4 on Taobao, together with strong sMAPE gains such as $z^i$5 versus $z^i$6 for CDiff on StackOverflow. It also emphasizes efficiency: fewer than 10 flow steps, with sampling times such as 7s versus 86s on Taxi, 16s versus 204s on Taobao, and 12s versus 175s on Amazon (Shou, 6 Aug 2025).

Transition Flow Matching is evaluated primarily in few-step image generation. On CIFAR-10 it reports FID $z^i$7 at 1 NFE, $z^i$8 at 2 NFE, $z^i$9 at 5 NFE, and $z^{<i}$0 at 10 NFE. On ImageNet-256 with a latent-space transformer backbone, it reports FID $z^{<i}$1 at 1 step and $z^{<i}$2 at 2 steps (Ma, 16 Mar 2026). These results matter because the model is not distilled from a pretrained fast sampler; the transition operator is trained directly.

WarmPrior, although not presented as a named TFM method, provides empirical evidence for temporally grounded source distributions within flow matching. The paper reports improved success rates over standard Gaussian-source policies across Robomimic and MimicGen, gains that are larger at lower NFE, improved performance on a Franka Research 3 across four tabletop tasks, and stronger prior-space RL fine-tuning on Robomimic Square and Transport, with WarmPrior variants exceeding $z^{<i}$3 on Square and around $z^{<i}$4 on Transport (Kang et al., 13 May 2026).

6. Conceptual boundaries, adjacent methods, and limitations

TFM should be distinguished from several adjacent temporal ideas in flow-based modeling. Temporal Pair Consistency couples paired timesteps along the same probability path to reduce estimator variance, but it does so entirely at the training-objective level and explicitly does not change the architecture, probability path, or solver. Its focus is variance reduction and temporal coherence of the learned vector field, not temporal forecasting or trajectory generation in the narrower TFM sense (Maduabuchi et al., 4 Feb 2026). Blockwise Flow Matching partitions the interval $z^{<i}$5 into temporal segments and assigns a specialized velocity block to each one; this is a temporal decomposition of the generative trajectory, but again not a general TFM formalism (Park et al., 24 Oct 2025).

Other nearby methods operate on temporal structure without using the name. Edit-Based Flow Matching for Temporal Point Processes models generation as a continuous-time Markov chain over insert, delete, and substitute edits; the supplied synthesis explicitly describes it as a TFM-style transport in discrete edit space, but the paper itself does not use the label as a named method (Lüdke et al., 7 Oct 2025). Distillation work on two-time flow models likewise learns maps $z^{<i}$6 between arbitrary times, but presents this as TTFM distillation rather than Temporal Flow Matching proper (Khungurn et al., 2 May 2025). In alignment for flow-matching image generation, VGPO is motivated by the temporally structured nature of the reverse-time process, yet it addresses temporal credit assignment and groupwise reward collapse rather than generative transport design (Shao et al., 13 Dec 2025).

Several limitations are explicit. For (Disch et al., 29 Aug 2025), only the abstract and a note about missing substantive paper content are available in the supplied material, so no verified statement can be made here about its objective function, architecture, benchmark details, or ablations beyond the abstract claims. UFM-TPP states a key modeling limitation directly: it assumes conditional independence among future events given the context, a simplification that enables non-autoregressive generation but may reduce expressiveness for highly structured futures (Shou, 6 Aug 2025). Trajectory Flow Matching notes that memory conditioning is not always useful on sparser or shorter trajectories and that the method does not yet provide causal representation learning; it also emphasizes the need for prospective validation, calibration, and bias assessment in clinical use (Zhang et al., 2024). Transition Flow Matching relies on Jacobian-vector products and time-pair sampling design, so its practical behavior depends on higher-order autodiff support, loss weighting, and the chosen $z^{<i}$7 sampling strategy (Ma, 16 Mar 2026).

The main conceptual boundary is therefore clear. Temporal Flow Matching is not simply “flow matching with a time variable,” because ordinary flow matching already has continuous time. What distinguishes TFM-style methods in the cited literature is that the temporal structure of the target problem becomes an explicit modeling object: trajectory identity must be preserved, future latent states must be causally extrapolated, future event horizons must be generated jointly, or finite-time transitions must be predicted directly. Where the term is used precisely, it names that shift from generic interpolation to structured temporal transport.