Flow Matching-based Foundation Model (FlowFM)

• FlowFM is a generative foundation-model paradigm that employs continuous-time flow matching to create reusable representations and priors.
• It jointly optimizes a representation encoder and a conditional flow-matching generator, achieving notable speedups and improved training efficiency over diffusion-based methods.
• The framework is extensible across various modalities such as wearable sensors, images, videos, and cryo-EM, enabling robust downstream alignment and inverse-problem solutions.

Flow Matching-based Foundation Model (FlowFM) denotes a foundation-model paradigm in which the core generative object is a learned continuous-time flow or velocity field rather than a diffusion denoiser. In current arXiv usage, the term has two closely related meanings. In the narrow sense, FlowFM is a specific self-supervised framework for wearable sensor data that jointly trains a representation encoder and a conditional flow-matching generator (Ukita et al., 17 Dec 2025). In the broader sense, it refers to large flow-matching generative foundation models such as FLUX, Qwen-Image, and WAN video, together with downstream alignment, adaptation, and inverse-problem pipelines built around them (Ping et al., 13 Feb 2026). Across both senses, the shared premise is that simulation-free conditional velocity regression can serve as the pretraining primitive for reusable representations, reusable priors, or both (Lipman et al., 2022, Lipman et al., 2024).

1. Terminological scope and conceptual definition

The term FlowFM is not used uniformly across the literature. One line of work introduces FlowFM as a concrete model for wearable-sensor SSL, with a decoupled representation encoder $f_\phi$ and conditional flow-matching generator $v_\theta$ trained jointly from unlabeled data (Ukita et al., 17 Dec 2025). Another line uses FlowFM as an umbrella label for large flow-matching foundation models in image and video generation, especially when discussing RL alignment, modular training systems, or parameter-efficient adaptation of models such as FLUX (Ping et al., 13 Feb 2026).

Related papers extend the same design logic to domain-specific foundation models and foundation priors. CryoFM learns a global prior over high-quality cryo-EM density maps and reuses it across denoising, missing-wedge restoration, and ab initio modeling without fine-tuning (Zhou et al., 2024). Flow Marching builds a generative PDE foundation model from a bridge-based flow-matching objective, a Physics-Pretrained Variational Autoencoder, and a Flow Marching Transformer (Chen et al., 23 Sep 2025). Cross-spectral detection work shows that a single pretrained flow-matching vision model, FLUX.1 Kontext, can be adapted with LoRA into an RGB$\rightarrow$IR or RGB$\rightarrow$SAR translator using only 100 paired examples per domain (Clouser et al., 7 Jan 2026).

A common misconception is that FlowFM names a single fixed architecture. The literature instead supports a narrower interpretation, in which it is a specific wearable-sensor model, and a broader interpretation, in which it is a family of foundation-model constructions built on flow matching. This suggests that the stable unifying feature is not a particular backbone but the use of flow-matching objectives to produce reusable generative operators, reusable priors, or jointly learned representations.

2. Mathematical basis in flow matching

The formal substrate of FlowFM is the flow-matching view of Continuous Normalizing Flows. A time-dependent vector field $v_t(x)$ defines a flow map $\phi_t$ by

$$\frac{d}{dt}\phi_t(x) = v_t(\phi_t(x)), \qquad \phi_0(x)=x,$$

and the induced density path $p_t$ satisfies the continuity equation

$$\frac{d}{dt}p_t(x) + \operatorname{div}(p_t(x)v_t(x))=0.$$

Flow Matching trains $v_t$ by regression to a target vector field along a chosen probability path, rather than by simulation-heavy likelihood optimization (Lipman et al., 2022).

The standard conditional path used throughout the literature is the linear interpolation

$v_\theta$0

together with a Conditional Flow Matching objective of the form

$v_\theta$1

The critical scalability result is the marginalization trick: the tractable conditional regression objective has the same gradient as the intractable marginal FM objective up to a constant, so training can proceed without integrating the model dynamics during optimization (Lipman et al., 2022, Lipman et al., 2024).

In the specific FlowFM formulation for wearable sensors, the generic condition $v_\theta$2 is replaced by a learned representation $v_\theta$3, yielding

$v_\theta$4

The paper further interprets this as

$v_\theta$5

emphasizing that the encoder parameters $v_\theta$6 do not merely supply side information but parameterize the effective flow problem itself (Ukita et al., 17 Dec 2025).

The broader FM literature also makes clear that diffusion is not a disjoint alternative. Diffusion paths are a special case of Gaussian conditional paths in FM, while OT displacement interpolation provides a straighter and often more efficient alternative (Lipman et al., 2022). The later guide generalizes this picture further, presenting FM as part of a generator-matching framework spanning Euclidean spaces, manifolds, discrete spaces, and arbitrary continuous-time Markov processes (Lipman et al., 2024).

3. Canonical FlowFM architecture and training regime

The most explicit architecture named FlowFM consists of two trainable components: a representation encoder $v_\theta$7 and a velocity field network / conditional flow matching generator $v_\theta$8 (Ukita et al., 17 Dec 2025). The encoder is Transformer-based, using a ViT-like architecture adapted to 1D wearable sensor signals by replacing the patch embedding with a 1D convolution. The generator is based on DiT, also adapted to 1D sensor signals via a 1D convolutional patch embedding, and receives the intermediate state $v_\theta$9, timestep $\rightarrow$0, and representation $\rightarrow$1. Conditioning is injected into each transformer block with adaLN-Zero.

The central architectural claim is a decoupled design with joint optimization. The encoder specializes in discriminative feature learning, the generator specializes in high-fidelity generation, and both are updated through the same flow-matching objective. This differs from pipelines that freeze a pretrained encoder and use it only as a static condition. FlowFM also introduces Dynamic Guidance Switching, in which the representation $\rightarrow$2 is randomly masked to zero with probability 50% during training, preventing over-reliance on the condition and improving downstream classification accuracy on all datasets compared with 0% masking (Ukita et al., 17 Dec 2025).

Pretraining follows the usual SSL pattern: sample $\rightarrow$3, construct $\rightarrow$4, compute $\rightarrow$5, feed $\rightarrow$6 into the generator, and optimize the squared error between predicted and target velocity fields. The pretraining corpus is Capture-24, and downstream evaluation spans five HAR datasets—ADL, Opportunity, PAMAP2, REALWORLD, and WISDM—as well as a text-to-signal setting (Ukita et al., 17 Dec 2025).

Empirically, this FlowFM reduces training time by 50.4% relative to a diffusion-based baseline: on RTX 4090 with batch size 1024, FlowFM: 3.2 hours versus DiffFM: 6.5 hours. For generation it reports $\rightarrow$7 s/sample versus $\rightarrow$8 s/sample for DiffFM (1000 steps), a 51.0x speedup, while maintaining FID = 5.014 compared with FID = 5.047 for DiffFM at 1000 steps (Ukita et al., 17 Dec 2025). On downstream recognition, it achieves the best accuracy and F1 on all five datasets in both transfer learning and fine-tuning, with the paper highlighting up to +19.75% on Opportunity in fine-tuning relative to SSL-Wearables (Ukita et al., 17 Dec 2025).

4. Representative instantiations across modalities

In vision, a prominent FlowFM-style instantiation starts from FLUX.1 Kontext, described as a 12B-parameter latent rectified-flow transformer trained on large-scale RGB data, and adapts it into a cross-spectral translator with LoRA (Clouser et al., 7 Jan 2026). The model is conditioned on an RGB image and a fixed instruction such as “Convert this to an IR image from the KAIST sensor” or “Convert this to a SAR image,” allowing scene structure to be preserved while shifting modality. LoRA modules are attached to the query, key, value, and output projections in every self-attention layer, and to the linear projections in the MLP sub-blocks, with only about ~1% of the base model trained. With only 100 paired RGB-target examples per domain, the adapted model translates RGB into pixel-aligned IR or SAR, so the original bounding boxes can be reused for target-modality detection (Clouser et al., 7 Jan 2026).

The same paper establishes a pragmatic model-selection result: LPIPS computed on only 50 held-out pairs is a strong proxy for downstream performance. Lower LPIPS consistently predicts higher mAP for YOLOv11n on both IR and SAR, and for DETR on KAIST IR test data, with reported correlations including Pearson $\rightarrow$9 for KAIST + YOLOv11n and Pearson $\rightarrow$0 for M4-SAR + YOLOv11n (Clouser et al., 7 Jan 2026). Using the best LPIPS-selected adapter, synthetic IR from external RGB datasets improves KAIST IR pedestrian detection for YOLOv11n from 0.50 ± 0.02 to 0.54 ± 0.02 [email protected], while synthetic SAR combined with limited real SAR improves M4-SAR detection from 0.19 ± 0.01 to 0.25 ± 0.01 [email protected] (Clouser et al., 7 Jan 2026).

In structural biology, CryoFM is a flow-matching foundation model for biomolecular density maps. It is trained unconditionally on 3479 density maps curated from EMDB and then reused as a prior through flow posterior sampling for spectral noise denoising, anisotropic noise denoising, missing-wedge restoration, and ab initio modeling, all without fine-tuning (Zhou et al., 2024). The model uses a hierarchical Transformer based on HDiT, adapted to 3D density volumes, and the paper reports 0 fail rate in the denoising tasks it studies, together with strong FSC-AUC improvements such as a missing-region FSC-AUC increase from 0.00 to 0.76 in missing-wedge restoration (Zhou et al., 2024).

In scientific computing, Flow Marching constructs a generative PDE foundation model by jointly sampling noise level and physical time step between adjacent states. The full system combines P2VAE, which compresses states from c3p128 to c16p16 at a 12× compression rate, with an FMT backbone that integrates diffusion forcing and latent temporal pyramids (Chen et al., 23 Sep 2025). The paper curates a corpus of about 2.5 million trajectories across 12 distinct PDE families, reports up to 15× greater computational efficiency than full-length video diffusion models, and shows lower long-horizon rollout error than VICON-88M, for example 0.3048 versus 0.5627 on PA-NS for FMT-L-138M (Chen et al., 23 Sep 2025).

Function-space and forecasting extensions broaden the conceptual perimeter of FlowFM. Functional Flow Matching generalizes FM to infinite-dimensional Hilbert spaces, replacing density-based reasoning with measure-theoretic paths and vector fields, and supports super-resolution and conditional generation on time-series and PDE benchmarks (Kerrigan et al., 2023). FlowTime applies autoregressive conditional flow matching to multivariate probabilistic forecasting, factorizing future trajectories stepwise and training a shared flow-conditioned next-step generator (El-Gazzar et al., 13 Mar 2025). These systems are not presented as universal foundation models in the same sense as FLUX or CryoFM, but they demonstrate that the FlowFM recipe is not limited to image-space generation.

5. Adaptation, alignment, distillation, and system infrastructure

A major practical issue for FlowFM is how to reuse strong pretrained models that were not originally trained in FM form. Diff2Flow addresses this by converting a pretrained diffusion model into a flow-matching model through timestep rescaling, interpolant alignment, and velocity-field conversion / objective change (Schusterbauer et al., 2 Jun 2025). The paper argues that naive FM finetuning on a diffusion checkpoint is weak, especially with LoRA, because the model must simultaneously adapt to a different objective, interpolant, and timestep scale. Diff2Flow makes the conversion algebraic rather than architectural, claiming no extra computation overhead, and reports superior or competitive results across text-to-image synthesis, Reflow-style trajectory rectification, and monocular depth estimation, with its advantage most visible under parameter-efficient constraints (Schusterbauer et al., 2 Jun 2025).

A complementary route is sampling acceleration by distillation. Flow Generator Matching distills a pretrained multi-step flow-matching teacher into a one-step student generator while targeting equality of induced vector fields and distributions (Huang et al., 2024). On CIFAR-10 unconditional generation, the one-step FGM model achieves FID 3.08, outperforming the original 50-step teacher at FID 3.67 and approaching the 100-step teacher at FID 2.93. The same paper distills Stable Diffusion 3 Medium into MM-DiT-FGM, which attains 0.65 overall on GenEval at NFE 1, compared with 0.70 for SD3-Medium at NFE 28 (Huang et al., 2024).

As flow-matching foundation models became targets for RL alignment, systems infrastructure also emerged. Flow-Factory provides a registry-based framework that decouples the model, the RL/alignment algorithm, and the reward model through BaseAdapter, BaseTrainer, BaseRewardModel, and SDESchedulerMixin abstractions (Ping et al., 13 Feb 2026). It supports GRPO, DiffusionNFT, and AWM across Flux, Qwen-Image, and WAN video models, along with pointwise and groupwise rewards, multi-reward aggregation, preprocessing-based memory optimization, and distributed training. On 8×H200 GPUs for FLUX.1-dev with PickScore reward over 100 steps, preprocessing reduces peak memory per device from 61.08 GB to 53.14 GB and time per step from 144.02 s to 82.68 s, corresponding to a 13.0% reduction and 1.74× speedup (Ping et al., 13 Feb 2026).

These developments show that FlowFM is not only an objective family but also an increasingly structured systems stack: conversion from diffusion priors, one-step distillation, RL alignment, modular rewards, and PEFT-based specialization all now exist as explicit research directions.

6. Empirical profile, misconceptions, and unresolved questions

A recurrent misconception is that a flow-matching foundation prior automatically transfers well to all downstream tasks. The most systematic counterexample appears in the jet-foundation-model study, which compares supervised classification, flow-matching generation, and masked particle modeling under controlled pretraining and fine-tuning regimes (Elsharkawy et al., 12 Jun 2026). There, pure Generator pre-training is the weakest configuration for top tagging, while flow matching must be in the pre-training objective to see a significant finetuning advantage for downstream generation. The paper therefore argues that classification and generation are empirically and conceptually orthogonal enough that a model intended to transfer to both must be pretrained on both (Elsharkawy et al., 12 Jun 2026).

Another misconception is that a foundation FM prior is already a strong inverse-problem prior by virtue of broad pretraining. The FMPlug line of work argues the opposite: foundation FM priors can lag behind domain-specific or even untrained priors unless the latent optimization respects both the similarity between measurement and target and the Gaussian structure of the source distribution (Wan et al., 1 Aug 2025, Wan et al., 20 Nov 2025). FMPlug introduces a time-adaptive warm-up strategy and sharp Gaussianity regularization, and the later paper reports gains such as 31.83 PSNR / 0.97 SSIM on linear inverse scattering and 23.26 / 0.48 on MRI ($\rightarrow$1), versus 17.15 / 0.66 and 8.94 / 0.15 for D-Flow (Wan et al., 20 Nov 2025). The implication is not that foundation FM priors are ineffective, but that plug-in optimization must remain compatible with the model’s Gaussian latent geometry.

The cross-spectral detection study also makes the limits of current FlowFM reuse explicit. Its method is few-shot, not zero-shot; it still requires 100 paired co-registered examples per modality, studies only RGB$\rightarrow$2IR and RGB$\rightarrow$3SAR, and uses only one backbone, FLUX.1 Kontext (Clouser et al., 7 Jan 2026). The paper further notes that LPIPS measures perceptual similarity rather than radiometric correctness, that synthetic-only SAR still lags real-only SAR, and that DETR gains are limited. These caveats materially narrow what can be inferred from improved mAP or low LPIPS.

Open technical questions recur across the literature. In PDE modeling, the authors of Flow Marching identify the compression bottleneck of P2VAE and leave fuller treatment of epistemic uncertainty for future work (Chen et al., 23 Sep 2025). In cross-spectral vision, future work is directed toward better physical constraints, more modalities, broader detector families, and less paired supervision (Clouser et al., 7 Jan 2026). In jet pretraining, the central unresolved issue is why pretraining loss is a poor proxy for downstream transfer and whether the observed objective hierarchy persists at much larger scales (Elsharkawy et al., 12 Jun 2026).

Taken together, the literature defines FlowFM less as a single model than as a technically coherent foundation-model paradigm: simulation-free flow matching supplies the training objective; large pretrained generators or priors supply reuse; and downstream efficacy depends on how well the chosen adaptation, conditioning, or optimization procedure respects the geometry of the learned flow.