RAW-Flow: Flow-Based Models for Raw Data Processing

• RAW-Flow is a family of methodologies that uses flow-based generative models to map, reconstruct, and enhance raw sensor data across different domains.
• The framework exploits invertible transforms and deterministic latent transport to address challenging inverse problems in RGB-to-RAW reconstruction, audio synthesis, and biomedical localization.
• Applications of RAW-Flow include low-light image enhancement, high-fidelity audio generation, and nanoscale sensing, achieving significant performance gains over conventional methods.

RAW-Flow refers to a family of methodologies and models that leverage flow-based generative frameworks for processing and modeling raw data, with applications ranging from RGB-to-RAW image reconstruction, audio synthesis, low-light RAW image enhancement, to analytical modeling of nanoscale sensing in biomedical systems. These approaches exploit invertible transforms, deterministic or stochastic latent transport, and physically grounded modeling to achieve state-of-the-art fidelity, efficiency, and interpretability in otherwise ill-posed inverse or generative problems.

1. Definitions and Conceptual Overview

RAW-Flow denotes frameworks where the transport or transformation between signal domains, or between corrupted and clean data, is formalized using a flow–based or ODE (ordinary differential equation) generative process in raw or latent space. In image applications, RAW-Flow addresses the challenge of mapping processed RGB images back to linear RAW camera measurements, as in the reverse modeling of ISP (Image Signal Processing) pipelines (Liu et al., 28 Jan 2026). In audio, it designates compact flow-based generative models for raw waveform synthesis directly via invertible mappings (Ping et al., 2019, Kim et al., 2018). In biomedical localization, RAW-Flow describes analytical models for the spatiotemporal statistics of raw event data transmitted by nanoscale sensors under flow-guidance in the bloodstream (Pascual et al., 2023).

Key distinguishing features are:

• Use of flow-matching or invertible mappings to model deterministic or stochastic transport;
• Emphasis on the geometry of data in latent or observation space, surpassing simple regression objectives;
• In several instantiations, explicit conditioning on physical models or sensor physics for higher realism.

2. Deterministic Latent Flow Matching for RGB-to-RAW Reconstruction

RAW-Flow in RGB-to-RAW modeling reframes the ill-posed inverse ISP problem as transport in a latent space, sidestepping issues of color deviation and detail loss inherent to regression-based approaches (Liu et al., 28 Jan 2026). Two autoencoders (one for RGB, one for RAW) map their respective domains into compact latent codes ($z_0, z_1$), between which a deterministic time-dependent vector field $\hat v_\theta(z_t, t)$ is learned. The latent path is the linear interpolation $z_t = (1-t)z_0 + t z_1$; supervised by the flow-matching loss,

$$\mathcal{L}_{\rm flow} = \mathbb{E}_{\,t\sim\mathcal U[0,1]} \|\hat v_\theta(z_t,t) - (z_1-z_0)\|_2^2,$$

and solved by integrating the ODE with Euler steps.

Critical architectural details include dual-domain autoencoders with feature alignment constraints, cross-scale context guidance injecting hierarchical RGB features, and skip connections between RGB and RAW branches. Training follows staged optimization: first autoencoder reconstruction and alignment, then flow matching, with an optional final fine-tuning combining all objectives.

RAW-Flow achieves superior fidelity on FiveK-Nikon, FiveK-Canon, and PASCALRAW, outperforming regression and diffusion-based baselines by 1.7–5 dB in PSNR (Liu et al., 28 Jan 2026).

3. Flow-Based Generative Models for Raw Audio Synthesis

In audio, RAW-Flow refers to invertible, maximum-likelihood generative models operating directly on high-dimensional raw waveform data. Both WaveFlow (Ping et al., 2019) and FloWaveNet (Kim et al., 2018) instantiate such architectures. The models transform the data $x$ to a latent $z$ via a composition of simple invertible flows, optimizing

$$\log p_X(x) = \log p_Z(z) + \sum_{l=1}^L \log|\det \partial f_l(x_{l-1})/\partial x_{l-1}|,$$

where $p_Z(z)$ is typically a standard normal prior.

WaveFlow employs a matrix “squeeze” operation and stacks of dilated 2-D convolutions to achieve a trade-off between parallelism and likelihood fidelity. Its inverse mapping is local affine and autoregressive in one dimension, approaching the expressivity of WaveNet while offering significant speedup ($42.6\times$ real-time with only 5.91M parameters) (Ping et al., 2019).

FloWaveNet similarly combines ActNorm, affine coupling layers, and channel permutations in a multi-scale hierarchy to obtain a single-stage, distillation-free, parallel likelihood model that matches the inference efficiency of teacher–student frameworks (Kim et al., 2018).

4. Physics-Guided Rectified Flow for Low-Light RAW Image Enhancement

RAW-Flow also arises in the design of restoration frameworks for low-light RAW images, exemplified by the Physics-Guided Rectified Flow (PGRF) model (Zeng, 10 Sep 2025). Here, a physics-based per-pixel noise synthesis, derived from first-principles modeling of CMOS noise—including both additive and multiplicative and pixel-level variations—is intimately coupled with a rectified flow ODE-based sampler.

Letting $X_{gt}$ denote the clean image and $X_L$ the noise-injected low-light observation, the ODE is defined over the interpolant $Z_t = tX_{gt} + (1-t)Z_0$ and is guided via a conditional velocity field $v_\theta(Z_t, X_L, t)$, realized as an SR3-style U-Net. Synthesis employs a sampling-step search over timesteps to optimize PSNR.

On the LLID dataset, PGRF surpasses both physically-based and learned baselines, yielding 45.95 dB (SSIM 0.97), outperforming ELD and P-G by significant PSNR margins and delivering superior texture fidelity and artifact removal (Zeng, 10 Sep 2025).

5. Analytical RAW-Flow Modeling in Flow-Guided Biomedical Localization

In nanoscale biomedical sensing, RAW-Flow supplies a complete analytical description of the statistics of raw time-elapsed and event-detection data as transmitted by nanodevices under flow-guided localization (Pascual et al., 2023). The model captures both communication reliability ($p_c$) and energy duty cycle ($\eta$), linking device-level constraints to the distribution of observed data:

$X = (t, b) = (n_1 T_1 + \dots + n_r T_r + Q,\, b).$

The derivation considers permutations of region visits, compound iteration times due to communication failures, and Bernoulli modeling of event detection per region, yielding joint densities for both detected and undetected events.

Comparison with hybrid ns-3/BloodVoyagerS-based simulation demonstrates that the analytical RAW-Flow model closely matches empirical distributions over a wide variety of device parameterizations, passing U-tests and exhibiting low CDF and KL divergence across regions (Pascual et al., 2023).

6. Quantitative Performance and Ablation

Across domains, RAW-Flow frameworks offer quantifiable improvements and allow principled ablations:

• In image reconstruction, deterministic latent flow matching provides $\sim2.5$ dB gain over diffusion or regression baselines; cross-scale context and dual-domain alignment yield further boosts (Liu et al., 28 Jan 2026).
• WaveFlow closes the log-likelihood gap with full autoregressive audio models, attaining MOS scores of $4.32$–$4.43$ while being $8$–$40\times$ faster than real time (Ping et al., 2019).
• In low-light enhancement, physics-guided flows outperform both simulation-guided and learning-based baselines by up to $2.8$ dB (Zeng, 10 Sep 2025).
• Analytical RAW-Flow modeling enables efficient and accurate data generation for downstream biomedical algorithms, supporting rapid design-space exploration (Pascual et al., 2023).

Application Domain	Key Metric	RAW-Flow Performance
RGB-to-RAW (FiveK-Canon)	PSNR / SSIM	32.55 / 0.9445
Audio Synthesis	MOS (WaveFlow)	4.32–4.43
Low-light RAW Enhance	PSNR / SSIM (LLID)	45.95 / 0.97
Nanoscale Localization	U-Tests / D_{KL}	>80% regions pass; KL <0.04

7. Limitations, Generalizations, and Future Directions

RAW-Flow methodologies, while state-of-the-art, exhibit several constraints:

• In image domains, deterministic latent flow may still be challenged by extreme quantization and information loss or by transfer to non-photographic sensors.
• Physics-guided flows rely on correct noise model calibration; Poissonian and dark current assumptions may break at very high ISO, with extreme photon sparsity, or in devices with unpredictable per-pixel process variations.
• Audio flows involve trade-offs between parallelism and likelihood, and may require unlikely permutations for full expressivity. Channel permutation and feature alignment choices can strongly affect final fidelity.
• In analytical sensing models, assumptions of independence and fixed region traversal may be violated in pathological vascular flow or multi-device interaction regimes.

Future work includes application to more diverse sensor modalities, adaptive or self-supervised recalibration in deployed imaging systems, resource-adaptive architectures (e.g., lightweight transformers), joint demosaicing-enhancement flows, and personalizing analytical models to individual anatomical data (Liu et al., 28 Jan 2026, Zeng, 10 Sep 2025, Pascual et al., 2023).

• "RAW-Flow: Advancing RGB-to-RAW Image Reconstruction with Deterministic Latent Flow Matching" (Liu et al., 28 Jan 2026)
• "Physics-Guided Rectified Flow for Low-light RAW Image Enhancement" (Zeng, 10 Sep 2025)
• "WaveFlow: A Compact Flow-based Model for Raw Audio" (Ping et al., 2019)
• "FloWaveNet : A Generative Flow for Raw Audio" (Kim et al., 2018)
• "Analytical Modelling of Raw Data for Flow-Guided In-body Nanoscale Localization" (Pascual et al., 2023)