Engineers who model blood flow and other fluids often assume the flow does not change over time. This "steady-state" assumption turns a nine-hour simulation into a twenty-minute one, but it is wrong exactly where it matters: in pulsing, swirling, reversing flow near narrowings, aneurysms and implanted devices. We propose to test that assumption, region by region, from a single kind of measurement nobody currently treats as evidence. That measurement is a long-exposure photograph of tracer particles in the flow. Each particle leaves a blur streak during the exposure, and the streak records the path the particle took. Its brightness along its length also records how long the particle lingered at each point, and therefore its speed history. In a steady planar flow, the paths of particles cannot cross, because streamlines cannot cross. Two streaks that intersect within one exposure are therefore direct evidence of unsteadiness, unless out-of-plane motion or particle inertia explains them. We will build a differentiable renderer that simulates these streaks from a candidate time-varying velocity field. Fitting it to blurred images, under fluid-physics constraints, recovers the velocity field inside the exposure window and shows where a steady model would be wrong.
The opportunity sits between three literatures that do not talk to each other. Work that recovers motion from blur treats a blurred image as a time integral of the sharp scene, but only for rigid objects or generic video pixels. Physics-constrained fluid reconstruction enforces conservation of mass and momentum, but only on sharp frames. Hemodynamic modeling uses cheap steady solutions and then catalogues their failures. Joining the three gives an instrument that keeps the cheap steady solution as its baseline and uses motion blur, normally discarded as degradation, to decide when that baseline is safe.
Core Claim
A long-exposure image of seeded flow contains enough information to separate a steady velocity field from an unsteady one, and to estimate the unsteady part within the exposure. This requires a forward model that renders the blur physically and a velocity model that conserves mass by construction. Concretely, we write the velocity as a steady mean $\bar{\mathbf{u}}(\mathbf{x})$ plus $K$ temporal modes:
$$\mathbf{u}(\mathbf{x},t)=\bar{\mathbf{u}}(\mathbf{x})+\sum_{k=1}^{K} a_k(t)\,\boldsymbol{\phi}_k(\mathbf{x}).$$
Here each spatial mode $\boldsymbol{\phi}_k$ is the curl of a stream function, so it is divergence-free (mass-conserving) automatically. The coefficients $a_k(t)$ are smooth functions of time over the exposure $[0,\tau]$, and $K=0$ is exactly the steady-state assumption. Tracer particles at positions $\mathbf{p}_i(t)$ are moved by $d\mathbf{p}_i/dt=\mathbf{u}(\mathbf{p}_i,t)$, and the predicted blurred image is
$$\hat B(\mathbf{x})=\frac{1}{\tau}\int_0^{\tau}\sum_i G\big(\mathbf{x}-\mathbf{p}_i(t)\big)\,dt,$$
where $G$ is the particle's imaged point-spread function. We fit by minimizing the gap between $\hat B$ and the observed image $B$, plus a penalty on the residual of the unsteady Navier–Stokes momentum equation.
The new element is the comparison between nested models, $K=0$ against $K>0$, scored on held-out exposures. It yields a map of where the steady assumption is violated and by how much, together with an estimate of acceleration within the exposure, which no steady method can provide.
Relation to Prior Work
Three lines of work motivate the gap. The first is steady hemodynamic modeling. Steady-state geometric multiscale FFR replaces transient coronary simulation with a steady 3D solver coupled to a lumped boundary model. It cuts compute from about nine hours to twenty minutes at 88.3% diagnostic accuracy, but with 68.18% sensitivity. Wang, Dong and Tian build a neural surrogate of steady stenosis flow that is more than 1,000 times faster than direct simulation, and name pulsatile flow as its main limitation. O'Brien et al. and Zettinig et al. show that steady models miss cyclic wall-shear peaks and flow reversal in diseased vessels. None of these papers offers a measurement that tells a practitioner, for a given geometry, whether the steady answer can be trusted. That measurement is what we propose to build.
The second line supplies our method: work on motion from blur. Parametric Object Motion from Blur shows that a non-uniform smear constrains rotation and acceleration. Wang et al. show that the ends of a streak mark the start and end of the exposure. BAD-Gaussians renders blur as an average over a continuous camera trajectory. We reuse this "render the exposure integral" idea, but replace a rigid camera path with thousands of particle paths advected by a fluid field, a setting the file identifies as where these methods break down. MoBGS handles dynamic scenes under blur but uses only photometric loss, with no conservation constraint. The Blur Interpolation Transformer (BiT) and conditional video diffusion unfolding learn to turn blur into sharp frames. Both are data-driven and fluid-agnostic, and the diffusion paper concedes failure on dynamics outside its training priors.
The third line is physics-constrained flow reconstruction, which both supplies and constrains our design. Lasinger et al. regularize multi-camera 3D flow with the stationary Stokes equations. That regularizer drops exactly the time-derivative term we need to recover, and it treats blur as noise. ScalarFlow and PI-NeRF enforce Navier–Stokes on reconstructed smoke, but capture deliberately sharp frames. Consensus-ADMM PIV refinement adds incompressibility to particle image velocimetry (PIV), the standard technique of tracking seeded particles between paired sharp frames, and assumes those frames are sharp. Divergence-free wavelets project MRI flow data onto mass-conserving fields after acquisition. Our stream-function parameterization is the same idea applied inside the forward model rather than as post-processing.
The physical-violation literature frames our evaluation. PhyWorldBench and PISA show that learned video generators routinely break conservation laws. STREAM shows that pixel metrics do not detect such violations. We therefore score velocities and physics residuals, not image sharpness.
Research Program
Forward model. We will build a differentiable streak renderer: GPU particle advection with a fourth-order time integrator, sub-steps inside $[0,\tau]$, and splatted Gaussian particle images. It will include an optional shutter-response curve and a sensor saturation model, because BiT shows real blur departs from frame averaging. Seeding density, particle positions at exposure start, and the velocity coefficients are optimized jointly. This is the expensive part, but it is far cheaper than the 64-hour network training reported for PI-NeRF, because the unknowns are a low-dimensional set of modes rather than a deep network.
Choosing the spatial modes. The modes come from cheap physics, which is how the steady-state direction enters. $\bar{\mathbf{u}}$ is initialized from a steady solver run on the known geometry. The unsteady modes $\boldsymbol{\phi}_k$ are taken from a small library of divergence-free fields that fit the geometry, for example stream-function B-splines restricted to the vessel. The library is therefore a correction to the steady answer, not a replacement for it.
Model selection. We fit $K=0,1,\dots$ and select $K$ by prediction error on held-out exposures. We also compute, per region, the drop in error from $K=0$ to the chosen $K$. That per-region drop is the "steadiness violation map."
Known confounds. Streak crossings can come from out-of-plane motion, so experiments use thin laser-sheet illumination and report sheet thickness. Crossings can also come from heavy particles that lag the fluid, so we use neutrally buoyant tracers with small inertia.
Direction ambiguity. An isolated streak cannot tell which way the particle travelled: reversing time turns acceleration into deceleration and leaves the brightness profile unchanged. We do not claim the Navier–Stokes penalty fixes this. In low-speed, viscosity-dominated flow the equations are nearly time-reversible, and any symmetry breaking at higher Reynolds number is an empirical question we will measure. The primary disambiguation comes from consecutive exposures, since the end of one streak must meet the start of the next across the camera's readout gap, and from the known sign of the mean flow at the inlet.
Baselines
The naive reference is classic streak velocimetry under the steady assumption. Velocity is streak length divided by exposure time, which is our model at $K=0$ without the physics penalty. We pair it with the steady solver's own prediction for the same geometry. Both use identical images and geometry. Our method beats them if it reduces error in instantaneous velocity and acceleration within the exposure in pulsatile regions, while matching them in genuinely steady flow. Matching in steady flow is the check that we have not overfit noise.
The strongest learned competitor is a two-stage pipeline. BiT turns each blurred frame into a dense sequence of sharp frames, and a standard optical-flow method is then run on consecutive unfolded frames. We retrain or fine-tune BiT on our synthetic fluid-blur data, with the same training images and comparable GPU hours. The best reported numbers for BiT are 34.97 dB PSNR on Adobe240 for BiT++ and 30.45 dB PSNR on the real-blur RBI dataset (BiT). The diffusion unfolder reports 30.01 dB PSNR on GoPro (Generating the Past, Present and Future). These are image metrics. We found no reported result on Lacuna for velocity error within an exposure on fluid imagery, so the velocity comparison will be established in this work. We will run the diffusion unfolder zero-shot as a second learned competitor, since fine-tuning a 2B-parameter model is outside a matched budget, and we will state that asymmetry. A win means lower velocity and acceleration error together with lower divergence and momentum residual.
The strongest physics-based competitor is Lasinger et al.'s stationary-Stokes regularized flow. We give it the same images and also, separately, sharp high-speed frames. The consensus-ADMM PIV refinement run on sharp high-speed frames serves as an upper reference rather than a competitor.
Ablations isolate each ingredient: dropping the brightness-along-streak term so that only streak geometry is used, dropping the Navier–Stokes penalty, dropping the divergence-free parameterization, and initializing $\bar{\mathbf{u}}$ from zero instead of the steady solver.
Why not just…?
Why not just use a high-speed camera? In a well-lit lab this is the real threat, and we treat it as one. Our claim is narrower: at a fixed light budget, long exposures gather more photons per frame, so short exposures become noise-limited in dim, confined or fluorescent settings. Cost is a secondary factor; TomoFluid notes high-speed tomographic cameras exceed $80,000 per unit. We will settle this with a matched-photon experiment. We will compare (a) one long exposure analyzed with our method against (b) many short, noisy exposures with the same total light analyzed by PIV, and sweep illumination downward. If (b) wins at every light level, the proposal's practical case collapses to reusing existing long-exposure data, and we will say so.
Why not just fine-tune a deblurring or video-diffusion model on fluid blur? This competitor is included above at a matched data budget. Evidence that it will produce physically inconsistent velocities comes from PISA: even after physics post-training, its trajectory error remains 0.067 L2, and the authors report visual degradation from physics losses. There is also a structural problem. A learned unfolder produces frames, not a velocity field, so its flow still has to be estimated afterwards and inherits optical flow's divergence errors. If the fine-tuned BiT pipeline matches our velocity error and passes the steadiness test, the explicit forward model is not justified. That is a real possible outcome.
Why not just rescale the steady solution with a measured inlet waveform? In a straight tube, the classical analytic solution for pulsatile pipe flow (Womersley flow) makes this nearly exact, and we expect to tie there. It cannot produce flow reversal or moving recirculation zones, which Zettinig et al. identify as the failure point downstream of stenoses. We include this rescaled-steady baseline and break down errors separately for the straight-tube and post-stenosis regions.
Why not just add a divergence-free constraint to streak velocimetry? That is our $K=0$ ablation. Mass conservation alone cannot represent time variation, so any gain from $K>0$ in pulsatile regions measures exactly what this simpler idea misses.
Proposed Experiments
Synthetic tests with full ground truth. We will simulate the 2D and 3D stenosis geometries parameterized as in Wang, Dong and Tian, using a lattice Boltzmann solver driven by pulsatile inflow over a range of pulse frequencies. We seed virtual tracers and render blurred exposures at exposure-to-period ratios from 0.01 to 0.3. To test behavior in buoyant 3D flow, we also advect tracers through ScalarFlow reconstructed velocity fields.
Real-blur pipeline. Following the beam-splitter design behind RBI, we will build a pulsatile stenosis phantom imaged by a long-exposure camera and a co-aligned high-speed camera. The high-speed camera provides ground-truth PIV. We will first validate the renderer's sensor model on RBI itself, by checking that synthetic blur rendered from RBI's sharp frames matches its real blurred frames.
Metrics. We measure endpoint error of velocity and acceleration at times within each exposure, peak wall-shear-stress error near the wall, divergence magnitude, and unsteady momentum residual. Detection quality of the violation map is scored as area under the ROC curve against ground-truth regions where the time variation exceeds a threshold. Image metrics are reported only as secondary measures.
Decisive checks. There are four. First, in synthetic steady flow, model selection must choose $K=0$; if it does not, the method hallucinates unsteadiness and fails. Second, recovery of acceleration within the exposure should degrade gracefully as the exposure-to-period ratio shrinks, because short streaks carry little curvature, and we will report the ratio below which it becomes uninformative. Third, the violation map should flag the post-stenosis recirculation zone where the steady solution's shear error is largest. Fourth, the matched-photon experiment against short-exposure PIV must show a crossover light level below which long-exposure streak analysis wins. If no such crossover exists, the result is negative and we will publish it as such.
References
- Non-Invasive Quantification of Fraction Flow Reserve Based on Steady-State Geometric Multiscale Models
- Fast prediction of blood flow in stenosed arteries using machine learning and immersed boundary-lattice Boltzmann method
- Constraining OCT with Knowledge of Device Design Enables High Accuracy Hemodynamic Assessment of Endovascular Implants
- 3D Velocity Field and Flow Profile Reconstruction from Arbitrarily Sampled Doppler Ultrasound Data
- Parametric Object Motion from Blur
- Video Frame Interpolation without Temporal Priors
- BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting
- MoBGS: Motion Deblurring Dynamic 3D Gaussian Splatting for Blurry Monocular Video
- Blur Interpolation Transformer for Real-World Motion from Blur
- Generating the Past, Present and Future from a Motion-Blurred Image
- Volumetric Flow Estimation for Incompressible Fluids Using the Stationary Stokes Equations
- ScalarFlow: a large-scale volumetric data set of real-world scalar transport flows for computer animation and machine learning
- Physics Informed Neural Fields for Smoke Reconstruction with Sparse Data
- Particle Image Velocimetry Refinement via Consensus ADMM
- Improved visualization and quantification of 4D flow data using divergence-free wavelets
- $PhyWorldBench$: A Comprehensive Evaluation of Physical Realism in Text-to-Video Models
- PISA Experiments: Exploring Physics Post-Training for Video Diffusion Models by Watching Stuff Drop
- STREAM: Spatio-TempoRal Evaluation and Analysis Metric for Video Generative Models