Tilt Matching: ET & Generative Modeling

• Tilt Matching is a dual-domain technique that aligns electron tomography projections and adapts generative model flows using tilt corrections.
• It employs deterministic corrections using fiducial markers and mathematical formulations like the Radon transform to correct translation and vertical-tilt errors.
• In generative modeling, TM guides velocity field evolution via reward-tilted sampling, achieving reduced variance without requiring reward gradients or trajectory backpropagation.

Tilt Matching (TM) refers to two technically distinct but conceptually analogous methodologies originating in electron tomography (ET) and generative modeling. In both domains, TM addresses the alignment or adaptation of a process to a target condition: in ET, achieving ideal geometric arrangement of tilted projections for accurate tomographic reconstruction; in generative modeling, adapting continuous-time flows or diffusions to target densities tilted by a scalar reward function without requiring reward gradients or trajectory backpropagation. TM protocols are characterized by either deterministic correction—removing systematic geometric tilt errors—or adaptive drift modification—solving for optimal velocity evolution under reward tilting.

1. TM in Electron Tomography: Alignment Protocols and Sinogram Correction

The electron tomography formulation of TM, as presented by Kim and Jun, establishes a pipeline to transform a stack of raw projections at nominal tilt angles $\theta_j$, $j=1\dots N$, into an ideally aligned set free of translation and vertical-tilt errors (Kim et al., 2017). TM leverages two or more fiducial markers (FPs) tracked throughout the projections. The alignment pipeline is as follows:

• Translation correction: Each projection $I_j(x, y)$ is horizontally shifted $\Delta x_j$ so FP #1 lies on the virtual rotation axis $x_0$, collapsing its sinogram trajectory to $s = x_0$.
• Vertical-tilt error correction: For FP #2, a per-projection rotation $\phi_j = \arctan[(y_{2j} - y^*)/(x_{2j} - x_0)]$ about $(x_0, y^*)$ aligns FP #2 to a common image row $y = y^*$ across all $\theta_j$.
• Parallel-tilt error diagnosis: Any residual deviation in FP #2’s sinogram trajectory, post vertical-tilt correction, indicates axis misalignment. The optimal fit may be elliptical, rather than purely sinusoidal; reconstruction remains possible but is not ideally focused layer-wise.

This TM process enforces the theoretical ideal for each fixed $z$-layer: its sinogram trajectory must follow

$T_{r,q}(\theta) = r\,\cos(\theta-\theta_0)$

for each point at radius $r$ with offset $\theta_0$. The practical algorithm is iterative, involving preprocessing, fiducial detection, translation and tilt corrections, validation via RMS sinogram residuals, tilt-range assessment, and final reconstruction via filtered-backprojection or iterative algorithms.

2. Key Mathematical Formulation and Error Correction Criteria

Central to TM in ET are explicit formulas for both alignment and tilt error quantification:

• Projection-to-sinogram: Projection data $f(x, y)$ at angle $\theta$ is related to sinogram via the Radon transform:

$$g(s, \theta) = \int\int f(x, y)\,\delta(s - x\cos\theta - y\sin\theta)\,dx\,dy$$

• Tilt correction from measured deviations: For observed phase shifts $\Delta s(\theta)$ in FP sinograms, the local tilt error is estimated by

$$\delta\theta(\theta) = \frac{\Delta s(\theta)}{r\,\sin(\theta-\theta_0)}$$

and applied as a corrective re-indexing $$\theta_j \leftarrow \theta_j - \delta\theta(\theta_j)$$.

• Tilt-angle coverage criterion: Ideal sinogram coverage requires continuous data over $180^\circ$ of view, especially around directions of maximal internal density gradient. Mechanical restrictions often limit ET to $\pm(60^\circ–70^\circ)$; missing these critical angles produces irreparable artifacts (“missing wedge”).

3. TM in Generative Models: Reward-Tilted Sampling and Velocity Field Evolution

In generative modeling, TM abstracts an ODE-level scheme for transporting a reference distribution $\rho_0$ to not just a target $\rho_1$ but one exponentially tilted by a scalar reward $R(x)$: $\rho_{1,R}(x) \propto \rho_1(x)\,e^{R(x)}$ (Potaptchik et al., 26 Dec 2025). The goal is to solve for modified drift $v_{t,a}(x)$ over a tilting parameter $a \in [0,1]$ such that the system

$\dot{X}_t = v_{t,a}(X_t), \quad X_0 \sim \rho_0$

transports $\rho_0$ to $\rho_{1,a}(x) \propto \rho_1(x)\,e^{aR(x)}$.

A stochastic interpolant $I_t^a = \alpha_t x_0 + \beta_t x_1^a$ (with $(x_0, x_1^a) \sim \rho_0 \otimes \rho_{1,a}$) induces a law $\mathrm{Law}(I_t^a) = \rho_{t,a}$, and drift evolution is given by an Esscher transform:

$$v_{t,a}(x) = \frac{ \mathbb{E}\left[ \dot{I}_t^0\,e^{aR(x_1)}\,|\,I_t^0 = x \right] }{ \mathbb{E}\left[ e^{aR(x_1)}\,|\,I_t^0 = x \right] }$$

The fundamental Covariance ODE dictates the infinitesimal update:

$$\partial_a v_{t,a}(x) = \mathrm{Cov}\left( \dot{I}_t^a,\,R(x_1^a) \mid I_t^a = x \right)$$

with initial condition $v_{t,0}(x) = v(t,x)$.

4. Implicit and Explicit Tilt Matching Algorithms

TM offers scalable regression objectives for estimating tilt-updated velocity fields, either to first-order accuracy (Explicit Tilt Matching, ETM) or by enforcing implicit all-orders cumulant expansion (Implicit Tilt Matching, ITM). Both operate over minibatches sampled from the appropriate base and reward-tilted endpoint couplings.

• ETM: Minimizes loss

$$L_{a\to a+h}^{ETM}(\theta) = \int_0^1 \mathbb{E}\left[ \| v_\theta(t, I_t^a) - T_{t,a,h} \|^2 \right] dt$$

for target $$T_{t,a,h} = v_{t,a}(I_t^a) + h[ \dot{I}_t^a R(x_1^a) - v_{t,a}(I_t^a) R(x_1^a) ]$$.

• ITM: Enforces fixed points via

$$L_{a\to a+h}^{ITM}(\theta) = \int_0^1 \mathbb{E}\left[ \| v_\theta(t, I_t^a) - T_{t,a,h}(v_{t,a}, v_\theta) \|^2 \right] dt$$

where $$T_{t,a,h}(v_{t,a}, v_\theta) = v_{t,a}(I_t^a) + (e^{hR(x_1^a)} - 1)[ \dot{I}_t^a - \mathrm{stopgrad}(v_\theta(I_t^a)) ]$$.

• Weighted Flow Matching (WFM): Specializes to $c=0$ control variate regime with higher gradient variance:

$$L_{a\to a+h}^{WFM}(\theta) = \int_0^1 \mathbb{E}[ e^{hR(x_1^a)} \| v_\theta(t, I_t^a) - \dot{I}_t^a \|^2 ] dt$$

Variance reduction is analytically confirmed: $$\mathrm{Var}(\nabla L^\mathrm{WFM}) \geq \mathrm{Var}(\nabla L^\mathrm{ITM})$$ for small $h$.

5. Connections to Stochastic Optimal Control and Mathematical Foundations

The drift update of TM relates directly to Doob’s $h$-transform and controlled stochastic differential equations (SDEs). For a value function

$$\psi_{t,a}(x) = \log\,\mathbb{E}[e^{aR(x_1)} \mid I_t^0 = x]$$

the optimal ODE drift for the controlled SDE is

$$v(t, x) + \frac{\sigma_t^2}{2}\,\nabla \psi_{t,a}(x)$$

where $\sigma_t$ matches endpoint laws. This coincides with $v_{t,a}(x)$ from the Covariance ODE, enabling reward-tilted transport without solving Hamilton-Jacobi-Bellman equations or backward SDEs. TM leverages regression with scalar rewards rather than gradients or trajectory-based derivatives.

6. Practical Benchmarks and Application Domains

TM in ET has become standard for high-fidelity tomogram reconstruction in presence of mechanical misalignment, notably when fiducial tracking is possible across a full or partial tilt range (Kim et al., 2017). In generative modeling, empirical studies verify efficiency and scalability:

• Lennard-Jones sampling (13 and 55 atoms): ITM yields effective sample size (ESS) of 0.507 for LJ-13 (prior best ≈0.23), 1D energy Wasserstein-2 distance (W₂) of 0.879 (prior >2.4), and geometric W₂ of 1.54 (prior >1.59). On LJ-55, TM matches or improves upon state-of-the-art samplers in energy and geometry distances (Potaptchik et al., 26 Dec 2025).
• Fine-tuning Stable Diffusion 1.5: Without reward scaling ($\lambda=1$), ITM achieves ImageReward=0.446 (base=0.187; adjoint matching=0.217), with competitive CLIPScore and HPSv2 metrics and superior text-image alignment. TM is competitive with adjoint methods even at higher reward scaling.

This suggests TM enables direct high-dimensional sampling and fine-tuning under explicit reward criteria, maintaining regularity and tractability across large model classes.

7. Implications, Limitations, and Diagnostic Observations

A plausible implication is that TM’s systematic correction (ET) and covariance-driven drift updates (generative models) provide rigorous deterministic and statistical guarantees of target approximation. In ET, the inability to achieve proper tilt-range or correct axis yields irreducible artifacts. In generative modeling, TM algorithms do not require reward gradients or trajectory backpropagation, producing lower-variance estimates and no discretization bias in tilt parameter $a$. However, mechanical limits, sample thickening, and insufficient tilt coverage remain practical constraints in tomography; reward function regularity and coupling tractability bound generative TM scalability.

Empirical troubleshooting indicates the necessity of distant fiducials for stable alignment (ET), subdivision of the tilt parameter for algorithmic stability (ITM), and validation over multiple layers/slices for assurance of global correction. Algorithmic pseudocode for ITM establishes an iterative update over tilt increments, sampling endpoint pairs and applying cumulant-centered regression loss to learn $v_{t,1}(x)$ for the fully tilted target.

TM methodologies span geometric correction and probabilistic flow modification, unified by their aim to regularize the projection or generative process with respect to a target transformation—be it translational/rotational (ET) or reward-based density tilting (generative models).