Projected Diffusion Models (PDMs)

Updated 6 April 2026

Projected Diffusion Models (PDMs) are generative frameworks that combine standard diffusion updates with explicit projection steps to enforce physical, geometric, and feasibility constraints.
They employ projection operators, including augmented Lagrangian methods, to convert unconstrained updates into solutions that meet strict convex and nonconvex requirements.
Applications in robotics, scientific synthesis, and inverse problems demonstrate PDMs’ ability to achieve near-zero constraint violations and improve metrics such as PSNR, SSIM, and FID.

Projected Diffusion Models (PDMs) are generative frameworks that interleave standard diffusion model denoising steps with explicit projections onto constraint sets, enabling direct enforcement of hard constraints—including physical, geometric, and feasibility requirements—throughout the sampling process. This approach addresses a critical limitation of classical diffusion samplers, which cannot guarantee constraint satisfaction in high-dimensional or safety-critical applications. Recent literature establishes several families of PDMs, including those for multi-agent motion planning, scientific synthesis, and inverse problems, employing a range of projection operators and optimization strategies to certify feasibility at every generation step (Liang et al., 2024, Liang et al., 5 Feb 2025, Christopher et al., 2024, Zirvi et al., 2024, Zheng et al., 27 May 2025, Yu et al., 2023).

1. Fundamental Principles and Motivation

PDMs originate from the observation that standard score-based diffusion models, trained to model complex data distributions via forward and reverse stochastic differential equations (SDEs), are inherently unconstrained in their sampling behavior. This unconstrained nature leads to violation of domain-specific requirements, such as collision avoidance in robotics, adherence to measurement models in inverse problems, or preservation of physical laws in scientific simulation (Liang et al., 2024, Christopher et al., 2024). PDMs remedy this by formulating each reverse diffusion iteration as a composite step: (1) an unconstrained gradient-based denoising update, and (2) an immediate projection of the tentative sample onto a prescribed feasible set $\Omega$ , defined via problem-specific constraint equations. The canonical update is:

$x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$

where $s_\theta$ is the learned score, $\gamma_t$ the step size, $z$ Gaussian noise, and $\Pi_\Omega$ denotes Euclidean projection onto $\Omega$ (Christopher et al., 2024, Liang et al., 2024). This process ensures all iterates remain feasible with respect to task constraints.

2. Mathematical Formulation and Projection Operators

The constraint set $\Omega$ is problem-dependent and may encode convex and nonconvex requirements:

Convex constraints: Linear start/goal anchoring, box or ball constraints, velocity bounds, or convex polytopes are directly handled via closed-form or efficient convex solvers.
Nonconvex constraints: Collision avoidance, distance-from-obstacle, nonconvex equalities—these require approximate projection methods such as augmented Lagrangian methods (ALM), slack variable introduction, or local nonconvex solvers (Liang et al., 2024, Liang et al., 5 Feb 2025, Christopher et al., 2024).

For instance, in multi-agent path finding, $\Omega$ enforces joint trajectory feasibility via: start/goal constraints, velocity bounds (both convex), inter-agent and agent-obstacle avoidance (both nonconvex quadratic). The ALM-based projection rewrites inequalities as penalized equalities using slack variables, optimizing the augmented Lagrangian:

$\mathcal{L}(\Pi, \nu_a, \nu_o) = \|\Pi - \bar{x}\|^2 + \nu_a^\top H_a(\Pi) + \nu_o^\top H_o(\Pi) + \rho_a \|H_a(\Pi)\|^2 + \rho_o \|H_o(\Pi)\|^2$

where $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 0 collect constraint violations and $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 1 penalize infeasibility (Liang et al., 2024, Liang et al., 5 Feb 2025). Dual ascent on $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 2 is employed to optimize feasibility during each projection step.

In the context of inverse problems, projection may be onto a data manifold or its low-rank tangent approximation (see Section 4). For latent-space PDMs, projecting onto the range of a decoder or latent submanifold replaces direct pixel-domain projection (Yu et al., 2023).

3. Sampling Algorithms for Projected Diffusion

All PDM sampling procedures follow a structured loop:

Initialization: Draw $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 3.
Reverse diffusion with projection: For each $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 4, perform $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 5 SGLD-like updates,
- Compute unconstrained update,
- Immediately project onto $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 6,
- Set $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 7.
Return: $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 8 as the final feasible sample.

Pseudocode for the generic PDM algorithm (as expressed in (Liang et al., 2024, Liang et al., 5 Feb 2025, Christopher et al., 2024)):

$s_\theta$ 9

In practice, projection subroutines employ convex solvers (for $x^{(i+1)} \;=\; \Pi_\Omega\left(x^{(i)} + \gamma_t s_\theta(x^{(i)}, t)\ +\ \sqrt{2\gamma_t}z \right)$ 9) and ALM for nonconvex parts. For efficiency and convergence, step sizes and penalty parameters are tuned as per application-specific schedules.

4. Manifold-Constrained and Subspace-Projected Variants

Beyond hard geometric constraints, PDMs have been extended to enforce broader forms of data consistency—in particular, ensuring samples remain close to the learned data manifold at every denoising step. Diffusion State-Guided Projected Gradient (DiffStateGrad) (Zirvi et al., 2024) replaces naively guided updates in inverse problems with projections of measurement-consistency gradients onto a state-specific low-rank subspace $s_\theta$ 0, estimated via SVD decomposition of the current (noisy) sample $s_\theta$ 1:

$s_\theta$ 2

with $s_\theta$ 3, $s_\theta$ 4 corresponding to the dominant singular vectors. Projecting measurement gradients in this way guarantees that reverse diffusion steps remain in the vicinity of the diffusion prior manifold, reducing artifacts and improving robustness to noise and hyperparameter tuning.

Further, projected latent diffusion models (PVDM) (Yu et al., 2023) structure diffusion directly in a learned low-dimensional latent space, with projections implicitly realized by the autoencoder architecture mapping between observed and latent domains. For inverse problems, DMILO and DMILO-PGD (Zheng et al., 27 May 2025) execute projection onto the extended model range plus sparse deviations, alternating measurement consistency gradient-descent steps with ILO-based projection, yielding improved convergence and memory efficiency over prior plug-and-play solvers.

5. Theoretical Guarantees

For convex feasible sets $s_\theta$ 5 and standard diffusion SDE conditions, projected SGLD chains are ergodic within $s_\theta$ 6 and converge to the correct constrained marginal distribution (Christopher et al., 2024, Liang et al., 2024). Dual ascent with augmented Lagrangian projection is guaranteed to converge to Karush-Kuhn-Tucker (KKT) points, provided convex subproblems attain their minima and penalty updates are appropriately scheduled (Liang et al., 2024, Liang et al., 5 Feb 2025). For nonconvex $s_\theta$ 7, global guarantees are absent, but empirical results demonstrate robust feasibility when projections are accurate.

Manifold-projected variants admit formal guarantees of reduced deviation from the data manifold: projected updates yield iterates that remain strictly closer to the manifold compared to unprojected steps, under mild smoothness and neighborhood assumptions (Zirvi et al., 2024). For DMILO-PGD, convergence bounds show that projected updates remain competitive with (potentially uncomputable) optimal projections onto the true model range (Zheng et al., 27 May 2025).

6. Applications and Empirical Evaluation

Robotics and Motion Planning

PDMs have been applied to Multi-Agent Path Finding (MAPF) and Multi-Robot Motion Planning (MRMP) in continuous and cluttered environments (Liang et al., 2024, Liang et al., 5 Feb 2025). Here, explicit projection ensures collision avoidance, velocity constraints, and boundary anchoring for all agents throughout the generated trajectory. Experimental benchmarks show that PDMs consistently achieve near-zero constraint violations—outperforming both standard diffusion models and soft-penalty-guided diffusion baselines in terms of feasibility and trajectory efficiency. For example, PDM achieves 0% violation rate in narrow-corridor and obstacle-dense scenarios, while standard diffusion produces 15-35% and up to 1% violations, respectively (Liang et al., 2024).

Scientific and Engineering Synthesis

General-purpose PDMs are validated for constrained synthesis of microstructures, physics-informed video, and motion planning, achieving exact feasibility and high generative fidelity (e.g., low FID and path-length difference metrics) as compared to post-hoc or conditional diffusion alternatives (Christopher et al., 2024). For micrograph synthesis under prescribed porosity, PDM yields <1% constraint error, outperforming conditional models (up to 30% violation).

Inverse Problems

PDM-based solvers such as DiffStateGrad (Zirvi et al., 2024) and DMILO-PGD (Zheng et al., 27 May 2025) demonstrate superior performance in diverse inverse problems, including image restoration, deblurring, inpainting, and super-resolution. For FFHQ random inpainting, DiffStateGrad enhances LPIPS (0.165 vs. 0.246), SSIM (0.898 vs. 0.809), and PSNR (31.68 dB vs. 29.05 dB) relative to unprojected solvers. DMILO-PGD yields up to +1.1 dB PSNR and 0.05 LPIPS improvement over DMPlug, with major gains also in FID and convergence stability.

Video Synthesis

Latent PDMs, such as PVDM (Yu et al., 2023), project high-dimensional video into factorized 2D-shaped latent planes, permitting efficient diffusion and reconstruction of high-resolution, temporally coherent videos. On UCF-101, PVDM achieves an FVD $s_\theta$ 8 of 639.7, outperforming previous methods by large margins, with lower memory and compute requirements.

7. Limitations and Open Directions

Current PDMs rely on feasible and computationally tractable projections; projection onto highly nonconvex or disconnected sets remains challenging and may only achieve local optimality. Hard constraint enforcement can trade off sample diversity or perceptual quality—e.g., in physics-informed video, FID may increase slightly compared to unconstrained models (Christopher et al., 2024). In latent PDMs, projection is tied to autoencoder architecture, so manifold misspecification remains a risk (Yu et al., 2023).

Ongoing research targets algorithmic optimizations (bespoke solvers for projection, more expressive latent manifolds), theory for nonconvex and disconnected constraint sets, and scaling to large vision, language, and multi-agent benchmarks. Extensions to handle stochastic constraints, adaptive metric projections, or further integration with feedback-guided diffusion models remain active investigations.

Key references: