Fast samplers for diffusion models (generators that turn random noise into an image by repeatedly denoising it) take a handful of large steps along a differential equation. The more accurate higher-order formulas used for these steps tend to blow up when steps are large, which shows up as oversaturated, exploding pixels. Today's fixes look a lot like the limiters used in shock-capturing fluid simulation. Fluid limiters are known to have one serious flaw. They cannot tell a real sharp feature from a numerical wiggle, so they flatten genuine smooth peaks and drop the scheme to first-order accuracy exactly where precision matters. We propose to import the cure that fluid simulation developed, flux correction, which pairs a safe low-order update with a high-order correction that is limited only as much as needed. We combine it with an idea from high-dimensional state estimation: do not limit everything uniformly. Split the correction into a few subspaces and limit each one separately, according to whether that subspace is actually misbehaving.
Concretely, at every sampling step we compute a cheap, stable first-order step and the usual high-order step. We treat their difference as the "anti-diffusive" correction, break it into a small number of fixed frequency bands, and give each band its own limiting coefficient between zero and one. Each coefficient is the largest value that keeps that band's energy within bounds set by the low-order solution. The core bet is that divergence artifacts live in only some directions of the latent space. If so, limiting band by band should remove artifacts the way momentum damping does, while keeping the high-order accuracy and fine detail that uniform damping or pixel clamping throws away. If instead a single per-step scalar coefficient does just as well, the subspace part of the idea is unnecessary, and the experiments are designed to show that clearly.
Why this connection, and what is new
The clipping bottleneck is well documented in hyperbolic conservation laws. Classical limiters judge each stencil by its local variation, so they cut smooth extrema and degrade to first order (Duan; Xu; Mabuza et al.). The method we borrow is the algebraic flux correction scheme of A linearity preserving nodal variation limiting algorithm for continuous Galerkin discretization of ideal MHD equations. It starts from a low-order artificially diffused solution, re-injects limited anti-diffusive fluxes, and is designed so that linear profiles are never clipped. With that design it keeps an experimental order of convergence near 2.0 on smooth Alfvén waves. Our sampler copies this structure: a bounded low-order step, a limited high-order correction, and an explicit rule that the correction passes untouched whenever the trajectory is locally well behaved.
Diffusion samplers currently stabilize themselves with tools that sit exactly where the clipping problem sits. DPM-Solver++ switches to predicting the clean image, which keeps the equation bounded, and adds dynamic thresholding at high guidance scales. Thresholding is an elementwise clamp of predicted pixels, the diffusion analogue of the crude limiter that flattens real peaks. Diffusion Sampling with Momentum for Mitigating Divergence Artifacts traces artifacts to large Jacobian eigenvalues falling outside the stability region. It enlarges that region with Heavy Ball momentum, which falls to first order, or with Generalized Heavy Ball momentum, whose damping parameter β interpolates between adjacent orders. Momentum acts as artificial diffusion applied uniformly to every latent dimension at a global strength. Our proposal differs from both. The limiting strength is computed rather than tuned, it is computed separately per subspace, and it is applied to the correction rather than to the state.
The subspace ingredient comes from state estimation, and two results there constrain our design. A low-rank ensemble Kalman filter for elliptic observations shows that an ad hoc locality-based regularizer destroys real physics in elliptic problems. Their filter becomes unstable for fewer than 40 ensemble members, whereas restricting the update to an informative subspace stays stable at 10. On dimension reduction in Gaussian filters shows that a subspace constraint should be applied conditionally and centered at the predicted mean, not a static global mean, to avoid bias. We take the general lesson: choose the regularized directions by where instability actually lives, and condition on the current low-order prediction. On the diffusion side, Subspace Diffusion Generative Models shows that high-frequency directions turn into pure noise earlier than low-frequency structure, which motivates a frequency-band decomposition. That paper uses projection only to save compute, never to stabilize a solver.
An obvious alternative subspace fails, and we state this as a design constraint. On the Trajectory Regularity of ODE-based Diffusion Sampling shows each sampling path lies in a very low-dimensional subspace. A multistep high-order update, however, is a linear combination of the current state and past model outputs, so it already lies in the span of the trajectory's own history. Projecting onto that span does nothing. The subspace therefore has to come from outside the trajectory. We consider three sources: fixed frequency bands, an offline principal-component basis of clean-image predictions at each noise level (in the style of Solonen et al.), and, as an expensive diagnostic only, the top Jacobian directions.
Core Claim
Let $x^{L}_{n+1}$ be the first-order (DDIM-type, data-prediction) update and $x^{H}_{n+1}$ the high-order multistep update from the same model evaluations. Let $P_1,\dots,P_K$ be orthogonal projectors onto $K$ fixed bands, with $K$ around 4 to 6. The corrected step is
$$x_{n+1} = x^{L}_{n+1} + \sum_{k=1}^{K} \alpha_k \, P_k\big(x^{H}_{n+1} - x^{L}_{n+1}\big),$$
where each $\alpha_k \in [0,1]$ is the largest value for which band $k$ of the implied clean-image prediction stays below an energy envelope. That envelope is the maximum of the band energy of the low-order prediction and of the previous step's prediction, inflated by a factor $(1+\epsilon)$. This is the counterpart of Zalesak-style local bounds, with "neighboring cells" replaced by "neighboring steps and the safe solution."
A preservation rule plays the role of the linearity-preserving property. If an embedded lower-order estimate agrees with the high-order step within band $k$ to tolerance, $\alpha_k$ is set to one. Because $x^{L}$ and $x^{H}$ reuse the same network outputs, the whole procedure adds no network evaluations, only a few transforms per step.
The claim is that this sampler matches or beats tuned momentum on artifact suppression at 10 to 20 steps. It should also keep an empirical convergence order close to that of the unlimited high-order solver, whereas momentum, thresholding, and a single global coefficient each lose order or detail.
Research Program
The first phase is diagnostic and decides whether the idea deserves the method. For ADM on ImageNet-256 at 10 steps, where divergence artifacts are documented, we run unlimited DPM-Solver++. At each step we decompose the high-minus-low correction into frequency bands, offline principal-component bands, and the top few Jacobian singular directions found by power iteration. We measure how unevenly the correction's runaway growth is spread across these decompositions. If growth is spread evenly across every decomposition, per-subspace limiting cannot beat a scalar limiter, and we report that outcome. If growth concentrates, we keep the cheapest decomposition that captures most of it.
The second phase builds the limiter above. It includes the envelope rule, the preservation rule, and a choice of whether the envelope is measured on the state or on the clean-image prediction. Following Solonen et al., we expect the prediction-centered version to be less biased.
The third phase measures the clipping cost directly. We estimate empirical convergence order against a 1000-step reference trajectory from the same noise, and quantify lost detail as the high-frequency energy deficit relative to that reference. To confirm that the mechanism really avoids clipping smooth peaks, we run the same per-band flux-corrected limiter on a 1D advection problem with both smooth extrema and discontinuities. We check whether its convergence order on the smooth part stays near the unlimited scheme's rather than dropping toward first order.
The fourth phase extends to stiff guidance. This means classifier-guided ADM at 20 steps and Stable Diffusion with classifier-free guidance at scales around 8 on MS-COCO. There, splitting methods and thresholding are the competitors.
Baselines
The naive reference is unlimited DPM-Solver++ multistep with the same model, steps, and noise seeds. On ImageNet-256 with ADM at 10 steps it scores FID 66.77, as reported in Diffusion Sampling with Momentum for Mitigating Divergence Artifacts. Any useful limiter must beat it clearly.
The strongest existing competitor for artifact removal without retraining or extra evaluations is Generalized Heavy Ball momentum. The same paper reports FID 37.43 for GHVB with β = 0.8 on ImageNet-256, ADM, 10 steps. Because that paper says the best β depends on model, step count, and guidance, a fair comparison tunes β per setting on a held-out FID set. GHVB gets the same number of tuning trials we give our only free parameter, ε, and both methods use the same seeds and evaluation sample count. Beating GHVB means lower FID than the tuned GHVB at 10 and 15 steps. It also means a measurably higher empirical convergence order, since GHVB interpolates toward a lower order. Matching its FID while keeping order would count as a partial win, worth reporting only if detail metrics favor us.
The thresholding baseline is DPM-Solver++ with dynamic thresholding, run at the percentiles recommended in DPM-Solver++. It matters most in the high-guidance Stable Diffusion setting. We found no reported FID on Lacuna for DPM-Solver++ with thresholding on MS-COCO at guidance 8, so we will measure it ourselves under our protocol.
For guided sampling, the reference is the splitting method STSP4 from Accelerating Guided Diffusion Sampling with Splitting Numerical Methods. It reports FID 4.49 at 20 steps on ImageNet-256, against 4.59 for 250-step guided sampling. We compare at equal network evaluations and equal wall time, because splitting evaluates the guidance twice per step.
The key ablation is a global flux-corrected limiter with $K=1$: the same envelope and preservation rules, but one scalar coefficient per step. It isolates whether the subspace split contributes anything beyond the flux-correction structure. A second ablation removes the preservation rule, to show that this rule is what protects convergence order.
Why not just…?
Why not just clamp or dynamically threshold the latents? Clamping is exactly the elementwise limiter whose failure mode, flattening genuine extrema and losing accuracy, is documented in Duan and Xu. In a sampler it hides blowups after they happen rather than preventing the high-order extrapolation that causes them. Our detail-deficit and convergence-order measurements test this directly. Thresholding should reduce saturation but leave a larger high-frequency energy gap to the reference than flux correction does. Since thresholding acts on the prediction and our limiter on the correction, we also test them combined.
Why not just tune GHVB's β? This is a real threat. A well-tuned β may already capture most of the FID gain at 10 steps. Our advantage has to show up in three places: no per-setting tuning, preserved convergence order, and better behavior when the best β shifts across guidance scales. The guided and classifier-free experiments are designed so that a single fixed ε is compared against β retuned for each scale. If tuned GHVB matches us on all three axes, we will say so.
Why not just use one adaptive scalar per step instead of subspaces? This is the most likely way the proposal partly fails, which is why the $K=1$ ablation is central rather than an afterthought. The first-phase diagnostic predicts the outcome in advance. If runaway growth is spread evenly across bands, the scalar limiter should tie, and the contribution narrows to adaptive flux correction for samplers, which is still distinct from momentum because it computes its strength and keeps a preservation rule.
Why not just use better step schedules or solver search? Geometry-based scheduling cuts five-step DDIM FID on CIFAR-10 from 49.66 to 28.05 (On the Trajectory Regularity of ODE-based Diffusion Sampling). Searched solver schedules reach 6.86 at five steps (A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models). Both change where steps go and which formula is used, not how an unstable correction is bounded, so they are complements rather than substitutes. We will run our limiter on top of a geometry-based schedule to check that the gains add up rather than overlap.
Decisive Checks
The idea stands or falls on three outcomes. The first is the phase-one diagnostic: the high-minus-low correction's growth must concentrate in a few bands of some cheap decomposition. The second is that, on ImageNet-256 with ADM at 10 steps, the band-wise limiter beats both tuned GHVB (37.43 FID reported) and the scalar $K=1$ limiter. The third is that its empirical convergence order against the 1000-step reference stays close to unlimited DPM-Solver++ while GHVB's and thresholding's do not, which also shows up as a smaller high-frequency detail deficit.
A positive answer on all three would show that the clipping bottleneck familiar from fluid limiters also affects diffusion samplers, and that subspace-selective limiting removes it at no extra network cost. A negative answer on the first or second would still give a clean result: the instability in diffusion samplers is not subspace-structured, and global adaptive flux correction is the right tool.
References
- A Discontinuous Petrov-Galerkin Method for One-Dimensional Hyperbolic Conservation Law Equations Based on HWENO Limiter
- Spectral/HP element method with hierarchical reconstruction for solving nonlinear hyperbolic conservation laws
- A linearity preserving nodal variation limiting algorithm for continuous Galerkin discretization of ideal MHD equations
- DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models
- Diffusion Sampling with Momentum for Mitigating Divergence Artifacts
- A low-rank ensemble Kalman filter for elliptic observations
- On dimension reduction in Gaussian filters
- Subspace Diffusion Generative Models
- On the Trajectory Regularity of ODE-based Diffusion Sampling
- Accelerating Guided Diffusion Sampling with Splitting Numerical Methods
- A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models