Tile Seams as Splitting Error: A Modified-Equation Account of Tiled Diffusion Sampling

Diffusion models generate images by starting from random noise and denoising it step by step. Mathematically, this is the numerical solution of a differential equation that moves the whole image at once. At 4K, 8K, or gigapixel scale the whole image does not fit in GPU memory, so practitioners denoise overlapping tiles separately and stitch them together. The stitching rule varies. URCDM inpaints each tile against its already finished neighbors (URCDM). Tiled Diffusion overwrites tile margins with neighbor values in round-robin order (Tiled Diffusion). FreeAdapt injects structure from a low-resolution reference image (FreeAdapt). All three are tuned by eye, and none can say why seams appear, how large they will be, or how they change when the number of sampling steps is cut.

We propose to treat a tiled sampler as what numerical analysts call an operator-splitting integrator. Such a method advances a big coupled system by updating its pieces one at a time. We then apply backward error analysis, the tool that finds which slightly different equation a discrete method actually solves. It was used to explain Nesterov acceleration (Understanding the acceleration phenomenon via high-resolution differential equations) and to design few-step diffusion schedules (Fast Diffusion Probabilistic Model Sampling through the lens of Backward Error Analysis). The analysis predicts that a seam has two separable causes. The first is a step-size-independent error from tiles seeing truncated context. The second is an error proportional to step size that comes from the order in which tiles are updated. Each cause calls for a different, cheap fix. The first fix comes from the coarse stage of a hierarchical pipeline. The second comes from symmetric tile ordering.

Core Claim

Let the full-image sampling equation be $\dot{x} = F(x)$, where $x$ is the whole image or latent and $F$ is the denoiser-defined velocity. A tiled sampler replaces $F$ by tile fields $F_i$. Each $F_i$ is the network's prediction on tile $i$'s crop, masked to that tile.

Sequential schemes, which include URCDM-style inpainting and Tiled Diffusion's round-robin constraints, apply these fields one after another within a step of size $h$. Standard splitting theory says such a scheme exactly follows the modified equation

$$ \dot{x} = \sum_i F_i(x) + \frac{h}{2}\sum_{i<j}[F_i, F_j](x) + O(h^2), $$

where $[F_i,F_j] = F_j' F_i - F_i' F_j$ is the commutator, built from the Jacobians $F'$ of the tile fields. The commutator is nonzero only where tiles interact, which is in the overlaps.

The gap between $\sum_i F_i$ and the true $F$ is a separate, step-size-independent truncation error. It is caused by receptive fields cut off at tile borders, the asymmetry SwinIR identifies at window edges (SwinIR).

This gives a testable law: seam magnitude ≈ $a + b\,h$ for one-sided orderings. Here $a$ is the truncation error and $b$ is the splitting error. A symmetric palindromic ordering (forward sweep for half a step, reverse sweep for the other half, as in Strang splitting) should cancel the $h$ term, leaving $a + O(h^2)$.

The commutator also explains the directional artifact build-up the research file attributes to URCDM's frozen borders: a one-sided ordering biases the error in the sweep direction.

Our design follows from this decomposition. We remove $b$ with symmetric tile composition, which costs no extra network evaluations. We shrink $a$ by adding the coarse stage's upsampled velocity as the missing low-frequency, long-range part of $F$. That lets overlaps become narrow, because the remaining high-frequency coupling is short-range.

The analogy to optimization is deliberate but limited. Shi et al. show that the low-resolution limit hides the step-size term that separates two algorithms. Their symplectic discretization work shows the right explicit integrator keeps that term's benefit at no extra cost (Acceleration via Symplectic Discretization). We borrow the method, which is to keep $O(h)$ terms and choose the integrator structure accordingly. We do not borrow their convexity-based guarantees, which do not apply to denoisers.

Research Program

The first component is a controlled testbed where the true answer is known. We use Stable Diffusion-class latent models at resolutions where full-image sampling still fits in memory, roughly 1024–2048 pixels. Here we can run the untiled solver as a reference and measure the tiled sampler's deviation from it directly, at every step and at every pixel distance from a tile border.

We sweep step count (8 to 200 function evaluations), overlap width, tile ordering (one-sided, random, palindromic) and stitching rule (overwrite, average, inpaint). We fit the $a + b h$ law per configuration. Separately, we estimate the commutator numerically with Jacobian-vector products at sampled states, to check that it predicts where seam error concentrates spatially and across noise levels.

Variational Control reports that early, high-noise timesteps set global structure (Variational Control for Guidance in Diffusion Models). We therefore expect the commutator to peak early. If it does, symmetric composition only needs to be applied there.

The second component is the sampler itself. It combines palindromic tile ordering with a coarse-anchored velocity correction. The correction replaces the low-pass band of each tile's velocity with the upsampled velocity of a low-resolution run of the same model. This differs from FreeAdapt, which substitutes the Fourier phase of the image or latent state. We correct the velocity being integrated, so the correction enters the equation as a drift term whose discretization error we can track.

We also derive a step schedule in the spirit of the Restricting Backward Error schedule. That schedule spaces steps to bound a global curvature estimate. Ours spaces steps to bound the estimated commutator norm in the overlaps.

The third component moves the sampler to the regimes where tiling is mandatory. These are UHD restoration and histopathology slide synthesis, where URCDM's cascade gives the hierarchical setting and its inpainting is the incumbent stitching rule.

Baselines

The naive reference is independent tiles with linear alpha blending across fixed overlaps, run with the same base model, step count, and tile size. It anchors the low end and lets us report the seam error the other methods remove.

The strongest boundary-only competitor is Tiled Diffusion, whose latent tiling and similarity constraints with round-robin scheduling reached a Tiling Score of 0.03 against 0.29 for the baseline on 1,000 LAION captions. We will run its released constraints on the same base models and captions, at matched function evaluations and matched overlap width, and report both its Tiling Score and our reference-deviation seam metric.

The strongest hierarchical competitor is FreeAdapt. FreeAdapt reports PSNR gains "typically exceeding 2 dB" over standard patch-based inference and about 26.3% better DISTS on dehazing on its UHD benchmarks. We will run it with the same backbones (StableSR or DiffBIR). Both methods get the same per-image wall-clock budget, and FreeAdapt's guidance strength is tuned with the same number of validation trials we use for our anchor weight.

For gigapixel histopathology, the reference is URCDM's sequential overlap inpainting. It reported a Patch-FID of 39.52 against 150.15 for outpainting on its kidney data. That dataset is proprietary, so we will re-run inpainting ourselves on the TCGA breast and glioma slides. No reported URCDM number exists on those slides, so the comparison is against our own re-implementation at matched compute.

Beating a baseline means one of two things. The first is a lower seam error and equal-or-better full-image FID (or PSNR and DISTS for restoration) at the same function evaluations. The second is matching its seam error with fewer evaluations or narrower overlaps.

Two ablations isolate the ingredients. Palindromic ordering without the coarse anchor should reduce only the slope $b$. The coarse anchor without palindromic ordering should reduce only the intercept $a$. If either ingredient moves the other coefficient, the decomposition is wrong.

Why not just…?

Why not just take more sampling steps? The theory itself says this works for the splitting term, since $b h \to 0$. That makes it the most honest competitor rather than a straw man. It does nothing for the truncation term $a$, and it multiplies cost in exactly the regime where URCDM already needs about 24 hours on an 8–9 A100 cluster per slide. Palindromic ordering removes the $O(h)$ term at zero extra evaluations, so the claim to test is a better seam-versus-compute curve, not a better limit. The step sweep in the testbed settles this directly. If the measured slope $b$ is negligible next to $a$, the splitting half of the proposal is moot, and we will say so.

Why not just widen overlaps or average them? Wider overlaps attack the truncation term by giving border pixels more context, but the cost grows with overlap area. Averaging also does not remove the ordering bias of sequential schemes. Tiled Diffusion already shows its context-window width trades seams against local variety. Our prediction is that the coarse anchor lets overlaps shrink without raising $a$. The overlap-width sweep with and without the anchor tests this.

Why not just randomly shift the tile grid each step? This is a real threat. Under our analysis, random offsets turn the coherent $O(h)$ commutator bias into spatially incoherent noise that later steps may wash out. That could match palindromic ordering on visible seams. We include random shifting as a baseline. We will distinguish the two by measuring deviation from the untiled reference, not just seam visibility, since random shifting should leave higher-variance residual error. If random shifting matches our method on both measures, it is the simpler recommendation, and our contribution reduces to explaining why it works.

Why not just use FreeAdapt-style low-resolution guidance? FreeAdapt already supplies global structure from a reference, which overlaps with our coarse anchor. Our claim is that this guidance addresses only $a$. The palindromic-versus-one-sided comparison layered on top of FreeAdapt shows whether the step-proportional term remains once guidance is present, especially at 8–20 steps.

Proposed Experiments

The decisive first experiment is the scaling fit in the controlled testbed. We plot the reference-deviation seam error against $h$ for each ordering. The theory predicts linear behavior for one-sided orderings, roughly quadratic behavior for palindromic ordering, and an intercept that drops only when the coarse anchor is added. A second check asks whether the measured commutator norm, mapped over space and noise level, correlates with where seam error appears. If it does not, the modified equation is not the operative mechanism, and the method loses its justification even if it helps empirically.

We then evaluate at scale on three settings. The first is LAION captions, using the Tiling Score and full-image FID. The second is UHD-LL, UHD-Haze, UHD-Blur, and 4K-Rain13k, using PSNR and DISTS against FreeAdapt. The third is TCGA-BRCA and glioma slide synthesis, using Patch-FID and wall-clock time against re-run URCDM inpainting.

The palindromic and anchor-only ablations are reported in every setting. Few-step schedules (8–20 evaluations) are the primary regime, since the predicted gains are largest there. Throughout, we report seams at 8, 20, and 50 evaluations so that the "just take more steps" alternative is visible on the same axes.