Seams as Splitting Error: Implicit Interface Coupling for Tiled Diffusion Sampling

Diffusion models generate an image by starting from random noise and denoising it over many steps. When an image or video is too large to fit in GPU memory, practitioners cut it into tiles, denoise each tile separately, and stitch the results back together. The stitching leaves visible seams. The current fixes are hand-built heuristics. One averages the predicted noise wherever tiles overlap, at every step. Another blends internal features of neighbouring tiles. Both require heavy overlap, which means many extra network evaluations, and both force every tile to stay in lockstep with its neighbours at every step. This note proposes treating the seam as what a numerical analyst would call a splitting error. Tiled sampling is a system of coupled subdomains, so the coupling between tiles can be handled the way partitioned PDE solvers handle it: the expensive local physics is stepped explicitly, and a cheap, stiff interface-agreement term is stepped implicitly through a closed-form linear solve.

The practical claim is simple to state. If the term that forces neighbouring tiles to agree is solved implicitly rather than applied as an explicit average, it can be made arbitrarily strong without destabilising the sampler. Thin overlaps should then suffice, and tiles can run asynchronously, exchanging boundary information only every few steps. Classical stability analysis predicts when that exchange interval becomes too long. Hierarchical narrative planning enters as the application regime and as the source of a slowly varying global term. In long story videos, temporal chunks are tiles too. A scene-level plan, in the style of outline-and-summary story generators, plays the role of the slow component that changes only at chunk boundaries.

Core Claim

Write the tiled reverse-diffusion state as the stacked set of tile copies $u$, where pixels in an overlap or halo band appear once per tile that contains them. One reverse step is split into two parts. The first is $f(u)$, the per-tile denoiser update, which is expensive, local, and non-stiff relative to the step. The second is $g(u) = -\lambda(t) L u$, where $L$ is the graph Laplacian connecting duplicate copies of the same pixel across tiles and $\lambda(t)$ is a coupling strength that may depend on the noise level $t$. The proposed step is the canonical first-order IMEX form:

$$ (I + \Delta t\,\lambda(t)\,L)\,u^{n+1} = u^{n} + \Delta t\, f(u^{n}), $$

Here $\Delta t$ is the sampler step. Because $L$ only touches the halo bands and decouples into tiny per-pixel systems over the duplicate copies, this solve costs no network evaluations and has a closed form. It is a weighted average whose weights depend on $\lambda\Delta t$.

Two consequences follow, and both are testable. First, the noise-averaging rule of Restoring Vision in Adverse Weather Conditions with Patch-Based Denoising Diffusion Models is recovered as the limit $\lambda\to\infty$ over the full overlap. The new freedom is to apply strong agreement on a thin band and schedule its strength across noise levels. Second, consider the explicit alternative, in which disagreement is damped by subtracting $\Delta t\lambda L u^n$ and boundaries are exchanged only every $k$ steps. Interface mismatch then evolves roughly like forward Euler on a linear decay, with per-exchange amplification $|1 - k\Delta t\,\lambda\,\mu|$ for each Laplacian eigenvalue $\mu$. It diverges once $k\Delta t\,\lambda\,\mu_{\max} > 2$. The implicit form has amplification $1/(1 + k\Delta t\,\lambda\,\mu)$, which is below one for every $k$. This is the diffusion-sampling analogue of the step-size restriction $\Delta t < C$ that Analysis of Long Time Stability and Errors of Two Partitioned Methods for Uncoupling Evolutionary Groundwater-Surface Water Flows proves for explicitly coupled Stokes–Darcy interfaces. It predicts a sharp threshold in exchange interval for explicit coupling and no such threshold for implicit coupling.

Research Program

The method borrows its structure from partitioned scientific computing. Layton, Tran and Trenchea step each subdomain implicitly and the interface explicitly (BEFE and BELF) and show long-time stability under a parameter-dependent step bound. We invert the assignment because the costs are inverted. Here the subdomain physics is a neural network that cannot be solved implicitly, while the interface term is linear and cheap. An Ultrascalable Solution to Large-scale Neural Tissue Simulation supplies the second ingredient: volumes integrated locally, with boundary "proxies" synchronised at a tunable interval. We adopt that tunable interval as the exchange period $k$. Exact linear solves for a structured stiff term, rather than iteration, are what make the plasma scheme of Entropy Stable Numerical Schemes for Two-Fluid Plasma Equations fast, and the same logic applies to our halo solve.

The build has three parts, all training-free on top of existing patch diffusion samplers. The first is the coupled step above, with a halo of width $w$ pixels and a schedule $\lambda(t)$. The intended default is strong coupling at high noise, where blurring is harmless, and weaker coupling near $t=0$, where fine texture forms. We will also apply the step to internal features at the blending point Patch-DM uses, to test feature-space versus noise-space coupling.

The second part is a multi-rate asynchronous mode. Tiles on different devices run $k$ local steps between halo exchanges, and the implicit solve is applied at each exchange. This directly targets the inference-latency problem that both WeatherDiff and Hierarchical Patch Diffusion Models for High-Resolution Video Generation report.

The third part is a slow global term. For still images, this term is a coarse low-resolution prediction refreshed every $m$ steps, in the spirit of HPDM's Deep Context Fusion but injected as an explicit slow forcing rather than a learned feature pathway. For story video, the tiles are temporal chunks. The slow term is a per-segment plan, consisting of the segment prompt plus an entity record of characters, clothing, and setting maintained as in Re3: Generating Longer Stories With Recursive Reprompting and Revision. The plan is updated only at segment boundaries, and the implicit halo coupling acts on the overlapping frames between chunks.

On the analysis side, we will measure the empirical amplification of seam mismatch as a function of $k\Delta t\lambda$ and compare it with the linear prediction. A good match would be the first quantitative stability statement for tiled diffusion sampling.

Baselines

The naive reference point is independent tiling with no overlap and no coupling: tiles are denoised separately and concatenated. It sets the seam floor and the cost floor, since every pixel gets one network evaluation per step.

The strongest existing training-free competitor is overlapped noise averaging from Restoring Vision in Adverse Weather Conditions with Patch-Based Denoising Diffusion Models, which averages the predicted noise over overlaps at every reverse step. It is the method our scheme reduces to in a limit, so it isolates exactly what the implicit, thin-band, scheduled coupling adds. We will run it with the same pretrained denoiser, the same number of reverse steps, and overlap ratios swept from 0% to 75%. Results will be reported against total network function evaluations (NFE) and wall-clock time, so that its higher overlap is charged honestly. The paper reports about 20 seconds for a 640×432 image but quotes no scalar quality score in the file. No reported result was found on Lacuna for its Snow100K metric, so we will rerun it. Beating it means matching its seam score and FID at a quarter of its NFE, or improving seam score at equal NFE.

The strongest trained method is Patch-DM from Patched Denoising Diffusion Models For High-Resolution Image Synthesis, which blends internal feature maps of neighbouring patches. It reports FID 20.369 on Nature-21K versus 44.173 for InfinityGAN. Because our method is training-free, the fair test is to apply our sampler to Patch-DM's own trained denoiser and compare it with Patch-DM's native sampling at equal NFE. Beating Patch-DM means equal or lower FID than 20.369 at lower NFE, or lower seam score at equal FID.

For video, HPDM reports FVD 66.32 on UCF-101 at 256². It is the reference for coarse-to-fine coupling. Our slow-term variant is compared with its Deep Context Fusion at equal patch budget. For story video, PHENAKI-CONT on StoryBench reports FVD 350 on Oops-CSV action execution, from StoryBench: A Multifaceted Benchmark for Continuous Story Visualization. It is the reference for temporal continuation across narrative segments.

Three ablations isolate the key ingredient. The first replaces the implicit solve with the explicit damping step at the same $\lambda$, which tests whether implicitness itself matters. The second fixes $\lambda$ constant to test the noise-level schedule. The third removes the slow global or plan term to test whether residual failures are local or global.

Why not just…?

Why not just average noise over larger overlaps? This works, and it is our main baseline. The cost is the problem: at 50% overlap in each spatial axis, every pixel is evaluated roughly four times per step. The averaging also has to happen synchronously at every step. If WeatherDiff-style averaging at equal NFE matches the implicit thin-band scheme on seam score and FID, the synchronous variant of this proposal adds nothing. The overlap-versus-NFE sweep settles that directly.

Why not just shift the tile grid randomly at each step? This is the most serious threat. It costs no extra evaluations and spreads seams so that none persists. No reported result was found on Lacuna for random tile shifting, so it will be run as a baseline in every image experiment. Our case against it rests on two points. Shifting requires globally reassigning pixels to tiles at every step, which rules out asynchronous multi-device sampling. It also does nothing about entity-level errors, such as an object bisected and reassembled inconsistently. If shifting matches us on seams and FID, we will narrow the claim to the asynchronous regime ($k>1$), where shifting cannot operate.

Why not avoid tiling with a bigger GPU or a latent-space model? HPDM argues that full-frame video models exhaust memory as resolution and duration grow, and that latent autoencoders lose fine detail. Its patch training sees as little as 0.7% of pixels per step. Whole-slide pathology images, the setting of NASDM: Nuclei-Aware Semantic Histopathology Image Generation Using Diffusion Models, are far beyond a single canvas. Tiling is a fixed constraint in these regimes, not a crutch, so this proposal targets the regime where tiling is unavoidable.

Why not just use a better global ODE solver with fewer steps? Higher-order samplers reduce the number of steps but treat every tile identically, and they still require global synchronisation at each step. They are orthogonal to interface coupling, so we will combine our halo solve with a second-order sampler to show the gains stack. If they do not stack, that is reported as a limitation.

Why use a narrative plan instead of a long context window? Re3 found that conditioning on distilled outlines and entity records beats feeding raw history, with 14% higher plot coherence. Here the plan is cheap to hold fixed within a segment, which is exactly what a slow explicit term requires.

Proposed Experiments

The image study uses Nature-21K and LHQ for texture-dominated scenes and FFHQ for global symmetry. FFHQ is where Patch-DM qualitatively fails with mismatched eyes, and no reported number was found on Lacuna for Patch-DM on FFHQ. Snow100K provides a restoration task. Metrics are FID on full images, FID on crops centred on tile borders, and a seam ratio: mean gradient energy on border pixels divided by that on interior pixels. Cost is measured as NFE, wall-clock time, and inter-device bytes exchanged.

The video study uses UCF-101 at 256² with spatial tiling, reporting FVD. The story study uses StoryBench story continuation on Oops-CSV and DiDeMo-CSV, generated at higher resolution than the benchmark's 160×96 so that spatial and temporal tiling are both active. StoryBench notes that FVD misses cross-scene inconsistency. We will therefore add a human study of entity consistency across chunk boundaries, judging whether clothing, identity, and setting stay the same, plus FVD restricted to windows straddling chunk seams.

Three checks decide the proposal. First, on Nature-21K the implicit thin-band sampler should reach the seam ratio of 50%-overlap averaging at no more than half its NFE. Second, sweeping the exchange interval $k$ should reveal a divergence threshold for explicit coupling near $k\Delta t\lambda\mu_{\max}\approx 2$, with no threshold for the implicit version. A missing threshold for explicit coupling would show the linear analogy is wrong. Third, on StoryBench, the plan term combined with implicit temporal halo coupling should improve human-rated entity consistency over chunked generation without the plan, at equal compute. If FFHQ asymmetry persists under every variant, the honest conclusion is that seams are a local splitting error but global structure is a separate problem that interface coupling cannot fix.