When a generator furnishes a room, it has to decide how far one object's style should reach. Current systems make one of two implicit bets. Anchor-based and message-passing methods assume style travels unattenuated: one reference object, or one global echo, sets the look for everything it touches. Local relational methods assume style does not travel at all. Device physics has a precise name for the first bet. The ballistic assumption treats carriers, such as electrons, as crossing a channel without scattering. It holds over distances shorter than the carriers' mean free path, the typical distance a carrier travels before scattering. Beyond that distance it overestimates how much gets through. We propose to measure, in designer-curated houses, whether stylistic agreement between furniture behaves ballistically or decays over a characteristic distance in the scene graph (the graph whose nodes are objects and whose edges are spatial or functional relations).
If style does decay over a measurable distance, a generator should propagate it with that same decay, so that each room stays coherent while adjacent zones are free to differ. We implement this as a single joint linear solve over the scene graph, with a learned absorption term that sets the decay length. This replaces sequential cascades, which drift, and global anchors, which flatten everything to one style. The physics supplies the question, the vocabulary of decay length and boundary coupling, and a falsification standard. It is a lens, not a claim that furniture obeys transport equations.
Core Claim
Stylistic compatibility in curated multi-room houses decays with scene-graph distance over a finite length $\ell$. That length is longer than a single object-pair relation and shorter than a whole house. A style-propagation rule calibrated to $\ell$ should beat two alternatives on multi-room generation: the infinite-reach (ballistic) rule used implicitly by global anchoring, and the zero-reach rule of purely local assembly. If $\ell$ measured on data turns out to exceed house size, the ballistic assumption is adequate for style, a single global style token suffices, and the proposal is falsified. We treat that outcome as a legitimate result.
Why this gap, and what the prior work says
The physics literature motivates the question by showing that the ballistic assumption fails in a structured, quantifiable way rather than uniformly. Deuschle et al. add electron–electron scattering to device simulation. They find that the quasi-ballistic case overestimates silicon nanowire transistor on-current by roughly 3.5× and underestimates barrier length (8.5 nm versus about 10 nm), while keeping currents conserved within 1%. Ducry et al. show that ballistic ab initio models consistently overestimate conductance in room-temperature boron-nitride memristors. The lesson we borrow is methodological: the error of a no-scattering model depends on the ratio of system size to mean free path, so that length should be measured before choosing the model. DeePTB-NEGF contributes a second idea we reuse. It obtains a 100–1000× speedup by predicting interactions locally within a cutoff radius and then solving the global open-boundary problem in one step, rather than iterating. Our style solver has the same shape: local learned couplings, one global linear solve.
The scene-synthesis literature supplies the evidence for the gap. Style-Consistent 3D Indoor Scene Synthesis with Decoupled Objects stylizes a first object from text and conditions each later object on the anchor's CLIP features. It reports drift over long propagation chains, which is the cumulative error of a sequential, open-loop process. EchoScene echoes noisy node states through a graph network at every denoising step. Its style exchange has no notion of distance or attenuation and outputs untextured geometry. Pair2Scene assembles dense scenes from local pairwise relations, with a collision rate of 0.02 and zero floating objects. Its authors state that it cannot enforce stylistic consistency.
ICAR, DeepStyle and Creating Capsule Wardrobes constrain what the propagated signal should be. They show that visual similarity is the wrong carrier: propagating "looks like" yields duplicates, whereas compatibility embeddings pair items that differ in appearance but share a design language. SceneTransporter is the closest "transport" work. It gates cross-attention with an optimal-transport plan that assigns image patches to 3D parts. That is a mass-assignment problem within one image, with no concept of propagation distance or decay, so it addresses a different question from ours.
Research Program
The first stage is measurement. We train a compatibility encoder on 3D-FUTURE objects inside 3D-FRONT, following ICAR's design so that it separates complementarity from similarity. 3D-FRONT contains 6,813 designer houses with 18,968 furnished rooms. For each house we build a scene graph whose edges cover within-room relations (support, functional adjacency) and between-room relations (doorways, open-plan adjacency). We then plot mean pairwise compatibility as a function of shortest-path graph distance, separately for within-room and cross-room paths.
To make the fitted decay meaningful, we compare against two controls that should show no structure. In the first, object styles are shuffled within each house. In the second, objects are swapped between houses. The ballistic hypothesis predicts a flat curve until the house boundary. The local hypothesis predicts compatibility falling to the shuffled level within one hop. We fit the decay length $\ell$ and report whether it differs from room-level and house-level scales.
The second stage is the generator component. Each object $i$ carries a style vector $z_i$ in the compatibility space. Anchors $a$ are reference styles attached to chosen nodes, for example a user's reference image for a zone. All style vectors are obtained jointly by solving
$$(L_W + \kappa I + \Gamma)\, z = \Gamma a,$$
where $L_W$ is the graph Laplacian with edge weights $W$ predicted by a small network from relation type and object categories, $\kappa \ge 0$ is a learned absorption (scattering) rate, and $\Gamma$ is a diagonal matrix of coupling strengths to the anchors.
Setting $\kappa = 0$ recovers the ballistic limit: every connected object converges toward the anchors' common style. Large $\kappa$ recovers local-only behavior. On a regular graph the decay length scales as $\ell \sim \sqrt{w/\kappa}$, where $w$ is a typical edge weight; this is the standard screening length of diffusion with absorption. That relation lets us check whether the $\kappa$ learned end to end matches the $\ell$ measured in stage one.
Because the solve is joint and closed-loop rather than sequential, there is no ordering along which errors can accumulate. That is the specific mechanism we propose against the drift reported for cascaded stylization. The resulting $z_i$ then condition asset retrieval from 3D-FUTURE, which is our primary and cheapest setting. Optionally they also condition depth-to-image texturing, following the decoupled-objects pipeline.
The third stage is the regime analysis. We vary house size and number of zones and plot each method's error against house diameter divided by $\ell$. This mirrors how device papers plot ballistic error against channel length divided by mean free path. The prediction is that global anchoring is competitive on single rooms and degrades as the ratio grows, while the absorbing solver stays flat.
Baselines
The naive reference point is per-object retrieval conditioned only on category and layout, with no style coupling, run on the same scene graphs and asset pool. We also include a Pair2Scene-style local rule that matches each object only to its direct neighbors. This represents the zero-reach limit. On its own task, Pair2Scene reports an FID of 65.92 on its 3D-Pairs evaluation, but no stylistic-consistency number. That is consistent with its stated limitation.
The ballistic baseline is a single global style token per house, which is equivalent to our solver with $\kappa = 0$ and one anchor. This is the strongest simple competitor, and the regime analysis is designed to show where it fails. We also include per-room global tokens with room boundaries supplied by an oracle. This baseline cheats on zone structure, and the solver must at least approach it without that information.
The strongest existing methods are two. The first is cascaded anchor stylization from Style-Consistent 3D Indoor Scene Synthesis, which reports a CLIP Score of 0.245 on its own Muji, Modern Light Luxury and Starry Night evaluation. The second is EchoScene-style learned message passing over the graph, which reports a 15% bedroom FID improvement over CommonScenes on SG-FRONT. For retrieval-level style cohesion, ICAR reports a 31.8% Style FID improvement on furniture and a 5.3–9.6% Fill-In-The-Blank accuracy gain. These figures come from different setups, so we re-run each method on our splits rather than compare against the published numbers.
For fairness, all methods share the same compatibility encoder, the same 3D-FUTURE asset pool, the same held-out houses, and the same hyperparameter-search budget. The EchoScene-style network is given the same parameter count as our edge-weight network plus solver. Beating these baselines means lower Style FID and higher held-out fill-in-the-blank accuracy on multi-zone houses than both the cascade and the learned message-passing model, with no loss on single rooms.
Two ablations isolate the proposed ingredient. The first fixes $\kappa = 0$ while keeping learned $W$; this tests whether absorption specifically matters. The second replaces the compatibility carrier with CLIP similarity; this tests whether the carrier matters, given the complementarity evidence from ICAR.
Why not just…?
Why not just use one global style token or a single anchor? This is the most serious threat, and it is the ballistic hypothesis itself. If the stage-one measurement finds that $\ell$ exceeds house diameter, the global token is the right model and we will report that. Our claim depends on 3D-FRONT houses showing cross-room style variation that is structured, not random. The house-diameter-to-$\ell$ sweep and the per-room oracle baseline settle this question directly.
Why not just let a graph network learn attenuation, as in EchoScene? A sufficiently expressive learned message-passing model could in principle learn distance-dependent decay. Our argument is not that it cannot, but that it is not constrained to do so. EchoScene's iterative echo has no explicit decay term, and sequential propagation is what the decoupled-objects paper links to drift. The capacity-matched comparison tests whether an explicit, interpretable $\kappa$ beats implicit learning. We also check whether the trained graph network's effective influence actually decays with graph distance. If it matches the solver, the contribution reduces to the measurement result.
Why not just prompt a large language model per room for style tags? The decoupled-objects paper found that text alone was too ambiguous to keep styles consistent and needed visual anchors. Pair2Scene reports that language-model layout systems are slow, at about 210 seconds for Holodeck and 373 seconds for LayoutVLM. We include a per-room prompted baseline with the same retrieval backend, to test whether coarse zone labels from a language model already capture most of the benefit.
Why not just minimize visual-similarity distance across the house? ICAR's decoupling of similarity from complementarity yields its reported gains, and similarity search returns duplicates rather than complementary pieces. Our CLIP-carrier ablation measures how much of any improvement comes from the carrier rather than from the propagation rule.
Proposed Experiments and Decisive Checks
The first decisive check is the stage-one curve on 3D-FRONT. If compatibility versus graph distance is indistinguishable from the shuffled control beyond one hop, or flat across whole houses, the finite-length hypothesis fails in that direction and the generator work is not pursued. A clear decay with $\ell$ between room and house scale licenses the rest.
The second check asks whether the $\kappa$ learned end to end reproduces the measured $\ell$ through $\ell \sim \sqrt{w/\kappa}$. Agreement would indicate that the solver captures real structure in the data rather than acting as a tunable regularizer.
The third check is generation on held-out multi-zone houses. Metrics are Style FID, held-out-object fill-in-the-blank accuracy, CLIP score against zone anchors, and a "boundary sharpness" score: the compatibility gap between within-zone and cross-zone pairs, compared with the ground-truth gap. The analysis is stratified by house diameter divided by $\ell$.
The fourth check is a 20-participant preference study following the protocol of the decoupled-objects paper. Participants compare our output with the cascade and the global token on whole-house renders.
The expected pattern is that methods tie on single rooms, the cascade degrades with chain length, and the global token degrades with the number of zones. The solver should remain within reach of the per-room oracle without being given room boundaries.
References
- Electron-electron interactions in device simulation via nonequilibrium Green's functions and the GW approximation
- An ab initio study on resistance switching in hexagonal boron nitride
- Deep learning accelerated quantum transport simulations in nanoelectronics: from break junctions to field-effect transistors
- Style-Consistent 3D Indoor Scene Synthesis with Decoupled Objects
- EchoScene: Indoor Scene Generation via Information Echo Over Scene Graph Diffusion
- Pair2Scene: Learning Local Object Relations for Procedural Scene Generation
- ICAR: Image-Based Complementary Auto Reasoning
- DeepStyle: Multimodal Search Engine for Fashion and Interior Design
- Creating Capsule Wardrobes From Fashion Images
- SceneTransporter: Optimal Transport-Guided Compositional Latent Diffusion for Single-Image Structured 3D Scene Generation
- 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics