DisCo-Layout: 3D Indoor Layout Synthesis

• The paper introduces a novel framework that disentangles semantic from physical refinement using a multi-agent coordination loop to ensure collision-free and coherent placements.
• It employs distinct refinement tools—SRT for semantic adjustments and PRT for physical corrections—to overcome the limitations of coupled optimizers.
• Empirical results on 45 indoor scenes demonstrate zero collision and out-of-bound violations, outperforming baselines like LayoutGPT and Holodeck.

Searching arXiv for the specified papers and closely related layout-generation work. First, retrieving the DisCo-Layout paper by arXiv id. DisCo-Layout is a framework for 3D indoor layout synthesis that disentangles semantic refinement from physical refinement and coordinates them within a multi-agent system (Gao et al., 2 Oct 2025). In its formulation, the input consists of a natural language prompt $T$, a set of retrieved 3D assets $O=\{o_1,\dots,o_n\}$, and a room with boundary walls $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$; the output is a set of poses $P=\{p_1,\dots,p_n\}$ with $p_i=(x_i,y_i,\theta_i)$ such that object placements satisfy both semantic coherence and physical plausibility (Gao et al., 2 Oct 2025). The framework’s central claim is that semantic richness and robust flexible refinement should not be handled by a single coupled optimizer: instead, abstract object relationships are revised by a Semantic Refinement Tool (SRT), while coordinate-level feasibility is enforced by a Physical Refinement Tool (PRT), with a Planner–Designer–Evaluator loop orchestrating both (Gao et al., 2 Oct 2025). A related but distinct usage of the same decoupling intuition appears in layout generation for 2D graphic scenes, where LDGM decouples diffusion across category, position, and size in a discrete diffusion model; that work does not reference DisCo-Layout explicitly, but it embodies the same high-level principle of separating attribute-specific corruption from coordinated generation (Hui et al., 2023).

1. Formal problem setting

DisCo-Layout addresses 3D indoor layout synthesis under open-ended prompts and heterogeneous asset sets (Gao et al., 2 Oct 2025). The task is defined as producing object poses $P=\{p_1,\dots,p_n\}$ with $p_i=(x_i,y_i,\theta_i)$ given a prompt $T$, retrieved assets $O=\{o_1,\dots,o_n\}$, and room boundaries $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$ (Gao et al., 2 Oct 2025). Two classes of requirements are imposed. Semantic coherence means that placements reflect functional and spatial relationships implied by the prompt, including patterns such as “chairs around table” and “TV faces sofa” (Gao et al., 2 Oct 2025). Physical plausibility means that objects lie within $O=\{o_1,\dots,o_n\}$0, avoid collisions, and satisfy explicit spatial constraints such as “against wall” (Gao et al., 2 Oct 2025).

The paper motivates this formulation by contrasting it with fixed-dataset approaches that rely on in-domain assets and fixed placement schemes, which impedes generalization to unseen assets and layouts as well as open-ended prompts and constraints (Gao et al., 2 Oct 2025). It further distinguishes semantic richness from robust flexible refinement. Semantic richness refers to high-level relational logic such as adjacency, grouping, facing, or functional pairings; robust flexible refinement refers to iterative, targeted corrections for collisions, out-of-bounds placements, and wall alignment without degrading semantic structure (Gao et al., 2 Oct 2025). The stated argument is that single-step LLM/VLM methods and coupled refiners either lack consistency or entangle objectives, causing interference (Gao et al., 2 Oct 2025).

A plausible implication is that the framework is designed less as a monolithic generator than as a decomposition of layout synthesis into semantically interpretable and physically verifiable subproblems.

2. System architecture and multi-agent coordination

The framework is organized around three agents and two refinement tools: a Planner, a Designer, an Evaluator, the Semantic Refinement Tool (SRT), and the Physical Refinement Tool (PRT) (Gao et al., 2 Oct 2025). The Planner derives high-level placement rules and semantic asset groups with an ordered placement schedule, producing constraints $O=\{o_1,\dots,o_n\}$1 and ordered groups $O=\{o_1,\dots,o_n\}$2 via

$O=\{o_1,\dots,o_n\}$3

The Designer then proposes an initial layout for each group using VLM reasoning over the current scene image and the relevant constraints:

$O=\{o_1,\dots,o_n\}$4

where $O=\{o_1,\dots,o_n\}$5 is the rendered top-down image of previously placed groups (Gao et al., 2 Oct 2025). The Evaluator assesses semantic coherence and physical plausibility using targeted VQA-style questions:

$O=\{o_1,\dots,o_n\}$6

with $O=\{o_1,\dots,o_n\}$7 (Gao et al., 2 Oct 2025).

The pipeline proceeds group by group. First, the Planner computes constraints per asset and groups them by placement priority. Second, the Designer places assets in group $O=\{o_1,\dots,o_n\}$8 into the scene and renders $O=\{o_1,\dots,o_n\}$9. Third, the Evaluator performs structured VQA on semantic constraints and physical issues, producing booleans $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$0. Fourth, if $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$1, SRT is applied; if $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$2, PRT is applied; the layout is then updated to $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$3. Fifth, the process iterates until all groups have been placed and refined (Gao et al., 2 Oct 2025).

The paper also specifies the coordination loop in pseudocode. Starting from $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$4, the system initializes $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$5 and an empty room render $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$6. For each group $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$7, it computes $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$8, forms $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$9, renders $P=\{p_1,\dots,p_n\}$0, evaluates $P=\{p_1,\dots,p_n\}$1, conditionally applies SRT and PRT, renders the updated scene, and continues until the final layout $P=\{p_1,\dots,p_n\}$2 is returned (Gao et al., 2 Oct 2025). The stopping criterion is explicit: per group, refinement continues until both booleans are False, after which the system proceeds to the next group (Gao et al., 2 Oct 2025).

This staged design separates planning, initial proposal, diagnosis, semantic repair, and physical repair into discrete operators. The paper’s interpretation is that such modularity avoids the objective interference of coupled optimization (Gao et al., 2 Oct 2025).

3. Semantic refinement and constraint representation

Semantic refinement is handled by the Semantic Refinement Tool, which operates when the Evaluator detects semantic issues, i.e., when $P=\{p_1,\dots,p_n\}$3 (Gao et al., 2 Oct 2025). Constraints are specified per object as

$P=\{p_1,\dots,p_n\}$4

where relationship types satisfy

$P=\{p_1,\dots,p_n\}$5

and the facing constraint $P=\{p_1,\dots,p_n\}$6 identifies the target object an asset should face (Gao et al., 2 Oct 2025).

When semantic failures are detected, the failed constraints $P=\{p_1,\dots,p_n\}$7 are converted to explicit textual feedback such as “The chair is not facing the table,” and the VLM proposes minimal position and rotation changes that satisfy the failed constraints while preserving the overall layout (Gao et al., 2 Oct 2025). The tool is defined as

$P=\{p_1,\dots,p_n\}$8

Its interface takes the current layout $P=\{p_1,\dots,p_n\}$9, failed semantic constraints $p_i=(x_i,y_i,\theta_i)$0, the current scene image $p_i=(x_i,y_i,\theta_i)$1, and asset sizes, and produces JSON patches only for the objects that need changes (Gao et al., 2 Oct 2025).

The described SRT procedure is explicit. It extracts all failed constraints and formulates feedback text for each one; prompts the VLM with the current layout encoded as JSON with object name, size, position, and rotation; enforces a restricted output schema containing only changed objects; parses the resulting JSON; and minimally updates $p_i=(x_i,y_i,\theta_i)$2 relative to prior values before returning the revised layout $p_i=(x_i,y_i,\theta_i)$3 (Gao et al., 2 Oct 2025).

In the planner’s JSON schema, each object may include against_wall, relative_position, relative_object, and rotation fields (Gao et al., 2 Oct 2025). The Evaluator’s structured VQA schema likewise returns per-question JSON with keys question, objects, reason, and answer (Gao et al., 2 Oct 2025). This suggests that semantic refinement is not merely an implicit latent correction mechanism, but an explicitly inspectable constraint-repair process.

4. Physical refinement and grid-matching optimization

Physical refinement is performed by the Physical Refinement Tool when $p_i=(x_i,y_i,\theta_i)$4 (Gao et al., 2 Oct 2025). The tool is designed to correct coordinate-level errors including collisions, out-of-bounds placements, and wall-alignment violations using a grid-matching algorithm (Gao et al., 2 Oct 2025). The update rule is

$p_i=(x_i,y_i,\theta_i)$5

The floor is discretized into a uniform grid $p_i=(x_i,y_i,\theta_i)$6 (Gao et al., 2 Oct 2025). Repositioning is constrained by three validity criteria. First, placements must be collision-free, with collision measured by

$p_i=(x_i,y_i,\theta_i)$7

where $p_i=(x_i,y_i,\theta_i)$8 is object $p_i=(x_i,y_i,\theta_i)$9’s oriented footprint rectangle on the floor (Gao et al., 2 Oct 2025). Second, each footprint must be fully contained in $P=\{p_1,\dots,p_n\}$0 (Gao et al., 2 Oct 2025). Third, if $P=\{p_1,\dots,p_n\}$1, the object’s back face must be flush with the nearest wall and its rotation must be updated to face inward (Gao et al., 2 Oct 2025).

For an invalid object $P=\{p_1,\dots,p_n\}$2 with current center $P=\{p_1,\dots,p_n\}$3, the grid-matching objective is

$P=\{p_1,\dots,p_n\}$4

subject to collision-free placement, room containment, and wall alignment if required (Gao et al., 2 Oct 2025). The solver therefore selects the nearest feasible grid cell to minimize displacement while preserving semantic relationships (Gao et al., 2 Oct 2025).

The paper specifies a three-stage PRT pseudocode. In the wall-alignment stage, if an object is against a wall, the nearest wall is found, rotation is set by wall2rotation, the object’s back center is computed, candidate wall grids are sorted by distance, and the object is either pulled to the nearest feasible wall grid or deleted if none exists (Gao et al., 2 Oct 2025). In the out-of-bounds correction stage, out-of-bounds objects are detected, general grid candidates are sorted by distance, invalid candidates are filtered, and the nearest valid position is chosen or the object is deleted if no candidate exists (Gao et al., 2 Oct 2025). In the collision-resolution stage, collision pairs are found, the smaller object in each pair is selected, candidate grids are sorted by distance, collision-inducing candidates are filtered, and the nearest feasible position is assigned or the smaller object is deleted if no feasible cell exists (Gao et al., 2 Oct 2025).

The paper notes that footprint handling uses object sizes and rotations restricted to $P=\{p_1,\dots,p_n\}$5 degrees to compute oriented bounding boxes, and that collision checks use occupancy maps with zero-overlap enforced by IoU (Gao et al., 2 Oct 2025). Complexity analysis is not provided (Gao et al., 2 Oct 2025). A notable consequence of the algorithm is that physical feasibility is guaranteed only up to the resolution of the chosen grid and the fidelity of the asset bounding boxes.

5. Models, interfaces, and implementation details

GPT-4o is used both for prompt generation in dataset construction and for agent reasoning in the Planner, Designer, Evaluator, and SRT through structured VQA and constrained JSON outputs (Gao et al., 2 Oct 2025). The Designer operates in a 2D coordinate system whose origin is at the bottom-left corner $P=\{p_1,\dots,p_n\}$6, with positions expressed as positive integers in centimeters and rotations restricted to $P=\{p_1,\dots,p_n\}$7 degrees, where $P=\{p_1,\dots,p_n\}$8 faces $P=\{p_1,\dots,p_n\}$9, $p_i=(x_i,y_i,\theta_i)$0 faces $p_i=(x_i,y_i,\theta_i)$1, $p_i=(x_i,y_i,\theta_i)$2 faces $p_i=(x_i,y_i,\theta_i)$3, and $p_i=(x_i,y_i,\theta_i)$4 faces $p_i=(x_i,y_i,\theta_i)$5 (Gao et al., 2 Oct 2025). Visual inputs are top-down rendered images with brown floors and white walls, and the Evaluator consumes these images together with JSON metadata (Gao et al., 2 Oct 2025).

The Planner outputs semantic asset groups such as group1, group2, and orders them by importance, typically placing larger or focal assets first (Gao et al., 2 Oct 2025). Its relationship vocabulary is exactly [near, side of, in front of, aligned with, opposite, around] (Gao et al., 2 Oct 2025). The Designer takes the previously placed objects (arranged_objects) and the top-down render $p_i=(x_i,y_i,\theta_i)$6 as inputs and is instructed to output JSON only (Gao et al., 2 Oct 2025). The Evaluator performs semantic assessment via targeted yes/no questions derived from $p_i=(x_i,y_i,\theta_i)$7 and $p_i=(x_i,y_i,\theta_i)$8, and a separate binary physical assessment that determines whether a physical refinement tool is required (Gao et al., 2 Oct 2025).

The asset catalog is retrieved from Objaverse and curated for semantic fidelity and geometric quality following the Holodeck protocol (Gao et al., 2 Oct 2025). The paper does not report grid resolution, architectural hyperparameters, training or fine-tuning details, or runtime and memory usage; PRT is explicitly heuristic rather than learned (Gao et al., 2 Oct 2025). Code is stated to be publicly available, and the appendix fully specifies prompts for the Planner, Designer, Evaluator, and SRT, along with the grid-matching pseudocode for PRT (Gao et al., 2 Oct 2025).

These interface decisions make the framework unusually transparent relative to end-to-end neural layout generators. A plausible implication is that reproducibility depends less on hidden weights in a trained physical module and more on prompt fidelity, rendering setup, and exact implementation of the grid-matching routines.

6. Evaluation, baselines, and empirical behavior

The evaluation uses a curated test set of 45 indoor scenes spanning 9 categories: five common categories—bathroom, bedroom, dining room, kitchen, living room—and four uncommon categories—buffet restaurant, classroom, children’s room, home gym (Gao et al., 2 Oct 2025). Prompts are synthesized by GPT-4o and range from simple labels to multi-clause specifications; assets are retrieved via Objaverse and manually screened (Gao et al., 2 Oct 2025). Baselines are LayoutGPT, Holodeck, and LayoutVLM (Gao et al., 2 Oct 2025).

Physical plausibility is measured with Collision Rate (CR), the proportion of object pairs with footprint IoU greater than zero,

$p_i=(x_i,y_i,\theta_i)$9

and Out-of-Bounds Rate (OOB), the average fraction of each footprint area that lies outside the room boundaries,

$T$0

Semantic coherence is evaluated with GPT-4o scores for Positional Coherency and Rotational Coherency (Gao et al., 2 Oct 2025).

The main quantitative results reported for the average over the nine categories are summarized below.

Method	Pos. / Rot.	Col. / OOB
DisCo-Layout	67.89 / 66.88	0.00 / 0.00
LayoutGPT	66.78 / 63.38	5.10 / 18.01
Holodeck	64.8 / 58.2	0.00 / 0.25
LayoutVLM	63.31 / 58.74	9.09 / 8.54

The paper highlights zero physical violations across all nine categories for DisCo-Layout, with Collision $T$1 and OOB $T$2 (Gao et al., 2 Oct 2025). It also reports category-specific results including Dining Room with Pos. $T$3, Rot. $T$4, Collision $T$5, OOB $T$6, and Living Room with Pos. $T$7, Rot. $T$8, Collision $T$9, OOB $O=\{o_1,\dots,o_n\}$0 (Gao et al., 2 Oct 2025).

Ablation studies attribute improvements to the disentangled refinement design. Removing SRT yields Collision $O=\{o_1,\dots,o_n\}$1, OOB $O=\{o_1,\dots,o_n\}$2, Pos. $O=\{o_1,\dots,o_n\}$3, Rot. $O=\{o_1,\dots,o_n\}$4; removing PRT yields Collision $O=\{o_1,\dots,o_n\}$5, OOB $O=\{o_1,\dots,o_n\}$6, Pos. $O=\{o_1,\dots,o_n\}$7, Rot. $O=\{o_1,\dots,o_n\}$8; removing both yields Collision $O=\{o_1,\dots,o_n\}$9, OOB $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$0, Pos. $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$1, Rot. $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$2; replacing structured VQA with open-ended VQA yields Collision $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$3, OOB $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$4, Pos. $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$5, Rot. $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$6; the full system achieves Collision $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$7, OOB $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$8, Pos. $B_{\text{room}}=\{b_1,b_2,b_3,b_4\}$9, Rot. $O=\{o_1,\dots,o_n\}$00 (Gao et al., 2 Oct 2025). The paper’s interpretation is that disentangled refinement coordinated by structured VQA yields the best overall quality (Gao et al., 2 Oct 2025).

Qualitatively, the framework is described as correcting orientations such as chairs facing a table, enforcing relationship types like near, aligned, and opposite, creating functional zones such as a bar cart adjacent to a dining table, resolving collisions, pulling large objects against walls, and ensuring in-bounds placement (Gao et al., 2 Oct 2025). When PRT cannot find a feasible grid position, it deletes the offending object, which prevents persistent physical violations (Gao et al., 2 Oct 2025).

7. Relation to decoupled layout generation, limitations, and significance

The name “DisCo-Layout” refers specifically to the 2025 framework for 3D indoor layout synthesis that disentangles semantic and physical refinement (Gao et al., 2 Oct 2025). The data also identifies a conceptual connection to LDGM, a 2023 method for 2D layout generation in formatted scenes. LDGM treats any coarsely specified or partially observed layout as an intermediate diffusion state and uses a decoupled diffusion process across category, position, and size, while learning a single joint reverse model with a transformer backbone and relation-aware attention (Hui et al., 2023). The paper summary states that LDGM and “DisCo-Layout” share the core idea of decoupling diffusion across layout attributes, but also states explicitly that LDGM does not reference “DisCo-Layout” and should be regarded as a different method that embodies the same decoupling principle in a discrete diffusion framework (Hui et al., 2023).

This comparison is significant because the two works instantiate decoupling at different levels. In LDGM, the separation occurs in the forward corruption process over layout attributes, followed by joint reverse denoising (Hui et al., 2023). In DisCo-Layout, the separation occurs between semantic correction and physical correction, mediated by a multi-agent orchestration loop (Gao et al., 2 Oct 2025). This suggests a broader design pattern in layout generation: decomposing heterogeneous constraints into specialized operators while retaining some coordinating mechanism.

The limitations of DisCo-Layout are stated directly. The framework depends on GPT-4o for reasoning and is sensitive to prompt design; open-ended VQA is less reliable than structured VQA (Gao et al., 2 Oct 2025). The discrete grid and restricted rotation set may limit placement resolution and nuanced orientations (Gao et al., 2 Oct 2025). PRT may delete objects when no grid cell satisfies all constraints (Gao et al., 2 Oct 2025). The method contains no explicit training or learned physical model, so the heuristic PRT may be sensitive to grid resolution and asset bounding-box quality (Gao et al., 2 Oct 2025). The approach also relies on manual asset screening and Objaverse, so domain shift may affect semantic judgments (Gao et al., 2 Oct 2025). Future work is identified as enhancing generalization to more complex scenarios and improving training for embodied AI applications, potentially with learned refinement components (Gao et al., 2 Oct 2025).

Within the scope defined by the paper, DisCo-Layout’s main contribution is the claim that semantic and physical layout quality benefit from being refined separately but coordinated explicitly. The reported zero collision and zero out-of-bounds rates, together with higher positional and rotational coherency than the listed baselines, support that claim on the 45-scene benchmark used in the study (Gao et al., 2 Oct 2025).