FurniMAS: Multi-Agent Furniture Decoration

• FurniMAS is a language-guided multi-agent system for automatic furniture decoration that converts empty furniture meshes into decorated 3D scenes.
• It integrates LLM-based semantic reasoning with non-LLM validation and constrained optimization to achieve precise, aesthetically coherent asset placement.
• The framework supports iterative editing by transforming natural-language instructions into asset selection, styling, and retrieval for complex 3D assembly.

FurniMAS is a term used in more than one arXiv context, but it most prominently denotes a language-guided multi-agent system for automatic furniture decoration. In that usage, FurniMAS takes an empty furniture item represented as a triangular mesh, a natural-language prompt, and a positive integer $N_{\text{assets}}$, and produces a decorated 3D furniture scene by selecting assets, assigning style and material, planning relative arrangements, solving exact placements through constrained optimization, and retrieving 3D models for instantiation (Nguyen et al., 7 Jul 2025). The system addresses furniture-surface decoration rather than room-scale furnishing, and is positioned as a hybrid of LLM-based semantic reasoning, non-LLM validation, optimization-based geometric realization, and retrieval-based 3D assembly.

1. Scope, task definition, and problem setting

FurniMAS addresses automatic furniture decoration, not room-level furnishing. The input object is a household furniture item such as a working desk or a TV stand, and the target output is a final decorated 3D furniture scene whose assets satisfy functionality, aesthetic coherence, and atmosphere preferences stated in natural language (Nguyen et al., 7 Jul 2025). The problem formulation explicitly includes three inputs: an empty furniture item represented as a triangular mesh, a natural-language human requirement, and the positive integer $N_{\text{assets}}$ specifying how many decorative or functional assets should be placed.

The task is defined at a finer spatial granularity than room-layout generation. The core difficulty is that furniture decoration requires denser packing, finer geometric reasoning, stronger clutter control, and more detailed style consistency than coarse room-scale furniture placement. The system must transform ambiguous language about practical use, mood, or style into concrete assets, styles, materials, relative arrangements, and physically realizable placements. It must also prevent assets from hanging outside support surfaces or colliding with one another. FurniMAS therefore treats decoration as a sequence of specialized subproblems rather than a single monolithic generation step (Nguyen et al., 7 Jul 2025).

The system also supports decorative editing. The supported editing operations include adding, removing, changing, resizing, rotating, or repositioning objects from natural-language instructions. A plausible implication is that FurniMAS is intended not only as a one-shot generator but also as an editable design assistant for iterative decoration workflows.

2. Multi-agent architecture and control structure

FurniMAS is implemented with AutoGen and uses GPT-4o as its default LLM backbone (Nguyen et al., 7 Jul 2025). Its architecture is explicitly hybrid: three LLM-based agents are paired with six non-LLM agents, and stage-specific communication is moderated by Admin roles. The LLM-based agents are the Asset Selector, Stylist, and Planner. The non-LLM agents are the System Admin, three Validators, the Arranger, and the Retriever.

The Asset Selector determines which assets should be placed, their approximate sizes, and the support surface assigned to each asset. The Stylist assigns style and material attributes. The Planner produces a scene-graph-like arrangement plan containing relative placements and orientations. These three stages are all validator-gated. The Asset Validator checks JSON compliance, verifies that the number of assets matches $N_{\text{assets}}$, ensures that all surfaces are populated if possible, and rejects any asset whose bounding-box width or length exceeds the maximum width or length of its assigned surface. The Style Validator ensures that styles and materials come only from the predefined bank. The Plan Validator checks that the planner’s output is structurally valid and usable downstream (Nguyen et al., 7 Jul 2025).

The stage protocol is iterative. If a stage output fails validation, the corresponding Validator issues continuous revision requests until the output passes. The system therefore does not propagate unchecked LLM output into later stages. The empirical motivation for this design is explicit: over 100 tasks, erroneous outputs were reported at frequencies of 26% for the Asset Selector, 18% for the Stylist, and 34% for the Planner (Nguyen et al., 7 Jul 2025). This makes validator-mediated correction central rather than auxiliary.

The agent decomposition is task-specific rather than conversationally open-ended. Communication is structured, serialized by stage, and schema-constrained. A plausible implication is that FurniMAS belongs to the class of neuro-symbolic pipelines in which LLMs are used primarily for semantic and aesthetic reasoning, while correctness, feasibility, and geometry are enforced by downstream algorithmic components.

3. Representations and end-to-end generation pipeline

The pipeline begins with surface extraction. Because a furniture item may contain multiple supporting surfaces, FurniMAS first extracts the polygon vertices, surface area, and surface height of each support surface using heuristic algorithms (Nguyen et al., 7 Jul 2025). These quantities are later used to determine how many assets should be placed on each surface, to validate whether asset sizes fit, and to compute feasible positions. The appendix further states that the system handles both flat and rugged surfaces by using a height tolerance of 2 cm in surface extraction and a drop operator in Blender so that assets conform to surface geometry.

Asset selection uses the user prompt, surface information, and $N_{\text{assets}}$. For each proposed asset, the Asset Selector outputs an asset name, a 3D bounding-box size in centimeters, and a surface index. The prompt constraints encourage wide variety, allow multiple instances of the same category, require that every surface be populated if possible, and specify that occupied area should be proportional to surface area. The selection procedure is also instructed to include a few large concept-defining assets together with several small detail assets, while preserving practical functionality (Nguyen et al., 7 Jul 2025).

Stylizing and materializing assign aesthetic attributes to the selected assets. The Stylist is restricted to a predefined bank of 31 styles and 17 materials. The style bank includes Contemporary, Coastal, Scandinavian, Shabby Chic, Transitional, Modern, Mid-century, Retro, Minimalist, Traditional, Farmhouse, Antique, Industrial, Rustic, Vintage, Mission, French, Art Deco, Victorian, Chippendale, Country, Craftsman, Shaker, Queen Anne, Hepplewhite, Louis XVI, Asian, Jacobean, Colonial, Federal, and Sheraton. The material bank includes wood, plywood, marble, paper, fibre, plastic, glass, textile, iron, steel, gold, silver, bronze, cotton, linen, leather, and paper; the duplication of “paper” is noted in the appendix (Nguyen et al., 7 Jul 2025).

Arrangement planning remains symbolic rather than metric. Global placement is represented by one of the $3 \times 3$ divided regions on a support surface. Local relations may be defined relative to another asset on the same surface. The supported relative positions are left of, right of, in front of, behind, and on top of; supported relative distances are near and far; and supported alignments are vertical-right, vertical-mid, vertical-left, horizontal-front, horizontal-mid, and horizontal-back. Orientation is discretized to forward, backward, left, and right (Nguyen et al., 7 Jul 2025). The planner is explicitly prompted to preserve functionality, as illustrated by the example of placing a monitor at the center, a keyboard in front of the monitor, and a mouse to the right of and near the monitor.

The final stage of symbolic-to-geometric instantiation is asset retrieval and 3D assembly. The Retriever uses OpenShape for text-based 3D asset retrieval from Objaverse. Retrieval queries are formed from proposed asset names, styles, and materials, and one asset is randomly selected from the top 10 semantically similar candidates to increase diversity (Nguyen et al., 7 Jul 2025). The retrieved assets are then placed in Blender according to the solved positions and orientations.

4. Geometric optimization, physical feasibility, and editing

FurniMAS does not ask the LLM to emit exact coordinates directly. Instead, exact arrangement is delegated to a non-LLM Arranger that solves a constrained optimization problem with Gurobi (Nguyen et al., 7 Jul 2025). For each asset $i$, the variables are a 2D position $\mathbf{p}_i = (x_i, y_i)$ on its support surface and a discrete orientation $\mathbf{O}_i$ represented by two Boolean variables $(r_{90}, r_{180})$. If both are true, the total rotation is 270 degrees.

The optimization objective is

$$\text{maximize}\: s \left (\mathbf{p}, \mathbf{O}; \mathbf{S} \right) \triangleq \sum_{(i,j)\in\mathcal{L}^{s}} s_{ij}\left (\mathbf{p}_{i}, \mathbf{O}_{i}, \mathbf{p}_{j}, \mathbf{O}_{j} \right ).$$

The constraints are

$N_{\text{assets}}$0

Here $N_{\text{assets}}$1 is the set of all assets, $N_{\text{assets}}$2 is the set of pairs with soft local distance or alignment relations, $N_{\text{assets}}$3 is the set of pairs with hard local placement relations, $N_{\text{assets}}$4 is the set of assets with specified global placements, and $N_{\text{assets}}$5 is the support-surface polygon for asset $N_{\text{assets}}$6 (Nguyen et al., 7 Jul 2025).

The interpretation of the constraints is direct. $N_{\text{assets}}$7 enforces surface containment, $N_{\text{assets}}$8 enforces collision avoidance among assets on the same surface, $N_{\text{assets}}$9 enforces membership in a specified $N_{\text{assets}}$0 surface region, and $N_{\text{assets}}$1 enforces hard local relations such as left of, right of, in front of, behind, and on top of. The optimization layer is therefore the component responsible for the perfect out-of-boundary and bounding-box collision performance reported in the quantitative evaluation.

FurniMAS is not trained end to end. The system is described as an integration-and-optimization pipeline using GPT-4o for language agents, Gurobi for arrangement optimization, and OpenShape for retrieval (Nguyen et al., 7 Jul 2025). A plausible implication is that its strengths are modularity and controllability rather than likelihood-based generation or end-to-end learned scene distributions.

5. Empirical performance and evaluation protocol

The main evaluation uses 200 furniture items from various household categories, rendered with Blender, and tested at three decoration densities: 8 assets, 16 assets, and 32 assets (Nguyen et al., 7 Jul 2025). For each decorated result, four images are rendered: left side view, right side view, front view, and back view. The baselines are adapted versions of LayoutGPT, Holodeck, and I-Design, all using GPT-4o as the LLM backbone for fairness.

Two feasibility metrics are used: OOB, the Out-of-Boundary Rate, and BBL, the Bounding Box Loss. Four GPT-4o-based quality scores are also reported on a 1-to-10 scale: Func for functionality, Layout for spatial organization, Scheme for style scheme and material harmony, and Atmos for aesthetic atmosphere. Each scene is graded 10 times by GPT-4o and the scores are averaged (Nguyen et al., 7 Jul 2025).

At 8 assets, FurniMAS achieves OOB $N_{\text{assets}}$2, BBL $N_{\text{assets}}$3, Func $N_{\text{assets}}$4, Layout $N_{\text{assets}}$5, Scheme $N_{\text{assets}}$6, and Atmos $N_{\text{assets}}$7 (Nguyen et al., 7 Jul 2025). At 16 assets, it achieves OOB $N_{\text{assets}}$8, BBL $N_{\text{assets}}$9, Func $N_{\text{assets}}$0, Layout $N_{\text{assets}}$1, Scheme $N_{\text{assets}}$2, and Atmos $N_{\text{assets}}$3. At 32 assets, it achieves OOB $N_{\text{assets}}$4, BBL $N_{\text{assets}}$5, Func $N_{\text{assets}}$6, Layout $N_{\text{assets}}$7, Scheme $N_{\text{assets}}$8, and Atmos $N_{\text{assets}}$9. Across all three densities, it outperforms the adapted baselines on the reported semantic and aesthetic metrics, while tying Holodeck on perfect OOB and BBL.

The system is also evaluated on decorative editing over 200 decorated scenes with four editing types: inserting or removing assets, changing assets, resizing assets, and rotating or repositioning assets. FurniMAS reports Func $3 \times 3$0, Layout $3 \times 3$1, Scheme $3 \times 3$2, Atmos $3 \times 3$3, and Edit $3 \times 3$4 (Nguyen et al., 7 Jul 2025). In the arrangement-only analysis on 100 surfaces, where each method receives the same text description and asset set and differs only in arrangement generation, FurniMAS obtains a GPT-4o score of $3 \times 3$5, compared with $3 \times 3$6 for LayoutGPT, $3 \times 3$7 for Holodeck, and $3 \times 3$8 for I-Design.

Ablations further support the specialization thesis. The best reported multi-agent configuration gives Func $3 \times 3$9, Layout $i$0, Scheme $i$1, and Atmos $i$2, while merged-task variants perform worse (Nguyen et al., 7 Jul 2025). Furniture-type analysis shows the strongest performance on workspace furniture, with Func $i$3, Layout $i$4, Scheme $i$5, and Atmos $i$6, and somewhat lower but still competitive results on living space, bedroom, and kitchen furniture. The LLM-backbone comparison shows that GPT-4o yields Func $i$7, Layout $i$8, Scheme $i$9, Atmos $\mathbf{p}_i = (x_i, y_i)$0, time $\mathbf{p}_i = (x_i, y_i)$1 min, and cost $\mathbf{p}_i = (x_i, y_i)$2, which is presented as the best quality-speed-cost tradeoff among GPT-3.5, GPT-4, GPT-4.5, and GPT-4o (Nguyen et al., 7 Jul 2025).

The main limitations are also explicit. Performance drops as asset count increases, the current system supports only basic placement operations, more complex operations such as hanging and draping are left for future work, surface complexity is only partially handled, and evaluation relies heavily on LLM judges rather than a traditional human-subject study (Nguyen et al., 7 Jul 2025).

6. Position in the research landscape and alternate use of the name

Within the broader arXiv literature on interior generation, FurniMAS occupies the furniture-surface regime rather than the room-scale regime. HSM, for example, frames indoor scene generation as recursive support-region population and explicitly targets dense object arrangements across room, furniture, and tabletop scales (Pun et al., 21 Mar 2025). Architect-Ant treats furnishing as an editable floor-plan layout problem using a coordinate-based DSL and deterministic rule scoring (Rodionov et al., 9 Jun 2026). FurniScene provides a dense object-centric dataset and a Two-Stage Diffusion Scene Model for intricately furnished rooms (Zhang et al., 2024). Earlier work on constrained furniture layout generation models room furniture through interpretable probabilistic procedural statistics, motifs, and rejection-sampled constraints (Henderson et al., 2017). Relative to these systems, FurniMAS is distinguished by its mesh-level furniture-decoration scope, multi-agent specialization, and optimization-backed realization of local arrangements (Nguyen et al., 7 Jul 2025).

The name “FurniMAS” is also used in a separate materials-science context. In that literature, FurniMAS denotes a 3D-printed, water-cooled, non-contact high-temperature furnace/sample-environment platform for multimodal in situ synchrotron X-ray experiments at ESRF ID03 (Lesage et al., 1 Jul 2025). That system is unrelated to furniture decoration: it is designed for stable operation up to $\mathbf{p}_i = (x_i, y_i)$3, heating rates above $\mathbf{p}_i = (x_i, y_i)$4, thermal stability better than $\mathbf{p}_i = (x_i, y_i)$5, full $\mathbf{p}_i = (x_i, y_i)$6 rotation, and $\mathbf{p}_i = (x_i, y_i)$7 tilt freedom, and supports DFXM, 3DXRD, MTT, PCT, and DCT (Lesage et al., 1 Jul 2025). The homonymy is therefore terminological rather than methodological.

In its principal usage, FurniMAS is best understood as a modular decoration framework that decomposes furniture-surface styling into support-surface analysis, semantic asset proposal, controlled aesthetic assignment, relational planning, constrained layout optimization, and retrieval-based 3D instantiation (Nguyen et al., 7 Jul 2025). This suggests a broader design pattern for controllable 3D content generation: use LLMs where language understanding and commonsense priors dominate, use validators where structural correctness matters, and use optimization where geometry must be exact.