IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 1

Eliminating Rasterization: Direct Vector

Floor Plan Generation with DiffPlanner
Shidong Wang and Renato Pajarola, Senior Member, IEEE

Abstract—The boundary-constrained floor plan generation generative model (e.g., CNN or GAN) to directly generate a
problem aims to generate the topological and geometric proper- raster image that contains the geometric and topological in-
ties of a set of rooms within a given boundary. Recently, learning- formation of the room(s), and eventually extracting the vector
based methods have made significant progress in generating
realistic floor plans. However, these methods involve a workflow information from this raster output. Recently, some approaches
arXiv:2508.13738v1 [[Link]] 19 Aug 2025

of converting vector data into raster images, using image-based have used a hybrid strategy, i.e., using graph neural networks
generative models, and then converting the results back into vec- (GNN) or Transformer, to directly predict vector information
tor data. This process is complex and redundant, often resulting about the layout, such as the coordinates of a bounding box
in information loss. Raster images, unlike vector data, cannot representing a room [8], [11], or a bubble diagram containing
scale without losing detail and precision. To address these issues,
we propose a novel deep learning framework called DiffPlanner the locations of nodes and their adjacencies [28]. However,
for boundary-constrained floor plan generation, which operates they still depend on image-based generative models, e.g., to
entirely in vector space. Our framework is a Transformer-based perform feature extraction on the input raster boundaries and
conditional diffusion model that integrates an alignment mecha- the ground truth raster images [8], [11], [28], or to generate a
nism in training, aligning the optimization trajectory of the model raster layout image to provide additional information for floor
with the iterative design processes of designers. This enables our
model to handle complex vector data, better fit the distribution plan prediction [8], [11].
of the predicted targets, accomplish the challenging task of floor Existing methods could generate some results comparable
plan layout design, and achieve user-controllable generation. to those of professional architects, thanks to the carefully
We conduct quantitative comparisons, qualitative evaluations, designed models. However, their workflow of converting vec-
ablation experiments, and perceptual studies to evaluate our tor information into images, then performing the generation
method. Extensive experiments demonstrate that DiffPlanner
surpasses existing state-of-the-art methods in generating floor process via an image-based generative model, and finally
plans and bubble diagrams in the creative stages, offering more converting the raster results into vector data is unnecessarily
controllability to users and producing higher-quality results that redundant and overly complex. Moreover, when converting
closely match the ground truths. vector data, like floor plans, to raster images, information loss
Index Terms—Floor plan generation, bubble diagram, deep often occurs [5], [13]. Vector data, which is defined by precise
generative modeling. geometry and topology, can scale infinitely without distortion.
In contrast, raster images are made of fixed-resolution pixels,
which can result in lost details and precision, and are scale-
I. I NTRODUCTION
dependent. Common issues include blurred edges, altered

D ATA-DRIVEN boundary-constrained floor plan gener-

ation, which attempts to generate the geometric and
topological properties of a set of rooms given the interior of a
angles, broken lines, and oversimplified complex structures
during the rasterization process.
Based on this, we considered whether a model could be
boundary, has recently attracted widespread interest within the developed that learns directly from vector data and outputs
computer graphics, vision, and architecture communities [4], layouts in the same format. This proves to be a challenging
[8], [11], [27], [30], [32]. Floor plan generation represents a task. For the input, learning directly from vector data is
typical coarse-to-fine layout design process. The input to this difficult as it is not as standardized as learning from raster
task is the exterior boundary of a building, and the output images. To illustrate this, when faced with raw vector data
should contain the number, categories, locations, adjacencies, (Figure 1(a)), humans often have only a superficial under-
and partitioning of rooms inside the boundary. Also, the output standing, but a raster image (Figure 1(b)) allows us to quickly
is expected to be in the format of vector data, which is critical grasp basic topological and geometric information. For the
for practical use by architects and designers, and which is also output, we expect the model not only to adapt to the early
the reason why a large number of researchers have investigated stages of design, offering rough designs that include geometric
the vectorization of raster floor plans [12], [17]. and topological information (e.g., bubble diagrams), but also
As illustrated in Figure 1(b), most of the existing previous to produce more detailed layout results (e.g., accurate floor
methods have adopted a raster-to-raster generation pipeline [4], plans). The output needs to be significantly more complex than
[27], [30], [32], i.e., first rasterizing the boundary informa- the input boundaries, meaning that the model must be capable
tion from a vector dataset, and then using an image-based of learning the distribution from more complex vector data
and ultimately producing detailed vector layouts from simple
S. Wang and R. Pajarola are with the Department of Informatics, University
of Zurich, Switzerland. E-mails: {shwang, pajarola}@[Link]. inputs (e.g., boundaries). Regarding the generation process,
Received 30 July 2024; revised 12 March 2025; accepted 5 April 2025. we anticipate that the model will be fully automatic, capable
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 2

Bubble diagram Floor plan

Input boundary & entrance Output room nodes Output room adjacencies Output room partitioning
{

"boundary": [[x, y], ... ],

Input
"entrance": [[x, y], ... ],
"nodes": [{
"id": int,
"category": int,
"size": int, (b) Raster images for floor plan generation
Output
"location": [x, y]}, ... ],
"adjacencies":
[[id, id], ... ],
"partitioning":
[[[x, y], ... ], ... ],

}
(a) Vector floor plan (c) Data tensors of vector-based floor plan generation

Fig. 1. Most previous methods with input and output vector data (a), internally rasterize the input into a series of images representing the layout and then
use image-based networks for the prediction process followed by vectorization of the output (b). In contrast, our DiffPlanner performs the prediction process
directly based on data tensors representing the vector space (c).

of autonomously generating the complete floor plan layouts based on diffusion models, which directly predicts vector
without any information beyond the boundaries. However, floor plans from vector boundaries. Previous methods
the process should be controllable on a high level, allowing depended (internally) on raster space, requiring the
users to input simple conditions such as the number of rooms conversion of vector information into a set of raster
and their categories. Additionally, it should be adjustable, images, followed by prediction processes using image-
permitting iterative user interventions at each step of the based generative CNN or GAN models.
design. 2) We have also designed an alignment mechanism to better
Based on this, we propose DiffPlanner, a novel deep learn- align the optimization trajectory of the model with the
ing framework for boundary-constrained floor plan generation. common iterative manual design processes during the
Our framework leverages recent popular diffusion models training stage.
(DM) [10], which have shown remarkable success in various 3) Our framework can address a coarse-to-fine floor plan
generative tasks, to handle the non-trivial challenge of floor layout design problem, even for boundary-unconstrained
plan layout design. It incorporates a Transformer [29] as the floor plan generation. It produces bubble diagrams that
backbone of the model, enabling the handling of complex represent geometric and topological information of de-
vector layout data. Additionally, it integrates a conditioning sign elements in the early stages, and detailed floor plans
mechanism that allows for controlled floor plan generation in the later stages. Our approach supports various levels
under specific conditions. We have designed an alignment of user interaction, including fully automatic, coarsely
mechanism during the training phase of the model, which en- controllable, and finely controllable options. Compared
ables the model to extract information from the iterative design to previous methods, our approach can generate more
processes of designers, thereby better fitting the distribution of layout information and provide users with richer inter-
predicted targets. actions.
Moreover, our approach also supports generating bubble
diagrams and floor plans without predefined boundaries. Pre- II. R ELATED WORK
vious methods that do not support predefined boundaries have A. Bubble Diagram Generation
been significantly limited in terms of user controllability – they
either fully automate the generation process without allowing Bubble diagrams establish a connection between the intent
user intervention [28] or only support bubble diagram-driven and solution of designer, and they are widely used in data-
floor plan generation [21], [22], [25], neglecting the genera- driven layout planning tasks such as architectural design and
tion of bubble diagrams representing the early design stages. interior scene synthesis [11], [21], [22], [25]. In previous
Compared to previous methods, our new approach produces work, they often serve as an input condition representing
higher-quality results that are closer to the ground truths and user intent, or as an intermediate representation between input
supports various forms of user interaction. conditions and final layout prediction. Almost no work focuses
on data-driven bubble diagram generation. Recently, Sun et
Overall, our contributions are as follows:
al. [28] proposed a data-driven approach (BubbleFormer) for
1) We frame the task of boundary-constrained floor plan generating bubble diagrams to better drive downstream layout
generation as a generative task fully in vector space, planning tasks. BubbleFormer requires rasterizing the input
for the first time, and propose a generation framework boundaries and the target bubble diagrams into images, then
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 3

TABLE I
C OMPARED TO EXISTING METHODS FOR BOUNDARY- CONSTRAINED BUBBLE DIAGRAM AND FLOOR PLAN GENERATION , OUR APPROACH IS BASED ON
STATE - OF - THE - ART GENERATIVE MODELS , SPECIFICALLY DIFFUSION MODELS (DM), AND PERFORMS PREDICTIONS DIRECTLY IN VECTOR SPACE
WITHOUT ANY INTERMEDIATE RASTERIZATION OPERATIONS . O UR METHOD SUPPORTS THE MOST EXTENSIVE RANGE OF USER INTERACTIONS , FROM
FULLY AUTOMATIC GENERATION , WHERE A VECTOR LAYOUT CAN BE OBTAINED ONLY FROM THE INPUT BOUNDARY, TO COARSELY CONTROLLABLE
PREDICTION , WHERE USERS CAN PROVIDE SIMPLE INSTRUCTIONS TO SPECIFY ROOM NODES (Rnode ), ROOM ADJACENCIES (Radja ), AND ROOM
PARTITIONING (Rpart ), TO FINELY CONTROLLABLE DESIGN , WHERE USERS CAN ITERATIVELY ADJUST EACH STEP OF THE DESIGN PROCESS .

Bubble Diagram Network w/o Fully Coarsely Controllable Finely Controllable

Generation Architecture Rasterization Automatic Rnode Radja Rnode Radja
BubbleFormer [28] CNN & Transformer ✗ ✓ ✗ ✗ ✗ ✗
DiffPlanner [Our] DM ✓ ✓ ✓ ✓ ✓ ✓
Floor Plan Network w/o Fully Coarsely Controllable Finely Controllable
Generation Architecture Rasterization Automatic Rnode Radja Rpart Rnode Radja Rpart
ArchiGAN [4] GAN ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗
ActFloor-GAN [30] GAN ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗
WallPlan [27] CNN ✗ ✓ ✗ ✗ ✗ ✗ ✗ ✗
RPLAN [32] CNN ✗ ✓ ✓ ✗ ✗ ✓ ✗ ✗
Graph2Plan [11] CNN & GNN ✗ ✓ ✓ ✓ ✗ ✗ ✗ ✗
iPLAN [8] CNN & GAN ✗ ✓ ✓ ✗ ✓ ✓ ✗ ✓
DiffPlanner [Our] DM ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

using a CNN to extract features and subsequently performing floor plans not as raster images but as geometric vectors of
the prediction process. Finally, it needs to employ the Hungar- the bounding box coordinates [8], [11]. However, these ap-
ian matching algorithm [16] to obtain unique matches between proaches still required starting from a rasterized boundary and
the generated nodes and edges. generating a rasterized layout image as auxiliary information
While BubbleFormer can automatically generate diverse for the final layout prediction.
results for users to browse and select their preferred de- Unlike all previous methods, our approach defines the
signs, the entire generation process does not allow for user boundary-constrained floor plan generation problem as a
interaction. But interactivity is very useful for an AI tool vector-to-vector generation task, directly predicting the vector
focused on layout planning [26]. Unlike BubbleFormer, our floor plan from the vector boundary without any rasterization
method supports various forms of user interaction, as shown operations, as shown in Table I. This enables our model
in Table I. It allows users to simply input instructions, such to generate higher-quality and more reasonable floor plans.
as the number and categories of nodes, and also supports Additionally, our model supports more diverse user-controlled
more detailed interactions, such as iterative design of nodes generation options, enhancing interactivity compared to previ-
and adjacencies. Additionally, our method does not require ous methods. This makes our approach more aligned with the
redundant rasterization operations. All predictions are made user needs and preferences in the design process.
directly in vector space, and no post-processing is needed.
C. Floor Plan Generation without Boundary Constraints
Floor plan generation without boundary constraints is of-
B. Boundary-constrained Floor Plan Generation
ten driven by bubble diagrams. Bubble diagrams effectively
Floor plan generation has been a prominent topic in the capture the design intent of architects during the initial stages
intersection of computer graphics [3], [19], [31] and architec- of design. Therefore, some learning-based approaches have
ture [1], [20], [24] for decades, aiming to achieve automated been proposed to enable architectural floor plan generation
design of architectural floor plans through computational by mapping bubble diagrams to floor plans [21], [22], [25].
design methods. The exterior boundary of a building is a However, these methods typically focus only on the final
crucial condition for floor plan generation [32]. In recent years, design stage, converting already finalized design ideas into the
more and more researchers have started using learning-based final layout, while neglecting the early creative stages, such as
methods to tackle the problem of boundary-constrained floor the bubble diagram generation phase.
plan generation [4], [8], [11], [27], [30], [32]. Architects indeed need an automated tool to help convert
Early researchers directly employed image-based generative bubble diagrams, which convey design intent, into final floor
models, such as Pix2Pix [14], to predict a rasterized floor plans to reduce their workload. However, they also need an
plan image from a rasterized boundary image. However, the AI design tool that collaborates with them in the creative
generated results were often too noisy to be vectorized [4]. stage, providing more inspiration and reducing their workload
Subsequently, some researchers began dividing the floor plan in conceptual design [26]. Unlike previous methods, our
generation process into multiple sub-tasks [27], [30], [32]. proposed approach collaborates with users from the initial
Starting with a rasterized boundary, they used image-based creative stage to the final design stage, first creating bubble
models to first predict intermediate raster representations, diagrams and then generating the final floor plans. Compared
such as location masks, activity maps, or wall maps, and to the state-of-the-art method in bubble diagram-driven floor
then predicted rasterized floor plans from these intermediate plan generation, HouseDiffusion [25], our approach produces
representations, resulting in more reasonable outputs that could higher-quality results that are closer to the ground truths and
be vectorized. Next, some researchers attempt to represent supports richer user interactions.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 4

(a) Vector GT (b) DiffPlanner (c) Previous methods

III. OVERVIEW

Problem. The goal of boundary-constrained floor plan gen- { boundary: 8x80

"boundary": [[x, y], ... ],
eration is to generate a set of geometric and topological "entrance": [[x, y], ... ]
}
properties of rooms within the given boundary. A floor plan entrance: 8x8

can be considered as a collection of boundary, entrance, and

{ is- c s (x, y)
a set of rooms, including information on their categories, "boundary": [[x, y], ... ],
"entrance": [[x, y], ... ],
sizes, locations, and other details (see Figure 2(a)). Previous "nodes": [{
"id": int,
approaches often re-frame the floor plan generation task as "category": int,
"size": int,
an image generation task, which first requires rasterizing the "location": [x, y]}, ... ]
} nodes: 8x5
vector floor plans into images representing various layout
{
information (see Figure 2(c)). Subsequently, a series of image- "boundary": [[x, y], ... ],
0
0 1 2 3 4 5 6 7
"entrance": [[x, y], ... ], 1
based networks perform the prediction and generation tasks. "nodes": [{ 2
"id": int, 3
In contrast, we define this task directly as a vector-to-vector "category": int, 4
5
"size": int,
generation task (with data tensors as shown in Figure 2(b)). "location": [x, y]}, ... ] 6
7
"adjacencies": [[id, id], ... ] adjacencies: 8x8
Dataset. This work utilizes the RPLAN dataset [32], which }

{ "boundary": [[x, y], ... ],

is a large open-source dataset that contains 80K real-world "entrance": [[x, y], ... ],
(x0, y0)(x1, y1)

"nodes": [{
floor plans. We use the same data preprocessing method as "id": int,
"category": int,
in [11] to extract the necessary vector information, such as the "size": int,
"location": [x, y]}, ... ]
room numbers, room categories, room sizes, room locations, "adjacencies": [[id, id], ... ],
"partitioning": [[[x, y], ... ], ]} partitioning: 8x4
room adjacencies, and other relevant details (Figure 2(a)). We
-1 1 Living room Bedroom Kitchen
further process this information into a data tensor format that

-
Bathroom Balcony Storage
our vector-based model can accept (Figure 2(b)) to obtain the
final dataset including 80K pairs of data, and randomly split Fig. 2. Overview and data representation of DiffPlanner and previous
methods. Our DiffPlanner decomposes the floor plan design process into
it into 56K-12K-12K for training-validation-test sets. several stages based on established design principles, allowing for human
Challenge. Achieving our goal is non-trivial. Firstly, it input at various levels of detail. Unlike previous methods that use raster
is challenging for the model to directly extract information images (c), which require the rasterization of vector data, our approach
operates directly on data tensors representing the vector space (b). Starting
from vector data and learn layout planning capabilities. For with input vectors for the room boundary and entrance, our model first predicts
example, humans find it easier to capture initial layout in- room nodes including their number, categories, sizes, and rough locations. It
formation from images compared to structured vector data then determines the adjacency matrix between these nodes to form a bubble
diagram, representing the initial design phase. Finally, our model predicts the
(Figure 2(a) vs. Figure 2(c)). Secondly, we expect this model to room partitioning to complete the floor plan.
handle diverse layout planning tasks, such as generating bubble
diagrams in the early design stages, producing final floor
plans in the later stages, and even tackling layout planning as specifying the number and categories of rooms, iteratively
tasks without boundary constraints. Finally, we want this controlling the room locations and their adjacencies, and
model to deeply interact with users. It should not only be determining the partitioning of rooms.
fully automatic but also allow for coarse user control, where
users can provide simple instructions, and fine control, where
IV. M ETHOD
users can iteratively control the process. Overall, the model
is expected to output complex layouts from simple input In this section, we first provide a brief introduction to the
information, with the entire generation process being highly diffusion models (DM) and then explain how our method
controllable by the user and performed entirely in vector space. progressively predicts room nodes, room adjacencies, and
Methodology. Similar to existing methods [8], [32], we also room partitioning from the input boundary.
appropriately decompose the floor plan design procedure into
several stages based on fundamental design principles and
widely adopted practices [23], allowing for human input at A. Diffusion Models (DM)
various levels of detail. Note that, unlike previous methods, The diffusion model [10] is a generative framework that
our approach performs predictions entirely in vector space, gradually transforms a simple initial distribution into a com-
without the need to rasterize the vector information. As shown plex target distribution through a series of iterative steps. This
in Figure 2(b), starting from a data tensor storing the input model operates by adding noise to the data and then learning to
vector boundary, that includes the entrance, our model first reverse this process, effectively denoising to generate realistic
predicts a set of room nodes including the information about samples. It is particularly effective for capturing intricate
number, categories, sizes, and approximate locations. Next, patterns and dependencies in data, making it suitable for
it predicts the adjacency matrix between these nodes to then tasks that require fine-grained detail and structure. Specifically,
create a bubble diagram, reflecting the initial design stage. diffusion models (DM) transform Gaussian noise xT into a
Subsequently, our model further predicts the room partitioning data sample x0 through a series of T denoising steps, involving
to generate the final floor plan. We allow the user to interact both forward and reverse processes during training. In the
with the model throughout the entire generation process, such forward process q(xt |x0 ), a data sample x0 is progressively
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 5

converted into a noisy sample xt at each time step t by adding

Gaussian noise ϵ ∼ N (0, I):
√ p xtinter xt-1
inter

xt = ᾱt x0 + (1 − ᾱt )ϵ. (1)

xinter
T-1

where αt = (1 − βt ), βt isQthe variance schedule controlling xT-1 x1inter

t
the noise level, and ᾱt = s=1 αs represents the cumulative x1
xt-1
data preservation. The reverse process pθ (xt−1 |xt ) begins with
pure Gaussian noise xT ∼ N (0, I), and iteratively denoises it xt
xT
step by step until it reaches x0 . During this process, the model x0
xt-1
takes xt and estimates xt−1 by inferring ϵ [10]:
xt
1 1 − αt
xt−1 = √ (xt − √ ϵθ (xt , t)) + σt z. (2) Iterative design process Forward process Reverse process Alignment mechanism
αt 1 − ᾱt of designers q (xt | x0) p (xt-1 | xt, c ) xt (xt, ) = xtinter

where ϵθ is a function approximator and is optimized by: Boundary Entrance Number Category Partial input

Lt-1 = ||ϵθ (xt , t) − ϵ||2 . (3)

Fig. 3. Architecture of NodeDiff. NodeDiff is a Transformer-based conditional

B. Node Generation diffusion model that integrates an alignment mechanism during training, align-
Problem. In this task, we aim to predict vector room nodes ing the optimization trajectory of model with the iterative design processes
of designers. This enables our model to handle complex vector data, better fit
including the number, categories, sizes, and locations of rooms the distribution of the predicted targets, accomplish the challenging task of
from a given vector boundary with entrance. layout planning, and achieve diverse user-controllable generation.
Representation. The vector boundary and entrance are rep-
resented by a set of corner coordinates. Previous methods [4],
[8], [11], [27], [28], [30], [32] often rasterize the vector bound- number, room categories, and/or partial input, as shown in
ary and entrance into a set of binary raster masks and then use Figure 3. Given an initial point x0 , we obtain an approximate
an image-based network to extract features (Figure 2(c)). In point xt by sampling from the posterior distribution q(xt |x0 )
contrast, we directly feed the vector boundary and entrance (1), and run a forward step with the inverse process ϵθ (xt , t, c)
into our model, avoiding the redundant step of rasterizing the conditioning on the constraints c to infer the representation
boundary and entrance to extract their embeddings. Note that, xt−1 at time t − 1:
while the entrance has a fixed number of four corner points,
the number of boundary corner points is variable. To handle 1 1 − αt
xt−1 = √ (xt − √ ϵθ (xt , t, c)) + σt z. (4)
this, we employ a simple corner augmentation strategy for the αt 1 − ᾱt
boundary. Specifically, we add a corner point at the midpoint
where ϵθ is optimized by:
of the longest edge of the boundary iteratively until the number
of corner points reaches 40, which is sufficient to represent the
exterior boundary of almost all real-world buildings. Next, we Lt-1 = ||ϵθ (xt , t, c) − ϵ||2 . (5)
replicate the boundary and entrance eight times, corresponding Alignment. By rewriting (1), we can obtain x̃θt , an approx-
to the maximum number of rooms in the RPLAN dataset [32]. imation of x0 at each time t, using the predicted noise ϵθ as:
Finally, we convert the vector boundary and entrance into
tensors with dimensions 8 × 80 and 8 × 8, respectively. √ √
x̃θt = (xt − 1 − ᾱt ϵθ (xt , t, c))/ ᾱt . (6)
Previous methods typically represent vector room nodes as
a multi-channel binary image [8], [28], [32], using channel Some recent work [2] additionally introduces a reconstruction
indices to indicate categories and masks to represent the loss to encourage plausible predictions of x0 at each time step
sizes and locations of rooms, or as a three-channel color t by computing the Mean Squared Error (MSE) between the x̃θt
image [4], [30], using different colors to denote different and x0 . Inspired by this, we designed an alignment mechanism
categories (Figure 2(c)). These rasterized results are then for NodeDiff. Layout planning is often an iterative design
used as prediction targets. Unlike these previous methods, we process where designers gradually add elements to refine the
directly represent the vector room nodes as an 8 × 5 tensor in design. We hypothesize that aligning the sampling trajectory
continuous space. Each row represents information for a single of the diffusion model with the iterative design process of
room, and the five columns indicate is-room (1 for room, 0 a designer during training can encourage the model to more
for padding), category (1 to 6 representing the living room, closely mimic the experience of designers. This alignment
bedroom, kitchen, bathroom, balcony, and storage), size, x aims to enhance the ability of model to generate coherent and
and y coordinates of the location of the room. Ultimately, we practical design outcomes.
normalize all the data tensors to a range of [−1, 1]. Design is often carried out element by element, so the iter-
NodeDiff. We propose NodeDiff, a Transformer [29]-based ative design process can be defined as a process of gradually
conditional diffusion model for predicting room nodes based adding elements as the time step t decreases (T ∼ 0). An
on constraints such as the boundary with entrance, room intermediate result xinter
t in the iterative design process can
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 6

be represented by elements that have been confirmed by the mask [28], combined with a rasterized node mask, to indicate
designer and elements that remain unconfirmed in the mind room adjacencies (Figure 2(c)). In contrast, we adopt a more
of the designer. For a design target x0 with n elements to be intuitive and efficient approach by using an adjacency matrix
predicted, we can easily calculate the number of confirmed to represent the adjacencies between rooms. In this matrix,
nec = min(n(1 − t/T ) + 1, n) and unconfirmed elements a value of 1 indicates that two nodes are connected, while a
neunc = n−nec in the intermediate result xinter
t at time step t. value of 0 indicates that they are not. We then normalize the
Clearly, the confirmed elements ec should match the prediction matrix to a range of [−1, 1]. This new representation allows
target values, while the unconfirmed elements eunc should also the prediction of room adjacencies to be carried out entirely
possess a certain level of plausibility, as the designer may in vector space, eliminating the need for any rasterization
already have a rough idea of them, even if they have not been operations.
finalized. Thus, we first initialize xinter
t with x0 . Then we AdjacencyDiff. We propose AdjacencyDiff to predict room
randomly select neunc indices and set the elements with these adjacencies under the conditions of boundary, entrance, room
indices to a weighted combination of x0 and random noise nodes, and/or partial input. Its network architecture, training
xnoise ∼ N (0, I), xinter
t eunc = x0eunc k + xnoiseeunc (1 − k), strategy, and implementation details are the same as those of
where k = (1 + nec /n)/2. As the number of confirmed ele- NodeDiff.
ments nec increases, k will approach 1. Because, as the design Alignment. The design of room adjacencies is also an itera-
becomes progressively refined, the unconfirmed elements eunc tive process. Therefore, AdjacencyDiff incorporates an align-
will increasingly resemble the final design x0 . ment strategy in training, with slight differences from NodeD-
Now, at each time step t, we have a rough estimate x̃θt of x0 iff in obtaining intermediate results xinter
t . We first extract a
and an intermediate result xinter
t corresponding to the iterative list containing n room adjacencies to be predicted, where each
design process of the designer. Thus, the total loss is defined room adjacency is a tuple of room IDs. We then calculate the
as L = Lt-1 + λLalign : number of room adjacencies nec = min(n(1−t/T )+1, n) that
should be confirmed at time step t. Next, we randomly select
nec adjacencies as confirmed adjacencies, while the remaining
L = ||ϵθ (xt , t, c) − ϵ||2 + λ||x̃θt (xt , ϵθ ) − xinter
t ||2 . (7)
neunc = n − nec adjacencies are considered unconfirmed.
where Lt-1 is the conditional DDPM [10] loss for time step Similar to NodeDiff, we first initialize xinter
t as x0 . Then, for
t, and Lalign is the alignment loss for time step t. This align- the indices represented by the unconfirmed room adjacencies
ment mechanism effectively guides the sampling trajectory at eunc , we assign a weighted sum of x0 and random noise
each step to match the iterative design process, distilling the xnoise ∼ N (0, I), xinter
t eunc = x0eunc k + xnoiseeunc (1 − k),
optimizer throughout the reverse process in training. where k = (1 + nec /n)/2.
User-controllable generation. Our NodeDiff is not only User-controllable generation. Similar to NodeDiff, our Ad-
fully automated but also easily supports user-controllable gen- jacencyDiff model also supports generating room adjacencies
eration. By simply modifying the conditions input to the model under various conditions, such as specifying the boundary or
in (5), various forms of user control can be achieved. NodeDiff not, and providing partial input. This can be easily achieved
supports a range of user inputs, such as specifying whether or by simply adding or removing conditions in (5).
not to include the boundary and entrance, providing the desired
number of rooms and their categories, and incorporating partial D. Partitioning Generation
input. This flexibility allows users to guide the generation Problem. In this task, our goal is to predict room parti-
process according to their specific needs and preferences. tioning from the input boundary with entrance, as well as the
Implementation details. We implemented our NodeDiff us- room nodes generated by NodeDiff and the room adjacencies
ing PyTorch, based on the public implementation of Guided- predicted by AdjacencyDiff. This will generate a floor plan
Diffusion [6]. The optimizer used is Adam [15] with decoupled that adapts to the final stage of the design process.
weight decay [18] for 300k steps with a batch size of 1024 on Representation. Unlike some previous methods that repre-
an NVIDIA RTX 2080 Ti GPU. The initial learning rate was sent room partitioning as a set of multi-channel binary raster
set to 1e-3 and was reduced by a factor of 10 after every masks [27], [32] or three-channel color images [4], [30], we
100k steps. We set the number of diffusion steps to 1000 follow the approach in [8], [11] by using the coordinates of
and uniformly sampled t during training. Please refer to the the top-left and bottom-right corners of room bounding boxes
supplementary materials for more details. to represent room layouts. This allows our model to perform
the prediction process entirely in vector space.
C. Adjacency Generation PartitioningDiff & Alignment. We propose PartitioningDiff
Problem. In this task, our goal is to predict room adjacencies to predict room partitioning under the conditions of boundary,
from the input boundary with the entrance, as well as the room entrance, nodes, adjacencies, and/or partial input. Its network
nodes generated by NodeDiff. The room adjacencies reflect architecture, training strategy, alignment mechanism, and im-
the relationships between rooms, forming an essential part of plementation details are the same as those of NodeDiff.
the bubble diagram and serving as a crucial condition for the User-controllable generation. By simply adding or remov-
subsequent floor plan design. ing conditions in (5), our PartitioningDiff also supports gen-
Representation. Previous methods represent room adjacen- erating room partitioning under various conditions, such as
cies as specific line segments in raster space to create a connect specifying the boundary or not, and providing partial input.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 7

TABLE II TABLE III

Q UANTITATIVE COMPARISON OF FID SCORES FOR BUBBLE DIAGRAM Q UANTITATIVE COMPARISON OF FID SCORES FOR FLOOR PLAN
GENERATION BETWEEN OUR METHOD AND ITS VARIANTS SUPPORTING GENERATION BETWEEN OUR METHOD (D IFF P LANNER ),
DIFFERENT TYPES OF USER INPUTS (D IFF P LANNER ), AND THE STATE - OF - THE - ART METHODS (A RCHI GAN, RPLAN, ACT F LOOR -GAN,
STATE - OF - THE - ART METHOD (B UBBLE F ORMER ). WALL P LAN , I PLAN, G RAPH 2P LAN , AND H OUSE D IFFUSION ), AND THEIR
VARIOUS VARIANTS UNDER DIFFERENT CONDITIONS .
Bubble Diagram Input Condition Output FID
Generation B Rn Rc Rs Rl w/ Proc Score Floor Plan Input Condition Output FID
BubbleFormer ✓ ✗ ✗ ✗ ✗ ✓ 14.90 Generation B Rn Rc Rs Rl Ra w/ Proc Score
DiffPlannerbI ✓ ✗ ✗ ✗ ✗ ✗ 1.59 ArchiGAN ✓ ✗ ✗ ✗ ✗ ✗ ✗ 117.59
RPLAN ✓ ✗ ✗ ✗ ✗ ✗ ✓ 4.64
DiffPlannerbII ✓ ✓ ✗ ✗ ✗ ✗ 1.39 ActFloor-GAN ✓ ✗ ✗ ✗ ✗ ✗ ✓ 4.54
DiffPlannerbIII ✓ ✓ ✓ ✗ ✗ ✗ 1.19 WallPlan ✓ ✗ ✗ ✗ ✗ ✗ ✓ 2.55
DiffPlannerbIV ✓ ✓ ✓ ✓ ✓ ✗ 0.05 iPLANI ✓ ✗ ✗ ✗ ✗ ✗ ✓ 5.38
DiffPlannerI ✗ ✗ ✗ ✗ ✗ ✗ 0.66 iPLANII ✓ ✓ ✓ ✗ ✗ ✗ ✓ 5.21
DiffPlannerII ✗ ✓ ✗ ✗ ✗ ✗ 0.64 iPLANIII ✓ ✓ ✓ ✗ ✓ ✗ ✓ 2.13
DiffPlannerIII ✗ ✓ ✓ ✗ ✗ ✗ 0.59 Graph2PlanI ✓ ✗ ✗ ✗ ✗ ✗ ✓ 2.03
DiffPlannerIV ✗ ✓ ✓ ✓ ✓ ✗ 0.05 Graph2PlanII ✓ ✓ ✓ ✓ ✓ ✓ ✓ 0.50
Graph2PlanIII ✓ ✓ ✓ ✓ ✓ ✓ ✗ 4.84
DiffPlannerbI ✓ ✗ ✗ ✗ ✗ ✗ ✓ 1.08
DiffPlannerbII ✓ ✓ ✗ ✗ ✗ ✗ ✓ 0.93
V. E VALUATION AND D ISCUSSION DiffPlannerbIII ✓ ✓ ✓ ✗ ✗ ✗ ✓ 0.72
DiffPlannerbIV ✓ ✓ ✓ ✓ ✓ ✗ ✓ 0.26
We conduct quantitative comparisons, qualitative evalua- DiffPlannerbV ✓ ✓ ✓ ✓ ✓ ✓ ✓ 0.25
tions, ablation experiments, and perceptual studies to thor- DiffPlannerbVI ✓ ✓ ✓ ✓ ✓ ✓ ✗ 0.86
HouseDiffusion ✗ ✓ ✓ ✗ ✗ ✓ ✗ 29.56
oughly evaluate our method, concluding with a detailed dis- DiffPlannerI ✗ ✗ ✗ ✗ ✗ ✗ ✓ 3.38
cussion at the end. DiffPlannerII ✗ ✓ ✗ ✗ ✗ ✗ ✓ 3.25
DiffPlannerIII ✗ ✓ ✓ ✗ ✗ ✗ ✓ 3.49
DiffPlannerIV ✗ ✓ ✓ ✓ ✓ ✗ ✓ 2.89
A. Quantitative Comparisons DiffPlannerV ✗ ✓ ✓ ✓ ✓ ✓ ✓ 2.86

1) Method variants: To comprehensively evaluate the per-

formance of our DiffPlanner in generating bubble diagrams Bubble diagram generation. We use FID scores to compare
and floor plans under various input conditions, we define our method and its variants supporting different types of user
multiple variants based on different input settings: inputs (DiffPlannerbI, . . . , IV ), with the most recent state-of-the-
b
• DiffPlannerI : Generates bubble diagrams and floor plans art method [28] (BubbleFormer). Table II shows that our
using only the input boundary with entrance (B), without method (DiffPlannerbI, . . . , IV ) significantly outperforms Bubble-
requiring any additional conditions. Former in generating bubble diagrams under boundary con-
b b
• DiffPlannerII : Extends DiffPlannerI by incorporating the straints. Moreover, unlike BubbleFormer, which only supports
number of rooms (Rn ) as an additional input condition. fully automated bubble diagram generation without any user
b b
• DiffPlannerIII : Further extends DiffPlannerII by including input, our method allows for user control over a range of
room categories (Rc ) as an additional input condition. conditions, including the number (Rn ), categories (Rc ), sizes
b b
• DiffPlannerIV : Builds upon DiffPlannerIII by adding room (Rs ), locations (Rl ) of room nodes. Note that, BubbleFormer
sizes (Rs ) and locations (Rl ) as input conditions. needs to employ the Hungarian matching algorithm [16] to
b b
• DiffPlannerV : Expands DiffPlannerIV by additionally in- obtain unique matches between the generated nodes and edges.
corporating room adjacencies (Ra ) as an input condition, While our method does not require any post-processing (Proc)
enabling the model to predict floor plans from given operations.
bubble diagrams. Our method also supports user-controllable bubble diagram
For the boundary-unconstrained generation of bubble dia- generation without boundary constraints (DiffPlannerI, . . . , IV ),
grams and floor plans, we define the variants (DiffPlannerI , and as more user inputs are incorporated, the performance
. . . , DiffPlannerV ) with the same setup as above, except that of the model improves, achieving lower FID scores. The low
the input conditions do not include the boundary with entrance FID scores in Table II clearly demonstrate the ability of our
(B). method to accommodate various user conditions and generate
2) FID comparison: A quantitative evaluation of generated high-quality bubble diagrams that closely resemble the ground
bubble diagrams and floor plans is challenging. Nevertheless, truths, even without boundary constraints.
we compare our approach to state-of-the-art techniques using Floor plan generation. We conduct a quantitative compari-
the Fréchet Inception Distance (FID) score [9], as most state- son of FID scores between our method (DiffPlanner), the state-
of-the-art methods [8], [25], [27], [28] utilized for evaluation. of-the-art methods (ArchiGAN [4], RPLAN [32], ActFloor-
The FID score is a global metric to calculate the distribution GAN [30], WallPlan [27], iPLAN [8], Graph2Plan [11],
similarity between real data and generated data. The lower the & HouseDiffusion [25]), and their various variants under
FID score, the more similar the generated data to the ground different conditions including boundary with entrance (B),
truths. It is important to note that the FID score is highly number (Rn ), categories (Rc ), sizes (Rs ), locations (Rl ), and
correlated with the number of examples. We evaluate each adjacencies (Ra ) of rooms, as shown in Table III.
method with different constraints on a test dataset containing For the boundary-constrained floor plan generation task, our
12K examples and calculate the respective FID scores. method supports the most user conditions and achieves the
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 8

TABLE IV TABLE V
S TATISTICS COMPARISON . E ACH STATISTIC IS THE RATIO CALCULATED Q UANTITATIVE EVALUATION OF HOW WELL OUR D IFF P LANNER ADHERES
BASED ON THE GROUND TRUTH BUBBLE DIAGRAMS AND FLOOR PLANS . A TO THE INPUT CONDITIONS UNDER DIFFERENT CONTROL MODES FOR
RATIO CLOSE TO 1 SUGGESTS THAT THE GEOMETRY AND TOPOLOGY OF BUBBLE DIAGRAM AND FLOOR PLAN GENERATION TASKS . A LOWER
THE MODEL - GENERATED RESULTS ARE MORE SIMILAR TO THE GROUND MAEavg INDICATES STRICTER ADHERENCE TO THE INPUT CONDITIONS .
TRUTHS .
Bubble Diagram Input Condition Output MAEavg
Bubble Diagram Statistics Generation B Rn,c,s,l w/ Proc Rn Rc Rs,l
n l r Ln La
Generation Ravg Cavg Cavg avg avg DiffPlannerbII ✓ Rn ✗ 0.0 - -
BubbleFormer 0.932 0.376 0.369 0.969 1.074 DiffPlannerbIII ✓ Rn,c ✗ 0.0 0.0 -
DiffPlannerbI 0.996 1.000 1.004 0.998 1.004 DiffPlannerbIV ✓ Rn,c,s,l ✗ 0.0 0.0 0.0
Floor Plan Statistics Floor Plan Input Condition Output MAEavg
Generation n
Ravg l
Cavg r
Cavg Ln La Generation B Rn,c,s,l,a w/ Proc Rn Rc Rs,l Ra
avg avg
DiffPlannerbII ✓ Rn ✓ 0.0 - - -
RPLAN 0.869 0.851 0.997 0.958 1.045
DiffPlannerbIII ✓ Rn,c ✓ 0.0 0.0 - -
ActFloor-GAN 0.904 0.864 0.965 0.935 0.902
DiffPlannerbIV ✓ Rn,c,s,l ✓ 0.0 0.0 0.008 -
WallPlan 0.998 0.968 0.973 0.970 0.911
DiffPlannerbV ✓ Rn,c,s,l,a ✓ 0.0 0.0 0.008 0.025
iPLANI 0.938 0.980 1.049 0.969 1.131
DiffPlannerbVI ✓ Rn,c,s,l,a ✗ 0.0 0.0 0.001 0.007
Graph2PlanI 0.980 0.971 0.988 0.969 1.029
DiffPlannerbI 0.996 0.997 1.001 0.998 0.971

dataset containing 12K examples and calculate its ratio with

lowest FID scores, as shown in Table III. The ArchiGAN, the corresponding data from the ground truths. A ratio close
based on Pix2Pix [14], directly predicts the raster floor plan to 1 suggests that the geometry and topology of the model-
from the input raster boundary, resulting in excessively noisy generated results are more similar to the ground truths.
outputs that cannot be vectorized, thus obtaining high FID We conduct a statistics comparison between our
score. The RPLAN, ActFloor-GAN, and WallPlan use location method (DiffPlannerbI ) and the state-of-the-art methods
masks, activity maps, and wall maps as intermediate prediction (BubbleFormer [28], RPLAN [32], ActFloor-GAN [30],
targets before further predicting the floor plans, resulting WallPlan [27], iPLANI [8], & Graph2PlanI [11]), focusing on
in comparatively lower and better FID scores. Similar to the tasks of generating bubble diagrams and floor plans with
our approach, iPLAN decomposes the floor plan generation boundaries as the only condition. This comparison is intended
process, using a set of node masks that includes room number, to evaluate the geometric and topological characteristics of
categories, and locations as the intermediate representation. the model-generated results.
This allows their models to support more user conditions. Both The results of the comparison are presented in Table IV.
Graph2Plan and our method employ a boundary and bubble Compared to the state-of-the-art method in bubble diagram
diagram-driven layout generation strategy. We also employ a generation, BubbleFormer, our method surpasses it in all five
post-processing (Proc) step as in Graph2Plan to eliminate gaps metrics, particularly excelling in the two statistics related
that may exist between the predicted rooms. Eventually, our to room adjacencies (C l & C r ). In the task of floor plan
method consistently achieves lower FID scores, both with and generation, our method achieves the best performance in four
without post-processing, surpassing Graph2Plan. out of five metrics compared to the existing methods (C l ,
For the task of floor plan generation without boundary con- C r , Ln & La ). The results thus indicate that our method can
straints, previous methods [21], [22], [25] often support only a generate outputs that most closely resemble the ground truth
single user condition, typically a bubble diagram-constrained geometry and topology, surpassing the existing state-of-the-art
floor plan generation. Compared to the previous state-of-the- methods.
art method, HouseDiffusion [25], our method supports a wider 4) Assessing controllability compliance: To effectively
range of user inputs and achieves significantly lower FID demonstrate the controllability of our method, we further
scores. This indicates that our method is more flexible and examine whether the generated results adhere well to the spec-
capable of generating higher-quality floor plans that closely ified input conditions, as shown in Table V. We first extract key
resemble the ground truths. attributes from the vector outputs of our DiffPlanner, including
3) Statistics comparsion: FID is a metric that assesses the the number (Rn ), categories (Rc ), sizes, locations (Rs,l ), and
similarity in overall distribution between generated results and adjacencies (Ra ) of rooms. Specifically, the Rn is represented
real data. However, it fails to specifically account for the as an integer, while the Rc , Rs,l , and Ra are represented as
intricate geometric and topological details in the quality of 8×1, 8×3, & 8×8 tensors, respectively, similar to the tensors
generated bubble diagrams and floor plans. Similar to [27], of nodes and adjacencies illustrated in Figure 2(b). Then, for
[28], we have gathered various statistical metrics to evaluate each input condition (Rn , Rc , Rs,l , & Ra ), we compute the
the geometry and topology of these generated bubble diagrams Mean Absolute Error (MAE) between the generated result and
and floor plans. the input for each sample in the test dataset, which contains
These statistics include the total number of rooms (Rn ), the 12K examples, and report the average (MAEavg ). A lower
number of rooms directly connected to the living room (C l ), MAEavg indicates that the model adheres more strictly to the
the ratio of rooms connected to the living room to all non- input conditions.
living rooms (C r ), the total number of living rooms (Ln ), and Table V shows that our DiffPlanner achieves consistently
the percentage of area occupied by the living room (La ). We low MAEavg across various control modes for boundary-
compute the average for each statistical metric across the test constrained bubble diagram and floor plan generation tasks.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 9

TABLE VI TABLE VII

Q UANTITATIVE EVALUATION OF THE OUTPUT DIVERSITY OF OUR METHOD Q UANTITATIVE EVALUATION OF THE ABILITY OF OUR THREE CORE
b
(D IFF P LANNER I ) ON THE AREAS OF SIX ROOM CATEGORIES , INCLUDING COMPONENTS (N ODE D IFF , A DJACENCY D IFF , AND PARTITIONING D IFF )
THE LIVING ROOM (Rliv ), BEDROOM (Rbed ), KITCHEN (Rkit ), TO PREDICT THE COMPLETE TARGET FROM PARTIAL INPUTS UNDER
BATHROOM (Rbat ), BALCONY (Rbal ), AND STORAGE (Rsto ). T HE DIFFERENT CONDITIONS . T HE TABLE SHOWS FID SCORES FOR EACH
D IVERSITYavg AND C OVERAGEGT avg RANGE FROM 0 TO 1, WHERE LOWER COMPONENT WITH VARYING PROPORTIONS OF TARGET INFORMATION IN
SCORES INDICATE HIGHER OUTPUT DIVERSITY. THE PARTIAL INPUTS : 25%, 50%, AND 75%.

Diversityavg Input Condition FID

Rliv Rbed Rkit Rbat Rbal Rsto B Rn Rc Rs Rl Ra Partial Score
DiffPlannerbI 0.57 0.53 0.26 0.19 0.34 0.00 NodeDiffI ✓ ✗ ✗ ✗ ✗ ✗ 25% 1.86
CoverageGT NodeDiffI ✓ ✗ ✗ ✗ ✗ ✗ 50% 0.76
avg
Rliv Rbed Rkit Rbat Rbal Rsto NodeDiffI ✓ ✗ ✗ ✗ ✗ ✗ 75% 0.16
NodeDiffII ✓ ✓ ✗ ✗ ✗ ✗ 25% 1.65
DiffPlannerbI 0.56 0.52 0.24 0.17 0.34 0.00
NodeDiffII ✓ ✓ ✗ ✗ ✗ ✗ 50% 0.65
NodeDiffII ✓ ✓ ✗ ✗ ✗ ✗ 75% 0.13
NodeDiffIII ✓ ✓ ✓ ✗ ✗ ✗ 25% 1.13
This demonstrates that our method effectively adheres to the NodeDiffIII ✓ ✓ ✓ ✗ ✗ ✗ 50% 0.46
NodeDiffIII ✓ ✓ ✓ ✗ ✗ ✗ 75% 0.11
input conditions, validating the effectiveness of its controllabil- AdjacencyDiff ✓ ✓ ✓ ✓ ✓ ✗ 25% 0.05
ity. Specifically, for input conditions Rn and Rc , our DiffPlan- AdjacencyDiff ✓ ✓ ✓ ✓ ✓ ✗ 50% 0.05
ner does not modify them during the prediction process; they AdjacencyDiff ✓ ✓ ✓ ✓ ✓ ✗ 75% 0.04
PartitioningDiff ✓ ✓ ✓ ✓ ✓ ✓ 25% 0.76
are directly preserved as part of the final output. For the input PartitioningDiff ✓ ✓ ✓ ✓ ✓ ✓ 50% 0.31
conditions Rs,l , they remains unchanged when predicting the PartitioningDiff ✓ ✓ ✓ ✓ ✓ ✓ 75% 0.07
bubble diagram and are directly carried over to the final output.
However, when predicting the floor plan, the final output
consists of room boxes inferred by the model based on the room categories. This can be expected, as the living room
input conditions and further refined through a post-processing and bedroom typically occupy the largest portions of indoor
(Proc) step. This introduces slight variations in the recalculated space in residential buildings. Additionally, their approximate
Rs,l and Ra , as they are derived from the predicted room locations are often constrained by the input boundary and
boxes, compared to the input Rs,l and Ra . To further analyze entrance, leading to a certain degree of similarity across
this effect, we conduct an ablation study by removing the post- different layout variations. However, since the scores lie in the
processing step (DiffPlannerbV vs. DiffPlannerbVI ). The results mid-range between 0 and 1, this already represents a strong
show that the slight variations in Rs,l and Ra are primarily diversity performance. Similarly, compared to the kitchen
caused by the post-processing module, while our model itself and bathroom, the balcony has more spatial constraints, as
still adheres well to the input conditions. it is often positioned along specific protruding edges of the
5) Output diversity: We further quantitatively evaluate the boundary. As a result, it exhibits a higher score than the
diversity of the results generated by our model to assess kitchen and bathroom.
whether it can produce diverse floor plans given the same input We further analyze the differences between the generated
condition. First, for each sample (s) in the test dataset con- result by our DiffPlannerbI (Our) and its corresponding ground
taining 12K examples (S), we let DiffPlannerbI generate five truth (GT) for each sample (s) in the test dataset (S), using
floor plan variants (Vi=1,...,K,K=5 ) using only the boundary as the following metric:
input condition. Then, to measure the similarity between two
variants Vi and Vj , we compute the Intersection over Union CoverageGT
avg = [Es∈S IoUs,r (Our, GT )]r∈R . (9)
(IoU) separately for the total area of each (r) of the six room
categories (R), including the living room (Rliv ), bedroom The CoverageGT avg ranges from 0 to 1, where a lower value
(Rbed ), kitchen (Rkit ), bathroom (Rbat ), balcony (Rbal ), and indicates lower similarity and greater divergence. Table VI
storage (Rsto ), between the corresponding variants (Vi & Vj ). shows that our method achieves a relatively low score on
For each sample (s), we compute the mean IoU across all CoverageGTavg , indicating that for each test sample, our model
pairwise comparisons among the five variants. Finally, we take does not simply generate results that closely approximate the
the average of these values over the entire test dataset to obtain GT. An ideal AI design tool should capture the diverse design
the following metric: space that maps a single input to multiple possible outputs,
rather than merely learning a deterministic mapping from a
single input to a single output. Given a boundary as input,
Diversityavg = [Es∈S E1≤i<j≤K IoUs,r (Vi , Vj )]r∈R . (8)
the corresponding floor plan design space should be diverse,
The Diversityavg ranges from 0 to 1. A lower value indicates rather than being limited to the specific GT solution.
a smaller IoU between the generated results, implying greater The low FID score in Table III suggests that the results
output diversity of the model. As shown in Table VI, our generated by our model are globally aligned with the data
method performs well on Diversityavg , achieving relatively distribution of GT. Meanwhile, the relatively low CoverageGTavg
low scores across different room categories. This indicates score, which measures the similarity between the generated
that our method generates outputs with high diversity. For result and its corresponding GT for each individual sample,
the living room and bedroom, the scores are 0.57 and 0.53, further demonstrates that our method does not simply memo-
respectively, which are slightly higher than those for other rize the input-to-single-GT mapping as a fixed solution during
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 10

BubbleFormer DiffPlannerI
b b
DiffPlannerII
b
DiffPlannerIII
b
DiffPlannerIV GT DiffPlannerI DiffPlannerII DiffPlannerIII DiffPlannerIV GT

Fig. 4. Qualitative comparison between our method based on various con-

ditions (DiffPlannerbI, . . . , IV ) and the state-of-the-art method (BubbleFormer)
for boundary-constrained bubble diagram generation. While BubbleFormer
can generate bubble diagrams fully automatically, it does not support user
interaction. In contrast, our method not only supports fully automatic gen-
eration (DiffPlannerbI ) but also allows users to apply various controls, such
as the number (DiffPlannerbII ), categories (DiffPlannerbIII ), sizes & locations
(DiffPlannerbIV ) of nodes. Fig. 5. Our method can generate bubble diagrams fully automati-
cally (DiffPlannerI ) or based on various user inputs, such as the num-
ber (DiffPlannerII ), categories (DiffPlannerIII ), and sizes & locations
(DiffPlannerIV ) of nodes, without boundary constraints.
training. Instead, it effectively captures the diverse design
space from input to output. The low Diversityavg score in
Table VI further validates this observation. not adjacent to any other rooms, bedrooms not connected to the
6) Partial input generation: We further evaluate the ability living room, and unreasonable kitchen placements, as shown
of our three core components (NodeDiff, AdjacencyDiff, & in Figure 4.
PartitioningDiff) to predict the complete target from partial Compared to BubbleFormer, our fully automated method
inputs based on different input conditions including the bound- (DiffPlannerbI ) performs excellently, predicting high-quality
ary with entrance (B), number (Rn ), categories (Rc ), sizes and reasonable bubble diagrams without any rasterization
(Rs ), locations (Rl ), and adjacencies (Ra ) of rooms. We set operations, as all predictions are made in vector space. Our
the proportion of target information in the partial inputs to method (DiffPlannerbI, . . . , IV ) also supports various user input
three values: 25%, 50%, and 75%. conditions, and as more user inputs are provided, our results
As shown in Table VII, even with partial inputs containing increasingly approximate the ground truths (GT).
only 25% of the target information, all three core components Figure 5 demonstrates the ability of our method to generate
achieve relatively low FID scores. Furthermore, as the pro- bubble diagrams fully automatically (DiffPlannerI ) or based on
portion of target information increases, the FID scores further various user inputs, such as the number (DiffPlannerII ), cate-
decrease. These experimental results demonstrate the ability gories (DiffPlannerIII ), and sizes & locations (DiffPlannerIV )
of our method to predict complete layouts from partial inputs, of nodes, without boundary constraints. This shows that
indicating strong support for user interaction. our method is capable of driving downstream boundary-
unconstrained layout planning tasks, allowing for the control-
B. Qualitative Evaluations lable generation of bubble diagrams in the early stages of the
1) Bubble diagram generation: Figure 4 presents the qual- design process.
itative comparison results of our method based on various 2) Floor plan generation: Figure 6 shows the qualita-
conditions and the state-of-the-art method BubbleFormer for tive comparison results between our method (DiffPlannerbI )
boundary-constrained bubble diagram generation. Please note and previous methods (Graph2PlanI [11], iPLANI [8],
that while BubbleFormer can generate bubble diagrams fully WallPlan [27], ActFloor-GAN [30], RPLAN [32], and Archi-
automatically, it does not support user interaction. In contrast, GAN [4]) for floor plan generation from boundary only. The
our method not only supports fully automatic generation door and window placements for all methods used the rule-
(DiffPlannerbI ) but also allows users to apply various controls, based algorithm in RPLAN [32] and Graph2Plan [11].
such as the number (DiffPlannerbII ), categories (DiffPlannerbIII ), ArchiGAN directly predicts rasterized floor plans from
and sizes & locations (DiffPlannerbIV ) of nodes. rasterized boundaries, resulting in noisy outputs that cannot
BubbleFormer is a VAE-based generative model that first be vectorized. Furthermore, methods like RPLAN, ActFloor-
rasterizes room nodes and adjacencies into image space, GAN, and WallPlan start with rasterized boundaries and first
extracts feature from the rasterized data, and then performs predict intermediate information, such as location masks,
subsequent predictions. This process often leads to errors activity maps, or wall maps, which contain the topological
in bubble diagram generation, especially in predicting room and geometric information of the layouts. They then perform
adjacencies. Common issues include bathrooms and balconies subsequent predictions, enabling them to generate floor plans
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 11

b
GT DiffPlannerI Graph2PlanI iPLANI WallPlan ActFloor-GAN RPLAN ArchiGAN

Fig. 6. Qualitative comparison of floor plan generation from boundary only between our method (DiffPlannerbI ) and previous state-of-the-art methods
(Graph2PlanI , iPLANI , WallPlan, ActFloor-GAN, RPLAN, and ArchiGAN). The unreasonable design is highlighted in the red box. Our method produces
floor plans closest to the ground truths (GT), demonstrating the effectiveness of our vector-based approach in generating accurate and high-quality floor plans.

b
GT DiffPlannerI in vector space, taking vectorized boundaries as input and
progressively generating the final high-quality vector floor
plans. As shown in Figure 6, our method produces floor
plans that are closest to the ground truths (GT), while other
methods exhibit varying degrees of unreasonable results. This
demonstrates the effectiveness of our vector-based approach
in generating accurate and high-quality floor plans.
Figure 7 qualitatively demonstrates that our method
(DiffPlannerbI ) can generate multiple different high-quality
floor plans from a single input boundary. This indicates that
our model successfully captures the diverse design space in the
floor plan generation task, mapping a single input to multiple
possible outputs. Within the diverse outputs of our model,
some results may closely resemble the corresponding ground
truths (GT), as seen in the samples shown in the second and
third rows of Figure 7. This is expected, as the GT itself is one
feasible solution within this diverse design space. For a given
input boundary, we do not expect the outputs of the model to
always match the corresponding specific GT. Instead, we aim
for the model to generate multiple reasonable and high-quality
floor plans, reflecting the inherent diversity of architectural
design solutions.
Our method can also be applied to another prominent task:
bubble diagram-driven floor plan generation. We conducted
a qualitative comparison with the state-of-the-art method,
Fig. 7. Qualitative evaluation of output diversity. Our method (DiffPlannerbI ) HouseDiffusion [25]. Although HouseDiffusion can predict
can generate multiple high-quality floor plans from a single input boundary. floor plans based on the number (Rn ), categories (Rc ), and
adjacencies (Ra ) of rooms provided in a bubble diagram,
it still tends to produce some unreasonable results, such as
that can be vectorized. Graph2Plan and iPLAN take rasterized missing rooms, illogical adjacencies, and noisy boundaries, as
boundaries as input but represent their prediction targets as shown in Figure 8. In contrast, our DiffPlannerIII can generate
vector bounding box coordinates, although they still require reasonable layout results even when only the number (Rn )
the model to predict a rasterized layout image as an auxiliary. and categories (Rc ) of rooms are provided as inputs. When
Unlike all previous methods, our approach operates entirely additional information such as the sizes (Rs ), locations (Rl ),
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 12

Input conditions DiffPlannerIII HouseDiffusion DiffPlannerV 25% 50% 75% GT

GT
Rn,c,s,l,a with Rn,c with Rn,c,a with Rn,c,s,l,a

NodeDiffI

NodeDiffII

NodeDiffIII

Adjacency
Diff

Partitioning
Diff
Fig. 8. Qualitative comparison of bubble diagram-driven floor plan generation
between our method and HouseDiffusion. HouseDiffusion often produces
unreasonable results like missing rooms and noisy boundaries. In contrast, Fig. 9. Our three components (NodeDiff, AdjacencyDiff, & PartitioningDiff)
our method generates realistic layouts with just the number and categories can predict complete layouts from partial inputs with varying proportions
of rooms (DiffPlannerIII ), and closely matches the ground truths (GT) when (25%, 50%, & 75%), enabling detailed user interaction and fine-grained
additional information is provided (DiffPlannerV ). iterative design. Each component produces high-quality results that closely
match the ground truths (GT) from the partial inputs. Elements with red
borders indicate user-specified partial inputs.
and adjacencies (Ra ) of rooms is added, our DiffPlannerV
produces results that are almost identical to the ground truths
(GT). This demonstrates the robustness and accuracy of our AdjacencyDiff, & PartitioningDiff) to demonstrate its effec-
approach in generating high-quality and realistic floor plans tiveness both qualitatively and quantitatively.
from bubble diagrams. 1) Qualitative comparison: Figure 10 presents visual com-
3) Iterative design with partial input: Real-world floor plan parisons of the intermediate outputs produced at various dif-
layout planning is an iterative design process, making it crucial fusion steps by our three core components (NodeDiffI , Adja-
for a data-driven layout planning tool to support user-driven cencyDiff, & PartitioningDiff) and their variants without the
iterative design. Unlike previous methods that decompose the alignment mechanism (NodeDiffI w/o Align, AdjacencyDiff
generation process into several sub-tasks, allowing users to w/o Align, & PartitioningDiff w/o Align). The input conditions
modify intermediate results for rough control, our goal is used by these components and their variants are listed in
to enable iterative design at each sub-task. This ultimately Table VIII. The comparison results indicate that the alignment
allows users to have fine-grained control over the entire design mechanism allows our model to more quickly and effectively
process, from initial concept to finalization, for every element approximate the ground truth distribution.
in the layout. The key to achieving this goal is equipping the For the room node generation task, NodeDiffI with the
model with the ability to predict the complete layout from alignment mechanism can already generate a preliminary
partial user inputs at each sub-task. result at step 50 that reflects the geometry and topology of
Figure 9 demonstrates the ability of our three components the final output. By step 20, it essentially produces results
(NodeDiff, AdjacencyDiff, & PartitioningDiff) to predict com- consistent with the final output and continues to refine the
plete layouts from partial inputs with varying proportions output in subsequent diffusion steps, finally producing geo-
(25%, 50%, & 75%). The input conditions used by each metrically and topologically coherent room nodes at step 0. In
component are listed in Table VII. Each component performs contrast, NodeDiffI w/o Align only starts to produce results
exceptionally well, predicting high-quality results that closely preliminarily consistent with the geometry and topology of
match the ground truths (GT) from the partial inputs. This the final output at step 20, achieving a result similar to the
indicates that our approach enables detailed user interaction final output at step 0 only by step 10. Comparing the results
with the model throughout the design process, supporting fine- at step 0, NodeDiffI with the alignment mechanism clearly
grained iterative design. generates room nodes with better geometry and topology,
while NodeDiffI w/o Align produces room nodes close to or
C. Ablation Experiments even on the boundary, which is uncommon in the dataset.
We conduct an ablation experiment on the alignment mecha- Once the room nodes are established, the target distribu-
nism for the three core components of our method (NodeDiffI , tion of room adjacencies is easily captured by the model.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 13

500 100 50 20 10 0 TABLE VIII

Time
Q UANTITATIVE COMPARISONS OF THE INTERMEDIATE OUTPUTS
NodeDiffI PRODUCED AT VARIOUS DIFFUSION STEPS BY OUR THREE CORE
w/o Align COMPONENTS (N ODE D IFF I , A DJACENCY D IFF , & PARTITIONING D IFF )
AND THEIR VARIANTS WITHOUT THE ALIGNMENT MECHANISM (A LIGN ).

Input Condition FID Score at Time t

NodeDiffI B Rn,c,s,l Ra t=50 t=20 t=10 t=0
NodeDiffI w/o Align ✓ ✗ ✗ 27.22 6.27 3.76 3.38
NodeDiffI ✓ ✗ ✗ 25.63 5.11 2.72 2.31
500 300 200 150 100 0 B Rn,c,s,l Ra t=200 t=150 t=100 t=0
Time AdjacencyDiff w/o Align ✓ ✓ ✗ 0.58 0.14 0.08 0.08
AdjacencyDiff ✓ ✓ ✗ 0.58 0.12 0.05 0.05
Adjacency B Rn,c,s,l Ra t=50 t=20 t=10 t=0
Diff PartitioningDiff w/o Align ✓ ✓ ✓ 16.27 6.05 3.03 1.71
w/o Align PartitioningDiff ✓ ✓ ✓ 16.14 5.39 2.22 0.85

Adjacency
Diff
of the ground truth distribution, and enhances the capability
500 100 50 20 10 0 of model to generate layouts that more closely resemble the
Time
ground truths.
Partitioning
Diff
w/o Align

D. Perceptual Studies
Partitioning
Diff
We conduct the perceptual studies with designers to evaluate
whether the bubble diagrams and floor plans generated by our
Fig. 10. Visual comparisons of the intermediate outputs produced at various DiffPlannerbI are comparable to those created by professional
diffusion steps by our three core components (NodeDiffI , AdjacencyDiff, & designers. We invite three professional designers, all experi-
PartitioningDiff) and their variants without the alignment mechanism (Align). enced in graphic design and floor plan layout planning, to
participate in the studies. Two groups of experiments are set
Therefore, both AdjacencyDiff and AdjacencyDiff w/o Align up to evaluate bubble diagrams and floor plans, respectively.
produce results consistent with the final output at step 0 by Each group includes 50 comparison tasks where each pair
step 100. However, compared to AdjacencyDiff w/o Align, of bubble diagrams (or floor plans) consists of one from the
AdjacencyDiff with the alignment mechanism is still faster, ground truths (GT) and one generated by our DiffPlannerbI with
achieving a result very close to that at step 0 by step 150. only boundary input. The participants are not aware of the
For the prediction of room partitioning, PartitioningDiff origin of the bubble diagrams or floor plans (being from GT
with the alignment mechanism has essentially moved all room or DiffPlannerbI ), and are asked to complete the forced-choice
boxes inside the boundary by step 50. Although at this time comparison task by rating each pair as “GT better/DiffPlannerbI
the rooms are not well-aligned and the shapes are somewhat better/equally good”. In total, we obtain 150 answers (50 pairs
irregular, it further refines the results, eventually producing × 3 participants) for each group (bubble diagrams and floor
reasonably distributed room boxes at step 0. In contrast, plans).
PartitioningDiff w/o Align has still not completely moved all We further compute Fleiss' Kappa [7], a statistical measure
room boxes inside the boundary by step 20, and by step 0 it of inter-rater agreement, obtaining 0.2 for floor plans and 0.1
produces some impractical designs, such as the balcony that for bubble diagrams, indicating slight agreement among raters.
extends beyond the boundary and does not fully occupy the This result is expected, as layout evaluation is inherently sub-
corners (depicted in yellow located at the bottom right corner), jective, and our studies allow users to make holistic decisions
as shown in Figure 10. without strict evaluation criteria. Interestingly, agreement is
2) Quantitative comparison: We further conduct quantita- higher for floor plans than for bubble diagrams, which aligns
tive comparisons of the intermediate outputs produced at vari- with expectations. Designers rely on well-established evalua-
ous diffusion steps by our three core components (NodeDiffI , tion criteria for floor plan layouts, such as spatial efficiency,
AdjacencyDiff, & PartitioningDiff) and their variants without functionality, and circulation, whereas bubble diagrams lack
the alignment mechanism (NodeDiffI w/o Align, Adjacency- standardized assessment frameworks.
Diff w/o Align, & PartitioningDiff w/o Align). We use the Despite individual user preferences, our perceptual studies
Fréchet Inception Distance (FID) as the evaluation metric be- are highly successful. Figure 11 shows the proportion of votes
cause it assesses the similarity in overall distribution between for the results generated by our DiffPlannerbI and from the GT
the generated results and real data. in the perceptual studies. The voting results for both groups
As shown in Table VIII, the FID scores for all three indicate that the results generated by our DiffPlannerbI are
components decreased at various steps when the alignment nearly indistinguishable from the GT, demonstrating that our
mechanism was included, indicating an improvement in per- method can produce high-quality, realistic bubble diagrams
formance. The results clearly demonstrate that the alignment and floor plans comparable to those designed by professional
mechanism enables faster and more accurate approximation designers.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 14

b b (b) Nearest neighbors in the training dataset

GT better Equally-good DiffPlannerI better (a) DiffPlannerI
1st 2nd 3rd 4th
b

42.7 % 42 % 40 % 40.7 %

15.3 % 19.3 %

Bubble diagram Floor plan

Fig. 11. The result of perceptual studies. The results indicate that the bubble
diagrams and floor plans generated by our method (DiffPlannerbI ) are nearly
indistinguishable from the ground truths (GT).

E. Discussion
1) New data representation: Previous methods [4], [8],
[11], [27], [28], [30], [32] often rasterize vector boundaries and
entrances into sets of binary raster masks, then utilize image-
based networks to extract features. In contrast, our method
feeds the vector boundary and entrance directly into our model,
as data tensors and not as images, avoiding the redundant step
of rasterizing them to extract their embeddings. Additionally,
Fig. 12. Qualitative comparison between the results generated by our
we also represent the prediction targets as data tensors, as DiffPlannerbI based on input samples from the test dataset (a) and their top
illustrated in Figure 2(b). This new representation for both four retrieved nearest neighbors in the training dataset (b).
input and output allows our method to operate entirely in the
vector space, eliminating the need for rasterization in bubble
diagram and floor plan generation tasks. identical solution to the closest neighbor, our model learns
In our approach, the room categories are mapped to a con- meaningful design patterns from the training data, enabling
tinuous range of [−1, 1] and represented as a one-dimensional it to generate reasonable and diverse results. For example, in
tensor. This choice is primarily made to maintain a unified the second row of Figure 12, our DiffPlannerbI , like its nearest
input format for diffusion models training and to mitigate neighbors, places a balcony on the protruding lower-left edge
the high-dimensional sparsity of one-hot encoding. Addition- of the input boundary. However, for other areas, our method
ally, this normalization stabilizes gradient optimization and explores alternative design choices within the diverse design
facilitates smoother training. However, we acknowledge that space instead of replicating the nearest neighbors. Similarly,
converting categorical variables into continuous values may in the fifth row of Figure 12, our method, like its nearest
introduce implicit numerical relationships that do not naturally neighbors, places bedrooms in the upper-left and lower-left
exist between discrete categories. Alternative approaches, such corners, away from the entrance. However, the placement
as discrete diffusion models [2] and hybrid representations, of other rooms differs significantly from that of its nearest
could be explored in future work to better preserve the discrete neighbors in the training dataset, further demonstrating that
semantics while maintaining the efficiency of the generative our model does not overfit but instead generalizes well to new
process. designs.
A natural concern is that data with a tensor-based represen- We also conduct a quantitative comparison of the differences
tation might be too easy to fit, potentially leading to overfitting. between the result generated by our DiffPlannerbI (Our) and its
To address this, we conduct both qualitative and quantitative first nearest neighbor (NN) in the training dataset on the areas
evaluations to demonstrate that our model does not suffer from of six room categories (R) for each sample (s) in the test
this risk. For each sample in the test dataset, we first use dataset (S), using the following metric similar to CoverageGTavg
DiffPlannerbI to generate a corresponding floor plan (shown in in (9):
Figure 12(a)) only from the input boundary. Then, we apply
the retrieval algorithm from Graph2Plan [11] to find a set CoverageN N
avg = [Es∈S IoUs,r (Our, N N )]r∈R . (10)
of the nearest neighbors from the training dataset based on
the similarity of input boundary (shown in Figure 12(b), with The CoverageN N
avg ranges from 0 to 1, where lower scores
similarity decreasing from left to right). indicate less similarity. As shown in Table IX, our DiffPlannerbI
As illustrated in Figure 12, the results generated by our achieves consistently low CoverageN N
avg scores across all six
DiffPlannerbI differ significantly from their nearest neighbors room categories. This indicates that the generated results by
in the training dataset. This suggests that our method does our method differ significantly from their nearest neighbors in
not suffer from overfitting; rather than simply memorizing the training dataset, further demonstrating that our method is
the input boundary-to-floor plan mapping and outputting an robust and does not suffer from overfitting.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 15

TABLE IX floor plans without boundary constraints, HouseDiffusion only

Q UANTITATIVE COMPARISON OF THE DIFFERENCES BETWEEN THE supports generating floor plans from bubble diagrams. How-
RESULT GENERATED BY OUR D IFF P LANNER bI AND THEIR FIRST NEAREST
NEIGHBOR (NN) IN THE TRAINING DATASET ON THE AREAS OF SIX ROOM ever, bubble diagrams contain a lot of user-driven creative
CATEGORIES , INCLUDING THE LIVING ROOM (Rliv ), BEDROOM (Rbed ), information, and can almost be considered the final stage
KITCHEN (Rkit ), BATHROOM (Rbat ), BALCONY (Rbal ), AND STORAGE of user creativity. For the initial stages of user creativity,
(Rsto ). T HE C OVERAGEN N
avg RANGES FROM 0 TO 1, WHERE LOWER
SCORES INDICATE LESS SIMILARITY. the generation of bubble diagrams, HouseDiffusion does not
offer support. Finally, HouseDiffusion also does not support
CoverageN N
avg user-controllable generation; it does not allow for partial user
Rliv Rbed Rkit Rbat Rbal Rsto
inputs, only supporting the generation of floor plans from a
DiffPlannerbI 0.40 0.31 0.12 0.10 0.13 0.00
complete bubble diagram.
In summary, unlike HouseDiffusion, which specifically sup-
ports the transformation from a complete bubble diagram to
2) Application scenarios: Previous methods [4], [8], [11],
a floor plan, our goal is to support the interaction with users
[27], [30], [32] were either specifically designed for boundary-
throughout the entire generative process in all its stages. We
constrained floor plan generation or for bubble diagram-driven
have thus proposed DiffPlanner to achieve our objectives.
floor plan generation. Among them, WallPlan [27] utilized
Both qualitative and quantitative comparisons demonstrate that
a retrieval approach to enable bubble diagram-driven floor
DiffPlanner can generate higher-quality results than HouseD-
plan generation. This approach involved adopting a retrieval
iffusion.
method to match the input bubble diagram and retrieve the
5) Limitation & Future work: In this paper, we represent
corresponding boundary, followed by a boundary-based lay-
a room using the coordinates of its top-left and bottom-
out prediction process. That is, it is still a boundary-driven
right corners. This representation has proven to be efficient
floor plan design method. Different from previous methods,
in previous methods [8], [11], enabling the model to quickly
however, we directly treat the vector boundary as a condition.
capture the target distribution. However, gaps invariably exist
By simply removing the boundary condition, our method can
between room boxes, necessitating a post-processing module
be easily adapted to generate floor plans without boundary
to align the boxes and eliminate these gaps.
constraints, and the entire prediction process will proceed
As shown in Table III, compared to Graph2Plan [11],
without any involvement of boundaries.
our method has improved fivefold in generating high-
3) Decomposition strategy: In floor plan generation, pre- quality floor plans without post-processing (Graph2PlanIII
vious methods [8], [11], [27], [30], [32] have adopted a vs. DiffPlannerbVI ). Nonetheless, it cannot completely elimi-
decomposition strategy, which breaks down the floor plan nate gaps between boxes, as evidenced in Figure 10 where
generation into several sub-tasks. This is a universal strategy some room boxes generated by PartitioningDiff do not align
that has been proven effective in improving the quality of well. Therefore, we adopt the post-processing method in
the generated results as well as enhancing user controllability. Graph2Plan [11] for floor plan generation task, which effec-
In the floor plan domain, the discussion on how to improve tively eliminates gaps between generated room boxes. These
the decomposition strategy to enable better user interaction is limitations may be addressed in future work through the in-
ongoing. troduction of smarter data representations and more advanced
Our approach differs from previous works in the following generative models.
ways: We further modified the decomposition strategy for floor In our approach, each conditioning setup requires training
plans, starting with room nodes, then room adjacencies, and a separate network to optimize performance for the specific
finally room partitioning. This allows our model to be used conditioning scheme. While effective, this may not be the most
to address two tasks: the data-driven generation of bubble efficient solution, especially for multiple conditioning setups.
diagrams and floor plans. This also enables our model to Developing a unified architecture that flexibly integrates arbi-
support the maximum user conditions during floor plan design. trary conditioning remains an interesting direction for future
Additionally, unlike previous methods that simply decompose work.
the tasks and then allow users to modify the outputs of model While our study focuses on a constrained task, floor plan
before proceeding with subsequent operations to achieve user- generation in residential buildings with limited number of
controllable generation, our method supports partial input in rooms, it remains highly combinatorial, requiring the model to
all three sub-tasks. This makes our framework fully user- learn meaningful spatial relationships rather than relying solely
controllable at every step, an achievement not realized by on retrieval. Furthermore, our results demonstrate generaliza-
previous methods. tion beyond training examples, producing diverse layouts that
Overall, our method follows the conventional decomposition align with real-world design principles. Beyond this specific
strategy used in floor plan generation but further updates the task, our approach has the potential to extend to broader
strategy to achieve richer user interactions. vector-based generative tasks, such as graphic layout design,
4) Comparison with HouseDiffusion: While being closely urban planning, or vector graphics generation, which similarly
related and previous state-of-the-art, HouseDiffusion [25] does benefit from direct structure generation without intermediate
not take boundary conditions into account, whereas boundaries rasterization. Future work could explore adaptations of our
are a crucial element in floor plan generation [32], which model to these domains, further testing its applicability.
our method accommodates. Secondly, in the generation of Regarding practical design utility, while our method aligns
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 16

with existing architectural heuristics, its real-world relevance [7] J. L. Fleiss. Measuring nominal scale agreement among many raters.
depends on its integration into design workflows. Conducting Psychological Bulletin, 76(5):378–382, 1971.
[8] F. He, Y. Huang, and H. Wang. iPLAN: Interactive and procedural layout
user studies with professional designers to assess its effective- planning. In Proceedings of IEEE Conference on Computer Vision and
ness in real applications is an important direction for future Pattern Recognition (CVPR), pages 7793–7802, 2022.
research. Also, while the bubble diagram serves as an intuitive [9] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter.
GANs trained by a two time-scale update rule converge to a local nash
and effective intermediate representation, we acknowledge equilibrium. In Advances in Neural Information Processing Systems,
that other representations may be better suited for certain pages 6629–6640, 2017.
application scenarios. Investigating alternative representations [10] J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models.
In Advances in Neural Information Processing Systems, pages 6840–
for improved model controllability and generation quality 6851, 2020.
remains an interesting avenue for further exploration. [11] R. Hu, Z. Huang, Y. Tang, O. van Kaick, H. Zhang, and H. Huang.
Graph2Plan: Learning floorplan generation from layout graphs. ACM
Transactions on Graphics, 39(4):118:1–118:14, 2020.
VI. C ONCLUSION [12] S. Hu, W. Wu, R. Su, W. Hou, L. Zheng, and B. Xu. Raster-to-
In this work, we addressed the challenges of boundary- Graph: Floorplan recognition via autoregressive graph prediction with
an attention Transformer. Computer Graphics Forum, 43(2):1–14, 2024.
constrained floor plan generation by proposing DiffPlanner, a [13] X. Huo, C. Zhou, Y. Xu, and M. Li. A methodology for balancing the
novel deep learning framework that operates entirely in vector preservation of area, shape, and topological properties in polygon-to-
space. By leveraging the diffusion models and a Transformer- raster conversion. Cartography and Geographic Information Science,
49(2):115–133, 2022.
based noise predictor, DiffPlanner effectively handles com- [14] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image trans-
plex vector data and produces high-quality floor plans. Our lation with conditional adversarial networks. In Proceedings of IEEE
approach eliminates the need for redundant rasterization pro- Conference on Computer Vision and Pattern Recognition (CVPR), pages
1125–1134, 2017.
cesses, avoiding common issues such as information loss and [15] D. P. Kingma and J. L. Ba. Adam: A method for stochastic optimization.
distortion. DiffPlanner integrates a conditioning mechanism arXiv preprint arXiv:1412.6980, 2014.
for controlled floor plan generation and an alignment mecha- [16] H. W. Kuhn. The hungarian method for the assignment problem. Naval
Research Logistics Quarterly, 2(1-2):83–97, 1955.
nism during training, aligning the optimization of model with [17] C. Liu, J. Wu, P. Kohli, and Y. Furukawa. Raster-To-Vector: Revisiting
the iterative design processes of designers. This ensures that floorplan transformation. In International Conference on Computer
the model can produce detailed vector layouts from various Vision (ICCV), pages 2214–2222, 2017.
[18] I. Loshchilov and F. Hutter. Decoupled weight decay regularization.
user inputs, and supports various stages of the design process, In International Conference on Learning Representations, pages 1–10,
from early-stage rough designs to detailed floor plans. 2019.
Our extensive experiments, including quantitative com- [19] P. Merrell, E. Schkufza, and V. Koltun. Computer-generated residential
building layouts. ACM Transactions on Graphics, 29(6):181:1–181:12,
parisons, qualitative evaluations, ablation experiments, and 2010.
perceptual studies, demonstrate that DiffPlanner outperforms [20] J. Michalek, R. Choudhary, and P. Papalambros. Architectural layout
existing state-of-the-art methods in generating both bubble di- design optimization. Engineering Optimization, 34(5):461–484, 2002.
[21] N. Nauata, K.-H. Chang, C.-Y. Cheng, G. Mori, and Y. Furukawa.
agrams and floor plans. It offers enhanced user controllability, House-GAN: Relational generative adversarial networks for graph-
supporting fully automatic, coarsely controllable, as well as constrained house layout generation. In Proceedings of European
finely controllable generation modes, thus providing users with Conference on Computer Vision (ECCV), pages 162–177, 2020.
[22] N. Nauata, S. Hosseini, K.-H. Chang, H. Chu, C.-Y. Cheng, and
richer interactions and higher-quality results. Y. Furukawa. House-GAN++: Generative adversarial layout refinement
Overall, our contributions highlight the importance of di- network towards intelligent computational agent for professional archi-
rectly learning from vector data and the potential of diffusion tects. In Proceedings of IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pages 13632–13641, 2021.
models in architectural design and floor plan generation. [23] R. Rengel. The interior plan: Concepts and exercises. Fairchild
DiffPlanner sets a new standard for boundary-constrained and Books/Bloomsbury Publishing Incorporated, 2016.
boundary-unconstrained floor plan generation tasks, show- [24] E. Rodrigues, A. R. Gaspar, and A. Gomes. An approach to the multi-
level space allocation problem in architecture using a hybrid evolutionary
casing the effectiveness of a vector-to-vector approach in technique. Automation in Construction, 35:482–498, 2013.
generating realistic and precise floor plans. [25] M. A. Shabani, S. Hosseini, and Y. Furukawa. HouseDiffusion: Vector
floorplan generation via a diffusion model with discrete and continuous
R EFERENCES denoising. In Proceedings of IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pages 5466–5475, 2023.
[1] S. A. Arvin and D. H. House. Modeling architectural design objectives [26] B. Shneiderman. Human-centered artificial intelligence: Reliable, safe
in physically based space planning. Automation in Construction, & trustworthy. International Journal of Human–Computer Interaction,
11(2):213–225, 2002. 36(6):495–504, 2020.
[2] J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. van den Berg. [27] J. Sun, W. Wu, L. Liu, W. Min, G. Zhang, and L. Zheng. WallPlan:
Structured denoising diffusion models in discrete state-spaces. In Synthesizing floorplans by learning to generate wall graphs. ACM
Advances in Neural Information Processing Systems, pages 17981– Transactions on Graphics, 41(4):92:1–92:14, 2022.
17993, 2021. [28] J. Sun, L. Zheng, G. Zhang, and W. Wu. BubbleFormer: Bubble diagram
[3] F. Bao, D.-M. Yan, N. J. Mitra, and P. Wonka. Generating and exploring generation via dual Transformer models. Computer Graphics Forum,
good building layouts. ACM Transactions on Graphics, 32(4):122:1– 42(7):1–13, 2023.
122:10, 2013. [29] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
[4] S. Chaillou. ArchiGAN: Artificial intelligence x architecture. Architec- Ł. Kaiser, and I. Polosukhin. Attention is all you need. In Advances in
tural Intelligence, pages 117–127, 2020. Neural Information Processing Systems, pages 6000–6010, 2017.
[5] R. G. Congalton. Exploring and evaluating the consequences of vector- [30] S. Wang, W. Zeng, X. Chen, Y. Ye, Y. Qiao, and C.-W. Fu. ActFloor-
to-raster and raster-to-vector conversion. Photogrammetric Engineering GAN: Activity-guided adversarial networks for human-centric floorplan
and Remote Sensing, 63(4):425–434, 1997. design. IEEE Transactions on Visualization and Computer Graphics,
[6] P. Dhariwal and A. Nichol. Diffusion models beat GANs on image 29(3):1610–1624, 2023.
synthesis. In Advances in Neural Information Processing Systems, pages [31] W. Wu, L. Fan, L. Liu, and P. Wonka. MIQP-based layout design for
8780–8794, 2021. building interiors. Computer Graphics Forum, 37(2):511–521, 2018.
IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 17

[32] W. Wu, X.-M. Fu, R. Tang, Y. Wang, Y.-H. Qi, and L. Liu. Data-driven
interior plan generation for residential buildings. ACM Transactions on
Graphics, 38(6):234:1–234:12, 2019.

Shidong Wang is currently working toward the PhD

degree with the Visualization and MultiMedia Lab
(VMML), Department of Informatics, University of
Zurich. His research interests include human-AI
interaction, deep generative modeling, and compu-
tational design.

Prof. Dr. Renato Pajarola has been a Professor in

computer science at the University of Zurich since
2005, leading the Visualization and MultiMedia Lab
(VMML). He has previously been an Assistant Pro-
fessor at the University of California Irvine and
a Postdoc at Georgia Tech. He has received his
Dipl. Inf-Ing. ETH and Dr. Sc. techn. degrees in
computer science from the Swiss Federal Institute
of Technology (ETH) Zurich in 1994 and 1998
respectively. He is a Fellow of the Eurographics
Association and a Senior Member of both ACM and
IEEE. His research interests include real-time 3D graphics, interactive data
visualization, and geometry processing.