acceptodds
Under review as a conference paper at ICLR 2027

RoomTokens: Masked Diffusion over Structured Tokens for Indoor Scene Synthesis

Abstract

Generating 3D indoor scenes requires a representation that captures object semantics, instance identity, height structure, and orientation while respecting room geometry. Methods that model scenes as sequences of object vectors learn spatial constraints implicitly, leaving collisions and boundary violations to losses or post-processing; semantic-map methods make occupancy explicit, but collapse instance and directional structure. We introduce RoomTokens, a sliced structured- token representation in which each token stores semantic class, instance-color identity, and facing field. This representation allows a scene to be generated as a discrete grid while preserving cues for instance grouping, orientation estimation, and 3D extent prediction. To learn the RoomToken distribution, we propose MDIRT (Masked Diffusion over RoomTokens), a masked discrete diffusion model that denoises semantic/instance-color/facing channels jointly with a shared cross-slice backbone. Architectural tokens are clamped as explicit floor- plan constraints, so the model samples furniture layouts that respect room geometry while retaining instance and object-attribute structure. Experiments on 3D-FRONT show that RoomTokens with MDIRT improve layout fidelity and physical consistency over sequence/vector- and map-based methods, suggesting structured spatial tokens as a practical choice for discrete 3D scene synthesis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Indoor scene generation, generative modelling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_TWqdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes RoomTokens, a sliced spatial representation containing semantic labels, instance-color labels, and body/front orientation cues. MDIRT jointly denoises these fields using masked discrete diffusion with a shared UNet, cross-slice attention, and fixed architectural tokens. Generated components are converted into object records using continuous attribute heads and retrieved 3D assets. Experiments on 3D-FRONT compare against object-vector and semantic-map diffusion methods, including an architecture-clamped SemLayoutDiff variant. The reported results show improved FID, relation statistics, collision rate, and opening clearance, with zero reported out-of-bounds rate. Component ablations and a 12-participant ranking study provide additional support.

Critical strengths
  1. A well-motivated representation. Adding instance separation, facing cues, and height slices directly addresses concrete information lost by a single semantic occupancy map. The representation and shared denoising architecture form a coherent approach to combining spatial constraints with object-level recoverability (Sections 3.1-3.4).

  2. Encouraging results against relevant baselines. Relative to SemLayoutDiff, Table 1 reports FID improving from 71.06 to 49.84, COL from 16.13 to 12.75, and opening clearance from 56.15 to 82.04. Including an architecture-clamped baseline is useful for assessing whether gains extend beyond simply fixing the floor plan. Table 2 ablations and Table 3's 79.6% first-place preference also support the practical value of the complete system, although KID and CKL do not improve over the original SemLayoutDiff.

Critical weaknesses
  1. The contribution of the multi-slice representation is not sufficiently isolated. Multiple height slices are a central component of RoomTokens, but the cross-slice ablation in Table 2 removes attention while retaining the multi-slice representation. This supports the usefulness of cross-slice attention within the proposed architecture, but does not directly establish the benefit of using multiple height slices. The comparison with SemLayoutDiff also changes the token fields and diffusion process. A single top-down baseline retaining the semantic, instance-color, and facing fields, with comparable model capacity and training, would more directly assess the contribution of vertical information. Without this comparison, the main-text evidence leaves an important part of the representation's claimed advantage insufficiently isolated.

  2. The explanation of final boundary compliance is incomplete. Architectural clamping constrains the generated grid, whereas Section 3.4 subsequently scales and rotates retrieved assets, which can change their occupied footprint. Section 4.2 acknowledges that these assembly operations can cause boundary violations for the clamped SemLayoutDiff baseline, yet attributes zero OOB across the MDIRT ablations to retaining architectural clamping. The reported results support boundary compliance on the evaluated scenes, but the main text does not sufficiently explain how the assembly stage preserves it or isolate the contribution of clamping from the rest of the pipeline. A controlled comparison holding scene assembly fixed would strengthen

Questions for authors
  1. What evidence shows that three instance colors are sufficient to reliably distinguish adjacent same-category objects across height slices?

  2. What mechanism keeps the final 3D assets within room boundaries after the scaling and rotation described in Section 3.4?

  3. How much does the multi-slice representation improve performance over a single top-down map with the same semantic, instance-color, and facing fields?

  4. What explains the worse KID and CKL scores than SemLayoutDiff in Table 1 despite the improvement in FID?

4 comments

Thank you for the careful review. We appreciate your recognition that the RoomTokens representation and MDiRT's shared denoiser jointly capture instance separation, orientation, and height. We also appreciate your positive assessment of the improvements in FID, collision, and opening clearance, and of the 79.6% first-place preference in the user study.

Q1: What evidence shows that three instance colors are sufficient to reliably distinguish adjacent same-category objects across height slices?

Answer: Our instance-separation analysis in Appendix C reports 0.92% color conflicts among adjacent same-category, different-instance token pairs and correct instance counts in 91.53% of scenes. The coloring procedure considers horizontal adjacency in every height slice and assigns one consistent color to each instance across slices (Appendix E). These statistics support reliable instance separation across the processed scenes.

To further assess this choice, we have added a three/four-color comparison to Section 4.2, Table 3, which supports the compact palette. See common response C3 for the full results. We have also clarified in Appendix E that the four-color theorem applies to connected regions in a single slice under edge adjacency.

W2 & Q2: What mechanism keeps the final 3D assets within room boundaries after the scaling and rotation described in Section 3.4?

Answer: MDiRT generates RoomTokens with architectural tokens fixed during denoising. During scene assembly (Section 3.4 and Appendix F), we convert the generated RoomTokens into object records and retrieve assets by category and size. The assets are scaled to the estimated extents, rotated to the predicted yaw, and placed at the recovered positions. No boundary correction is applied during assembly; zero OOB is measured after these operations. We have expanded the description in Sections 3.4 and 4.1 to clarify that evaluation follows asset placement without subsequent correction.

Thank you for suggesting further experiments to explain boundary compliance after asset placement. We have added diffusion and clamping comparisons to Section 4.2, Table 3, keeping RoomTokens and scene assembly unchanged. Without clamping, masked diffusion reduces scene OOB from 34.60% to 1.17% compared with multinomial diffusion. Architectural clamping further reduces MDiRT's scene OOB to 0.00% after asset placement. With clamping enabled, masked diffusion also has lower COL (12.75 versus 32.16). These results support both design choices for boundary compliance and physical plausibility (see common response C1 for the full results). We also provide details for the shared assembly and evaluation procedure in common response C2.

W1 & Q3: How much does the multi-slice representation improve performance over a single top-down map with the same semantic, instance-color, and facing fields?

Answer: Thank you for suggesting this comparison. We added a single top-down ablation to Section 4.2, Table 3, projecting the full scene onto one map while retaining the semantic, instance-color, and facing fields, three instance colors, conditioning, backbone, and training settings. MDiRT with multi-slice RoomTokens improves OC from 74.67 to 82.04 and reduces COL from 13.75 to 12.75 while generating more objects. These results directly support retaining height information. See common response C1 for results for all metrics and the other controlled comparisons.

Q4: What explains the worse KID and CKL scores than SemLayoutDiff in Table 1 despite the improvement in FID?

Answer: FID and KID compare feature distributions of rendered scenes using different statistics (Appendix A), so an improvement in FID does not imply an improvement in KID. CKL measures object-category frequencies and is unchanged when the same objects are rearranged. Improved spatial arrangements therefore need not yield lower CKL. We have added this discussion to Section 4.2, emphasizing that we follow the evaluation protocol used by prior work.

Since the distributional metrics do not necessarily reflect the quality of the generated scenes, we also include spatial plausibility metrics and qualitative comparisons, as well as a user study in our evaluation. These metrics indicate that scenes generated by MDiRT are more plausible. For instance, in Table 1 (which we updated with results across multiple sampling seeds), MDiRT reduces COL from 16.24 to 12.64 and increases OC from 56.02 to 82.31 compared with SemLayoutDiff. Most notably, in our user-study, 79.6% ranked scenes generated by our method first.

AuthorsOpenReview ↗

We sincerely hope response from the reviewer before the end of the author-reviewer discussion period.

Dear Reviewer TWqd,

Thank you for the careful evaluation of our paper. We want to know whether our previous response has successfully addressed any of your initial concerns. If not, and you have any further questions, please feel free to share your remaining concerns with us so that we can use our last chance to address them. We want to thank you again for your valuable time and efforts. We will be more than happy to receive your response and resolve any further problems.

Thank you for the detailed responses and additional experiments. They have addressed some of my concerns. However, I am still unclear about how rotation, scaling, and placement interact, and how these factors affect the different evaluation metrics. These questions leave me unconvinced that my main concerns have been fully addressed, so I will keep my score unchanged.

AuthorsOpenReview ↗

Response to Reviewer TWqd

Thank you for clarifying your concern. Below, we provide additional details on how the generated RoomTokens are used to obtain the rotation, scaling, and placement for each object, and how they impact the evaluation metrics.

1. From RoomTokens to the final scene

We group connected RoomTokens with the same semantic and instance-color labels across height slices into object instances. This gives us a set of object instances, each with a semantic label (providing the object category) and covering a set of tokens. The spatial extent of the tokens, together with height-slice information and facing cues, allows us to determine the position, dimensions, and orientation for each object. For each generated object instance (we use a set of prediction heads to estimate the yaw and object dimensions from the pooled features), we then retrieve an asset by category and size, scale and rotate it using these attributes, and place it at the recovered position. See Appendix F (Token-to-object assembly) for details.

Scene assembly steps follow prior work, including DiffuScene and SemLayoutDiff. We use SemLayoutDiff’s asset database and rendering procedure, without subsequent optimization or scene correction.

2. Effects on the evaluation metrics

Scaling sets the dimensions of a retrieved asset. Rotation changes which part of the room it occupies, even when its position stays the same. For example, turning a sofa by 90° can make its bounding box extend past a wall (OOB). The sofa can also overlap a coffee table (COL), block a passage (NAV), or obstruct a door (OC). Rotation and scaling can also change the spatial relationships between objects (Pair-Rel) and how the scene looks in the renderings used for FID and KID.

To test these effects directly, we kept the generated scenes, retrieved assets, and object positions fixed. We set all objects to zero yaw or kept each asset at its retrieved size. Table 10 is reproduced below. We added Figure 6 to Appendix F to illustrate how rotation and scaling affect collisions and navigability. Each panel shows side-by-side renderings of the same scene, comparing MDiRT placement with fixed yaw in (a) and with no scaling in (b).

Asset placementFID ↓KID × 10³ ↓CKL × 10² ↓Pair-Rel ↓#Obj.OOB_S ↓OOB_O ↓COL ↓NAV ↑OC ↑
MDiRT placement49.8410.416.4312.138.360.000.0012.7599.3082.04
Fixed yaw65.5617.456.4318.098.3630.9020.6435.8290.9279.42
No scaling58.2014.436.4312.348.3610.505.3819.8597.0280.17

Rotation. With MDiRT placement, all object boxes stay inside the room. With fixed yaw, 30.90% of scenes contain an out-of-bounds object. COL rises from 12.75% to 35.82%, and NAV falls from 99.30% to 90.92% (Table 10). In Figure 6(a), the rotated sofa and TV stand collide with the coffee table and cut off the purple region from the rest of the walkable floor.

Scaling. Keeping assets at their retrieved size raises scene OOB to 10.50% and COL to 19.85%. NAV falls to 97.02%, a smaller drop than with fixed yaw (Table 10). In Figure 6(b), the taller TV stand overlaps the lower one. This illustrates how scaling can affect collisions even when the walkable floor in that scene stays connected.

CKL and object count stay the same because all three placements use the same objects.

The generative process for RoomTokens outputs facing cues directly. In contrast, SemLayoutDiff uses a two-stage process that first generates a semantic map (without instance information) and then predicts object instances and estimates the orientation from the generated semantic map. This joint reasoning over semantics, instances, and facing cues helps our model learn consistent occupancy and orientation before asset placement. Our experiments show that removing facing cues increases the collision rate considerably – COL from 12.75 to 89.07 (Table 2), supporting the importance of including facing cues.

Our experiments show that our model can generate more consistent object attributes and more plausible assembled scenes.

Reviewer_3uFXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RoomTokens represents indoor scenes as stacked spatial grids. Each cell stores a semantic label, an instance color, and a facing label. A masked diffusion model generates furniture tokens while keeping architectural tokens fixed. Connected components and learned attribute heads recover object layouts. Retrieved assets are then placed to assemble the scene.

Critical strengths
  1. The representation addresses concrete limitations of semantic maps. Instance colors can separate touching objects of the same category. Height slices and facing labels provide additional geometric cues.
  2. Architectural clamping provides a simple way to preserve the observed room structure during generation.
  3. The paper includes a clamped SemLayoutDiff baseline and several representation ablations. The reported physical metrics and human rankings favor the proposed method.
Critical weaknesses
  1. The choice of three instance colors is not well explained. The four-color theorem does not show that three colors are enough to separate neighboring objects.
  2. The zero OOB result needs more explanation. Keeping grid cells inside the room does not ensure that the final furniture stays inside after rotation and scaling.
  3. It is hard to tell what brings the gains over SemLayoutDiff. The method also changes the masking process, training losses, and decoder.
  4. It is unclear whether the baselines were retrained. Several baseline numbers are exactly the same as those reported in SemLayoutDiff.
  5. The sampling process is hard to follow. The paper does not clearly explain when each field is filled or how partially masked tokens are handled.
Questions for authors

Are OOB and COL measured before or after final asset placement?

3 comments

Thank you for the constructive review. We appreciate your recognition that instance colors, height slices, and facing labels address limitations of semantic maps, and that architectural clamping preserves the observed room structure. We also appreciate your positive assessment of the clamped baseline, representation ablations, physical metrics, and human rankings.

W1: The choice of three instance colors is not well explained. The four-color theorem does not show that three colors are enough to separate neighboring objects.

Answer: Three instance colors are an empirical choice, supported by the data analysis in Appendix C: 0.92% of adjacent same-category, different-instance token pairs have color conflicts, while 99.08% have different colors. We have clarified the theorem's scope in Appendix E: four colors suffice for connected regions in a single slice under edge adjacency.

To further assess the three-color choice, we have added a three/four-color comparison to Section 4.2, Table 3: three colors achieve COL 12.75 versus 17.49 for four colors. A fourth color increases the valid semantic–instance-color–facing combinations from 209 to 277. In cases where three colors suffice, the additional combinations would be unused, resulting in sparse training data for the encoder while adding parameters and computation. We provide the full comparison in common response C3.

W2 & Q1: The zero OOB result needs more explanation. Are OOB and COL measured before or after final asset placement?

Answer: OOB and COL are measured after asset retrieval, scaling, rotation, and placement, with no subsequent clipping, optimization, or hand-designed correction. We use the same assets, rendering, and evaluation scripts as SemLayoutDiff. We have clarified this procedure in Sections 3.4 and 4.1 and Appendices A and F.

We have added diffusion and clamping comparisons to Section 4.2, Table 3, keeping RoomTokens and scene assembly unchanged. Without clamping, masked diffusion reduces scene OOB from 34.60% to 1.17% compared with multinomial diffusion. Architectural clamping further reduces MDiRT's scene OOB to 0.00% after asset placement. Please see common responses C1 and C2 which provide the comparison and evaluation details.

W3: It is hard to tell what brings the gains over SemLayoutDiff. The method also changes the masking process, training losses, and decoder.

Answer: Thank you for asking us to clarify these contributions. The component ablations in Table 2 examine the token fields, semantic-priority sampling, cross-slice attention, and class-balancing loss. For example, removing class balancing increases CKL from 6.43 to 21.31. We have also added representation, diffusion, and clamping comparisons to Section 4.2, Table 3, keeping the remaining settings and scene assembly procedure fixed. With RoomTokens and clamping unchanged, masked diffusion reduces COL from 32.16 to 12.75 and improves OC from 62.53 to 82.04 relative to multinomial diffusion. Please see common response C1 which provides the full results.

W4: It is unclear whether the baselines were retrained. Several baseline numbers are exactly the same as those reported in SemLayoutDiff.

Answer: We used the checkpoints from SemLayoutDiff, including those of prior methods retrained under its unified training setting. We followed the same generation and evaluation procedures and reproduced the reported results, explaining the matching entries. The clamped SemLayoutDiff variant reported in Table 1 was separately retrained. We have also added results over multiple sampling seeds to Table 1. Our method uses the same assets, rendering, and evaluation scripts (see common response C2). We will make the checkpoint sources explicit in Section 4.1.

W5: The sampling process is hard to follow. The paper does not clearly explain when each field is filled or how partially masked tokens are handled.

Answer: We have expanded Section 3.2 and Appendix F to clarify our semantic-priority sampling and how denoising happens with partially masked tokens. All fields to be generated start masked. At each reverse step, semantic labels are updated first; instance-color and facing predictions are then recomputed using those labels. Assigned labels are retained. If the semantic label remains masked, instance color and facing also remain masked.

Once a token is assigned “chair,” its instance-color and facing fields may be filled in that step or remain masked until a later step, according to the schedule. Mask means unknown; null means inapplicable. If a token is assigned a background or architectural category, its instance-color and facing fields are set to null, since these attributes apply only to furniture. Validity projection enforces compatible field values while preserving unresolved masks, and architectural tokens remain fixed.

AuthorsOpenReview ↗

We sincerely hope response from the reviewer before the end of the author-reviewer discussion period.

Dear Reviewer 3uFX,

Thank you for the careful evaluation of our paper. We want to know whether our previous response has successfully addressed any of your initial concerns. If not, and you have any further questions, please feel free to share your remaining concerns with us so that we can use our last chance to address them. We want to thank you again for your valuable time and efforts. We will be more than happy to receive your response and resolve any further problems.

Thanks for the clarification. I will adjust my recommendation.

Comments on the paper

8 comments
AuthorsOpenReview ↗

Common response — Part 1: Contributions and ablations

We thank the reviewers for recognizing the representation and the improvements in scene quality and physical plausibility. To summarize, our contributions are:

  • RoomTokens: a compact, sliced structured-token representation with semantic, instance-color, and facing fields. It preserves the information needed for instance grouping, orientation estimation, and 3D extent prediction for indoor scene synthesis.
  • MDiRT: a masked discrete diffusion model that jointly denoises these fields with a shared cross-slice backbone. Architectural tokens remain fixed during denoising to preserve the supplied floor plan.
  • Empirical evidence: improved spatial plausibility, zero reported OOB after asset placement, and 79.6% first-place preference in the user study.

We follow prior work in training and evaluating on the 3D-FRONT dataset with established splits. Specifically, we follow the setup used in SemLayoutDiff, including the rendering and evaluation protocol (C2). Below, we address concerns shared across reviewers, presenting additional experiments and clarifications. We have also revised our paper based on reviewer feedback, with revisions in blue. We thank the reviewers for their suggestions.

C1: Contributions of the representation, diffusion, and architecture clamping

We provide more information about existing ablations we performed as well as additional experiments to isolate the contributions of different components.

Addresses: fk5r W1, W2, Q1, Q3; 3uFX W2, W3, Q1; TWqd W1, W2, Q2, Q3.

With architecture clamped in both methods, MDiRT achieves lower COL than SemLayoutDiff (12.64 versus 40.89, Table 1). Removing instance color or facing increases COL to 88.77 or 89.07 (Table 2); removing class balancing increases CKL from 6.43 to 21.31 (Table 2). Appendix C reports 99.47% object coverage, 0.18 instance-count MAE, and 1.13° direct orientation MAE.

We added the following experiments to examine the effects of multiple height slices, diffusion choice, and clamping:

  • Single top-down slice — fk5r W1/Q1; 3uFX W3; TWqd W1/Q3.
  • Masked versus multinomial diffusion, with and without clamping — fk5r W1/W2/Q1/Q3; 3uFX W2/W3/Q1; TWqd W2/Q2.

All additional ablation models were trained from scratch with the same data split and training settings. The top-down variant replaces the multiple height slices with one top-down scene map (i.e., a single height slice), retaining the three token fields and backbone. The other variants use multi-slice RoomTokens and the same scene assembly but change the diffusion model and whether architectural clamping is on or off. All use the architectural condition and the same rendering and evaluation protocol. We use the same sampling seed for all conditions.

MethodRepresentationDiffusionClampingFID ↓KID × 10³ ↓CKL × 10² ↓Pair-Rel ↓#Obj.OOB_S ↓OOB_O ↓COL ↓NAV ↑OC ↑
MDiRT (Ours)Multi-slice RoomTokensMaskedOn49.8410.416.4312.138.360.000.0012.7599.3082.04
MDiRT variantSingle top-down RoomTokensMaskedOn62.5718.6411.119.026.080.000.0013.7597.3574.67
MDiRT variantMulti-slice RoomTokensMaskedOff57.6114.028.718.986.591.170.1817.6497.4470.94
MDiRT variantMulti-slice RoomTokensMultinomialOn70.5716.145.3211.028.000.030.004232.1693.1062.53
MDiRT variantMulti-slice RoomTokensMultinomialOff70.1418.5912.9216.788.6734.6015.8817.7496.0969.65
SemLayoutDiffTop-down semantic mapMultinomialOn73.5810.544.6322.027.230.500.2640.7795.3765.37
SemLayoutDiffTop-down semantic mapMultinomialOff71.068.654.7820.797.8916.636.6216.1396.3656.15

We show that representation, masked diffusion, and clamping all contribute to spatial plausibility. MDiRT with multi-slice RoomTokens achieves the highest opening clearance (82.04 versus 74.67) and a lower collision rate than the top-down variant while generating more objects. Without clamping, masked diffusion reduces scene OOB from 34.60% to 1.17% compared with multinomial diffusion. Architectural clamping further reduces MDiRT's scene OOB to 0.00% after asset placement. With clamping enabled, masked diffusion also achieves lower COL (12.75 versus 32.16) and higher OC (82.04 versus 62.53). For multinomial diffusion, clamping reduces OOB but increases collisions. Together with the field ablations, these results support the contributions of RoomTokens, masked diffusion, and architectural clamping.

We added these comparisons of height slices, diffusion models, and clamping to Section 4.2, Table 3.

AuthorsOpenReview ↗

Common response — Part 2: Evaluation and instance colors

C2: Final-scene evaluation, boundary compliance, and baseline reproduction

Addresses: fk5r W2, Q3; 3uFX W2, W4, Q1; TWqd W2, Q2.

As described in Appendices A and F, all scene-quality metrics are computed after asset retrieval, scaling, rotation, and placement. All methods use the same category- and size-based asset retrieval, followed by scaling, rotation, and placement, with the same 3D-FUTURE assets, SceneToolkit rendering, and evaluation scripts as SemLayoutDiff.

No post-processing, clipping, optimization, or hand-designed scene correction follows asset placement. Architectural clamping preserves the supplied room geometry. Validity projection enforces compatible semantic, instance-color, and facing fields during denoising. Appendix B reports that 99.49% of tokens, including 100% of furniture tokens, are already valid before projection at the final sampling step. Thus, the final projection corrects only a small fraction of token combinations. The evaluated scenes achieve zero OOB after asset retrieval and placement, without subsequent boundary correction. With scene assembly unchanged, the added comparisons in C1 show that masked diffusion reduces OOB and architectural clamping brings MDiRT's OOB to zero after asset placement.

Section 4.1 specifies SemLayoutDiff's data split and unified training setup. We followed its generation and evaluation procedures and reproduced its reported baseline results, explaining the matching entries. We also evaluated the baselines with multiple sampling seeds (Table 1). We have clarified baseline reproduction in Section 4.1 and scene assembly and evaluation in Sections 3.4 and 4.1, with details in Appendix F.

C3: Three instance colors and separation across height slices

Addresses: 3uFX W1; TWqd Q1.

We provide statistics to justify the choice of three instance colors in Appendix C. Appendix C reports 99.08% conflict-free adjacent same-class, different-instance token pairs and correct instance counts in 91.53% of scenes (MAE 0.18). For instance coloring, we consider horizontal adjacency in every height slice and assign each instance a consistent color across slices (Appendix E).

Appendix E shows that four colors suffice for connected instance regions in a single slice under edge adjacency, motivating a compact instance-color palette with a small number of instance colors. Our choice of three colors is supported by our data analysis (Appendix C) and generation results with three versus four instance colors (thanks for the suggestion).

Note that increasing from three to four instance colors expands the valid semantic–instance-color–facing combinations from 209 to 277. When three colors already separate adjacent instances, the fourth color’s body/front combinations remain unused, making the encoding sparser while still adding embedding and output parameters and computation.

Three/four-color comparison — 3uFX W1; TWqd Q1: we trained the four-color variant from scratch with the same training settings, changing the palette size. Reported three-color MDiRT results are included for reference.

PaletteFID ↓KID × 10³ ↓CKL × 10² ↓Pair-Rel ↓#Obj.OOB_S ↓OOB_O ↓COL ↓NAV ↑OC ↑
Three instance colors49.8410.416.4312.138.360.000.0012.7599.3082.04
Four instance colors58.0314.069.748.406.670.000.0017.4997.2667.74

Three colors achieve COL 12.75 versus 17.49 for four colors, with better navigation and opening clearance; four colors have lower Pair-Rel. These results and Appendix C's diagnostics support the compact three-color choice.

We added the three/four-color comparison to Section 4.2, Table 3. Appendices C and E provide the instance-separation analysis and the four-color result.

AuthorsOpenReview ↗

Common response — Part 3: Sampling uncertainty and user study

C4: Sampling uncertainty and user-study analysis

Here, we address questions from R-fk5r about the variation across runs and details of the user-study analysis.

Addresses: fk5r W2, Q4, Q5.

Five sampling seeds — fk5r W2/Q4. We evaluated MDiRT with five random sampling seeds and added the mean and standard deviation for all ten metrics to Table 1. The results show consistent scene quality across five sampling seeds, including zero reported OOB.

Sampling seedsFID ↓KID × 10³ ↓CKL × 10² ↓Pair-Rel ↓#Obj.OOB_S ↓OOB_O ↓COL ↓NAV ↑OC ↑
Single sampling seed49.8410.416.4312.138.360.000.0012.7599.3082.04
Five random sampling seeds (mean ± SD)50.14 ± 0.2710.35 ± 0.216.81 ± 0.6612.09 ± 0.078.25 ± 0.180.00 ± 0.000.00 ± 0.0012.64 ± 0.3499.25 ± 0.2482.31 ± 0.34

The five-seed results show limited variation in collision and opening clearance, with zero reported OOB across all five seeds.

Participant-level user-study analysis — fk5r W2/Q5.

As detailed in Appendix B, 12 participants each ranked 20 randomly selected sets of four anonymized scenes, covering 50 scene sets across three room types. Scene order was randomized, with each method appearing equally often in each display position. We account for repeated ratings using each participant's mean rank for each method. The Friedman test gives χ²(3) = 34.0 (p < 2 × 10⁻⁷); paired Wilcoxon tests remain significant against every baseline after Holm correction (p = 0.0015). All 12 participants preferred MDiRT on average, with the best mean rank in each room type.

Room typeOurs: mean rank ↓SemLayoutDiff: mean rank ↓DiffuScene: mean rank ↓MiDiffusion: mean rank ↓Ours: first-place (%) ↑
Bedroom1.242.163.193.4181.7
Living room1.421.923.173.5066.7
Dining room1.202.233.083.4982.1

We added the standard deviation across five sampling seeds to Table 1 and the user-study analysis to Section 4.2 and Appendix B (Table 8).

Contribution to the community

To conclude, we contribute a compact representation carrying the information needed for indoor scene synthesis and masked discrete diffusion with architectural clamping. RoomTokens preserve instance separation, orientation, and height information for scene assembly. The ablations and spatial evaluation support these design choices and provide a basis for further research on physically plausible scene synthesis using generative models.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.