Pith. sign in

REVIEW 5 major objections 6 minor 25 references

RoomCraft: Controllable and Complete 3D Indoor Scene Generation

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A four-stage pipeline converts text, sketches, and photos into complete 3D room layouts while keeping dense constraints from forcing furniture removal.

desk verdict Coherent systems paper with solid heuristics but an evaluation that doesn't yet support the 'significantly outperforms' claim. read the letter →

arxiv 2506.22291 v1 pith:ZLY2OPFF submitted 2025-06-27 cs.CV cs.AI

Mengqi Zhou , Xipeng Wang , Yuxi Wang , Zhaoxiang Zhang This is my paper
classification cs.CVcs.AI
keywords 3Dindoorscenegenerationlayoutprocedurallargelanguagemodelsvision-languageconstraintoptimizationcollisionavoidancespatialrelationships
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

RoomCraft claims to generate complete 3D indoor scenes from a text description, a hand-drawn sketch, or a photo of a real room, and to keep the room complete even when the user piles on many spatial constraints. Its central idea is to turn any input into a structured description—room type, furniture list, and pairwise spatial relations—and then treat layout as an ordering plus collision-avoidance problem rather than as a purely generative one. A heuristic-based depth-first search decides which furniture to place first, and a conflict-aware positioning strategy adjusts placement weights on the fly so that dense constraints do not force items to be deleted. On its reported benchmark of 150 scenes, RoomCraft obtains the best out-of-bound rate (19.3), orientation correctness (89.3), and CLIP similarity (27.7) among seven methods, and a 100-student user study ranks it first on aesthetics, layout coherence, and overall preference.

What carries the argument

The load-bearing machinery is the pairing of a constraint-ordering heuristic with a conflict-aware placement objective. The HDFS heuristic $f(V_i)$ counts, with weights, how many spatial constraints each furniture item participates in and sorts the list in descending order, so the most constrained objects (a table, a sofa) anchor the layout before smaller objects (a cup, a book) are placed on or beside them. The CAPS objective $\mathcal{L}_{plmt}(p)$ balances distance from the opposite wall against distance from nearby objects within radius $\mu$; when a furniture-to-wall collision is detected, $\alpha$ is increased and $\beta=1-\alpha$ adjusts accordingly, and conversely for furniture-to-furniture conflicts. The paper's key step is that these two mechanisms work in sequence: HDFS decides the order, CAPS decides the position, and the dynamic weights are what prevent a growing number of constraints from escalating into item removal.

What would settle it

Feed RoomCraft a set of 50 instructions that use spatial relations outside its predefined taxonomy—'diagonal to,' 'angled 45 degrees from,' 'overlapping but not colliding with'—and compare each generated layout with a human-annotated target; if the layouts systematically fail to reflect these relations, the claim of controllability across diverse inputs is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that controllable 3D indoor generation can be recast as a constraint-satisfaction problem with a fixed planning discipline. Any input is first parsed by a vision-language model into a scene organization $O=\langle R,V,E\rangle$ that names the room type, the furniture inventory, and the spatial relations among objects and between objects and the architecture. A spatial relationship network built from these relations feeds a heuristic-based depth-first search (HDFS) whose heuristic $f(V_i)=\sum_j w_i \mathbb{I}_j(V_i,V_j,E_{i,j})$ orders furniture so that strongly constrained items are placed first. Placement itself is guided by a unified constraint representation—a 5-tuple $C=(T,O,P,R,W)$ covering type, objects, parameters, relation, and weight—so that formal requirements and natural-language adjustments are handled in one loop, while the Conflict-Aware Positioning Strategy (CAPS) minimizes collisions through the objective $\mathcal{L}_{plmt}(p)=\alpha\mathcal{L}_{dist}(p)+\beta\mathcal{L}_{obj}(p,\mu)$ with $\alpha+\beta=1$ and dynamic weight adjustment by conflict type. The authors argue that this combination is what lets RoomCraft preserve all specified furniture under dense multi-constraint conditions, and they point to the best OOB, ORI, and CLIP-Sim scores in their seven-method comparison as evidence.

Load-bearing premise

The pipeline depends on the vision-language model converting every input—text, sketch, or photo—into the correct room type, furniture list, and spatial relations using a fixed vocabulary; if that conversion is wrong, the room can be complete but semantically wrong.

Editorial extensions

If this is right

  • On RoomCraft's reported benchmark, layouts generated from text, sketch, and image inputs all remain complete under multi-constraint conditions, so users can specify dense scenes without expecting furniture to disappear.
  • Because HDFS places strongly constrained furniture first and CAPS shifts weights per conflict type, adding more constraints moves placement quality instead of triggering deletions—directly targeting the collision-then-removal failure the paper identifies in procedural generators.
  • The unified constraint representation lets numerical specifications (Distance(sofa, tv) in [2.0 m, 3.5 m]) and vague phrasing ('the distance feels unnatural') enter the same adjustment loop, which brings interactive refinement of generated rooms closer.
  • If the reported metrics are representative, RoomCraft sets the best published numbers on this comparison set: out-of-bound rate 19.3, orientation correctness 89.3, and CLIP similarity 27.7.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Our inference: the same ordering-plus-collision-avoidance machinery could be transplanted to other constrained placement tasks, such as warehouse layout, robotic rearrangement, or exhibition design, because the constraint 5-tuple and the dynamic weight rule do not encode room-specific semantics.
  • Our inference: the pipeline's ceiling is set by the parser; as vision-language models improve, all downstream metrics should shift, so swapping GPT-4o for a weaker or stronger VLM is a natural experiment to separate parser gains from planner gains.
  • Our inference: the predefined spatial-relation vocabulary (Appendix B.1) is the real boundary of controllability; relations such as diagonal, angled, or 'close but not touching' would likely not be captured, and a stress-test benchmark built on those omissions would be a direct way to probe the claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. RoomCraft is a multi-stage pipeline that generates complete 3D indoor scenes from text descriptions, hand-drawn sketches, or real RGB images. A VLM (GPT-4o) first extracts room type, furniture list, and spatial relations into a structured organization O=<R,V,E>; a scene graph encodes pairwise constraints; a heuristic-based depth-first search (HDFS) orders furniture placement; basic room geometry is generated procedurally in Blender following Infinigen Indoors; and a constraint-driven optimization module with a unified constraint representation, an action space, and a Conflict-Aware Positioning Strategy (CAPS) refines placement to avoid collisions and preserve completeness. The authors report quantitative comparisons against six baselines on 150 scenes (Table 1), a human study (Figure 6), and ablations of CAPS and HDFS, and they conclude that RoomCraft significantly outperforms existing methods in generating realistic, semantically coherent, and visually appealing room layouts across diverse input modalities.

Significance. If properly supported, RoomCraft would be a practically useful contribution: it targets a real limitation of procedural generators (loss of furniture under dense constraints) and of neural generators (repetitive elements and limited global reasoning). The paper is strong on system design: it provides detailed prompt templates (Appendix B), clear algorithmic descriptions (Algorithm 1, Section 3), and explicit statements of limitations (Conclusions). It also attempts an ablation of the CAPS component and a human study. However, the evidence base for the headline claim of significant superiority is currently thin: the two main automatic metrics are undefined, the baseline comparison protocol is underspecified, no uncertainty quantification is provided, and the CAPS hyperparameters appear to be tuned on the test benchmark. These issues are fixable within the scope of a revision, but they are load-bearing for the central claim. The paper does not release code or data, and the evaluation dataset is not fully specified (e.g., '50 RGB images obtained from the website'), which further limits reproducibility.

major comments (5)
  1. [Section 4.1, Table 1] The two headline metrics OOB (Out-of-Bound Rate) and ORI (Orientation Correctness) are named as 'inspired by Chat2Layout' but are never formally defined in the paper: there is no equation, no annotation protocol, and no statement of whether they are computed geometrically, by a VLM, or by human raters. Since the constraint-feedback loop uses a VLM to assess rendered scenes (Section 3.2), a VLM-based metric would risk circularity. Please provide exact definitions and the measurement protocol for both metrics, and specify how the reported numbers in Table 1 were obtained.
  2. [Section 4.1, baselines] The paper states that baselines were evaluated with 'manually modified text descriptions to align with the input specifications of different baseline models', but for the sketch and image conditions it never explains how text-only baselines (LayoutGPT, Holodeck, AnyHome, InstructScene) receive equivalent information. Manual prompt adaptation can bias the comparison by implicitly encoding the expected layout. Please specify the exact prompts and adaptations used per baseline and per input modality, and ideally include a sensitivity analysis over prompt variants.
  3. [Table 1 and Figure 6] No error bars, confidence intervals, or significance tests are reported anywhere, including the human study. The margins over Chat2Layout in Table 1 are only 1.7 OOB, 4.5 ORI, and 0.6 CLIP-Sim on 150 scenes, and the CAPS ablation (Ours vs Ours w/o CAPS) differs by 1.4 OOB, 2.0 ORI, and 0.3 CLIP-Sim; without variance estimates these differences may be within run-to-run noise. Please report per-scene mean ± std, the number of independent runs, and paired significance tests where appropriate.
  4. [Section 4.3 and Figure 8] The CAPS weights α and β in Eq. (7) are tuned through a parameter analysis shown in Figure 8, which is conducted on the same benchmark used to produce the final numbers in Table 1; this is test-set hyperparameter selection and can inflate the reported gains relative to the baselines. Use a held-out validation split for any hyperparameter choice. Additionally, the 'Layout Coherence' metric reported in Figure 8 is never defined elsewhere in the paper; please define it or remove it.
  5. [Section 4.3, 'The influence of HDFS'] The HDFS ablation is presented only qualitatively (Figure 7, comparing RoomCraft to Infinigen Indoors). The claim that HDFS ensures 'the inclusion of all specified furniture and objects' is not quantified. Please provide a quantitative measure such as the furniture-preservation rate, the fraction of scenes that contain all required items, or a layout-completeness score, comparing the pipeline with and without HDFS.
minor comments (6)
  1. [Equation (5)] The notation is inconsistent: the heuristic f(V_i) uses an indicator I_j(V_i,V_j,E_i,j), but the text defines 'I_i(V_i) is a binary indicator function'; also the weight w_i appears to depend on the constraint type and the pair (i,j), so it should be written w_i,j (or defined accordingly).
  2. [Equation (8)] The update rule is unclear: if α increases for furniture-to-wall collisions and β increases for furniture-to-furniture collisions, then the equation β = 1 − α means that increasing α automatically decreases β; please clarify how both can increase simultaneously, or rewrite the update rule to match the stated behavior.
  3. [Section 4.1, dataset] The dataset description states '50 RGB images obtained from the website' without giving any source; please specify the website or dataset used so that the benchmark is reproducible.
  4. [Section 4.1, CLIP-Sim] The CLIP-Sim metric is not sufficiently specified: the exact CLIP model, image resolution, and rendering views used to compute the similarity are missing; please provide these details.
  5. [Appendix B.2, Figures 14 and 15] The text says these figures 'showcase the computational details' of L_dist(p) and L_obj(p, μ), but they appear to be code screenshots; consider including the actual code in the paper or a supplement to make the implementation fully reproducible.
  6. [References] The reference list contains 'Dai et al. 2024a' (Digital Cousins) and 'Dai et al. 2024b' (ACDC) with the same arXiv identifier and nearly identical titles; these are the same work and should be cited once.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the pipeline equations are algorithmic construction choices, the headline comparisons rest on external or human judgments, and the mild test-set hyperparameter selection is a validity concern rather than an input-output equivalence.

full rationale

RoomCraft's claimed derivation chain is a constructive pipeline: Eq. 1-2 extract a structured scene organization O=<R,V,E> from the input via GPT-4o, Eq. 5 defines a heuristic ordering cost, and Eq. 7 defines a placement objective with tunable weights. None of these equations defines a predicted quantity in terms of a fitted parameter, and no reported number in Table 1 is algebraically forced by Eq. 5 or Eq. 7. The CAPS weights alpha and beta are selected via the parameter analysis in Fig. 8, but the paper states in the same paragraph that the alpha-to-beta ratio has minimal effect on OOB and ORI, so the headline OOB/ORI improvements are not by construction consequences of that selection; at worst this is test-set hyperparameter tuning, which is a statistical validity issue, not circularity. The OOB and ORI metrics are never formally defined, so one cannot exhibit the reduction that would be needed to show the evaluation is the same as the optimization objective; that omission is flagged as a correctness risk but is not evidence of circularity. The self-citations to SceneX, CityX, and FurniScene appear in related-work context and are not load-bearing for the method's claims. The Limitations section concedes dependence on pretrained VLMs and on the granularity of predefined spatial relations, but these are external dependencies, not self-referential derivations. The CLIP-Sim metric uses CLIP, which is not part of the generation pipeline, and the user study uses human raters, providing independent evidence. Overall, the paper's central derivation is self-contained as an algorithmic system and no claimed result reduces to its own inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The system leans on GPT-4o, the predefined spatial-relation taxonomy, and two placement heuristics; the paper supplies no independent validation of these assumptions beyond the reported scenes. The free parameters are concentrated in the HDFS weights and the CAPS objective, and their values are either unspecified or tuned on the evaluation metrics.

free parameters (4)
  • constraint weights w_i in HDFS heuristic
    Eq. 5 defines f(V_i) = sum w_i * I_j; no values or update rule are given, and the authors say only that w_i indicates the importance of satisfying the constraint.
  • CAPS weights alpha and beta = ratio near 1.0 recommended from parameter analysis (Fig. 8)
    Eq. 7 defines the placement objective with alpha + beta = 1; the parameter analysis tunes alpha/beta against layout coherence, the same criterion used in the final comparison.
  • CAPS object-distance radius mu = 3
    In Eq. 7, mu is set to 3 without justification or sensitivity analysis.
  • CAPS adjustment rate k and increment delta_alpha
    Eq. 8 defines the dynamic weight update, but k and delta_alpha are never specified.
assumptions (5)
  • domain assumption GPT-4o reliably extracts room type, furniture list, and spatial relations from text, sketches, and RGB images.
    Section 3.1 Eq. 1; the entire pipeline depends on O = GPT-4o(I, P_I) being accurate, and the authors list VLM capability as a limitation.
  • domain assumption The predefined spatial-relation taxonomy in Appendix B.1 is expressive enough to capture user intentions.
    The paper defines furniture-to-architecture, furniture-to-furniture, and distance constraints, but acknowledges fine-grained control is limited by the granularity of these relations.
  • ad hoc to paper HDFS greedy ordering by f(V_i) yields a globally coherent layout.
    Eq. 5-6; no optimality or consistency guarantee is provided, and the ordering assumption is the core algorithmic claim.
  • ad hoc to paper CAPS placement objective is a valid surrogate for layout completeness and collision avoidance.
    Eq. 7; maximizing distance from the opposite wall and minimizing proximity to nearby objects within radius 3 is a heuristic, not derived from analysis.
  • domain assumption Manually modified prompts give baselines an equivalent fair comparison.
    Section 4.1; baselines have different input formats and the authors manually modify text to align, so fairness is assumed rather than controlled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RoomCraft: Controllable and Complete 3D Indoor Scene Generation." pith.science (2026). https://pith.science/paper/ZLY2OPFF

@misc{pith2026250622291,
  author       = {Pith},
  title        = {Pith review of: RoomCraft: Controllable and Complete 3D Indoor Scene Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZLY2OPFF}},
  note         = {Machine review of arXiv:2506.22291}
}
read the original abstract

Generating realistic 3D indoor scenes from user inputs remains a challenging problem in computer vision and graphics, requiring careful balance of geometric consistency, spatial relationships, and visual realism. While neural generation methods often produce repetitive elements due to limited global spatial reasoning, procedural approaches can leverage constraints for controllable generation but struggle with multi-constraint scenarios. When constraints become numerous, object collisions frequently occur, forcing the removal of furniture items and compromising layout completeness. To address these limitations, we propose RoomCraft, a multi-stage pipeline that converts real images, sketches, or text descriptions into coherent 3D indoor scenes. Our approach combines a scene generation pipeline with a constraint-driven optimization framework. The pipeline first extracts high-level scene information from user inputs and organizes it into a structured format containing room type, furniture items, and spatial relations. It then constructs a spatial relationship network to represent furniture arrangements and generates an optimized placement sequence using a heuristic-based depth-first search (HDFS) algorithm to ensure layout coherence. To handle complex multi-constraint scenarios, we introduce a unified constraint representation that processes both formal specifications and natural language inputs, enabling flexible constraint-oriented adjustments through a comprehensive action space design. Additionally, we propose a Conflict-Aware Positioning Strategy (CAPS) that dynamically adjusts placement weights to minimize furniture collisions and ensure layout completeness. Extensive experiments demonstrate that RoomCraft significantly outperforms existing methods in generating realistic, semantically coherent, and visually appealing room layouts across diverse input modalities.

Figures

Figures reproduced from arXiv: 2506.22291 by the authors.

Figure 1
Figure 1. RoomCraft generates 3D room layouts based on user-provided inputs, including real-world images, sketches, or textual descriptions. The resulting [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. RoomCraft: A Four-Stage Pipeline for Converting Multi-Modal Input into 3D indoor Scenes. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Constraint violation detection and correction results. Our framework identifies and resolves various constraint violations including spatial conflicts, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (18 more)
Figure 4
Figure 4. Figure 4: Results for instruct-to-room generation. We produce realistic 3D room scenes guided by three instruction types: textural description, sketch, and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Influence of varying weights 𝛼 and 𝛽 on spatial arrangement in room layout optimization. each furniture item, an initial position is chosen on predefined sur￾faces, such as the floor or walls. This initial position is then adjusted based on spatial constraints to ensur…
Figure 6
Figure 6. Figure 6: Human evaluation results for scene quality in RoomCraft, including [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Qualitative Comparison of generated indoor scenes between Room [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Visual comparison between Infinigen and RoomCraft. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Prompt example for spatial relationship analysis in indoor scenes. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Prompt example for constraint-based furniture relationship formalization. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Prompt example for room type classification. [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Prompt example for furniture enumeration and quantification. [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Implementation of the wall distance constraint function. [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Implementation of the object density component. [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]
Figure 16
Figure 16. Figure 16: Diverse 3D indoor scene generation through varied input text. [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: Diverse 3D indoor scene generation through varied input text. [PITH_FULL_IMAGE:figures/full_fig_p016_17.png]
Figure 18
Figure 18. Figure 18: Diverse 3D indoor scene generation through varied input text. [PITH_FULL_IMAGE:figures/full_fig_p017_18.png]
Figure 19
Figure 19. Figure 19: Diverse 3D indoor scene generation through varied input text. [PITH_FULL_IMAGE:figures/full_fig_p018_19.png]
Figure 20
Figure 20. Figure 20: Diverse 3D indoor scene generation through varied input text. [PITH_FULL_IMAGE:figures/full_fig_p019_20.png]
Figure 21
Figure 21. Figure 21: Diverse 3D indoor scene generation through varied input text. [PITH_FULL_IMAGE:figures/full_fig_p020_21.png]
Figure 22
Figure 22. Figure 22: Gallery of generated 3D indoor scenes [PITH_FULL_IMAGE:figures/full_fig_p021_22.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 21 canonical work pages

  1. [1]

    Identify furniture and objects: Recognize all the furniture and objects present in the picture

    work page

  2. [2]

    Describe spatial relationships: Clearly explain the spatial arrangement and relationships between the items (e.g., which items are adjacent, aligned, in front of, behind, on top of, under, or beside others)

    work page

  3. [3]

    Format: Describe Spatial Relationship in the form of a paragraph

    Categorize placements: Distinguish whether furniture or objects are placed against the wall, in the center of the room, or in any other notable position. Format: Describe Spatial Relationship in the form of a paragraph. Requirement: Your analysis should provide a comprehensive description that captures the overall layout and the relationships between obje...

    work page

  4. [4]

    For example, two cabinets placed side by side create a larger storage area in a dining room

    obj, side by side: Objects are placed next to each other along their sides, forming a continuous line or larger unit. For example, two cabinets placed side by side create a larger storage area in a dining room

    work page

  5. [5]

    obj, left/right to : One object is placed to the left or right side of another, indicating lateral positioning. For instance, a side table placed to the left of a bed provides surface area for essentials, or a floor lamp positioned to the right of an armchair offers optimal lighting for reading

    work page

  6. [6]

    For example, two bedside tables on either side of a bed, providing symmetry and convenient access to essentials from both sides

    obj, aligned with : This involves placing objects symmetrically on both the left and right sides of another object, creating balance and harmony. For example, two bedside tables on either side of a bed, providing symmetry and convenient access to essentials from both sides

    work page

  7. [10]

    global constraint (The relationship between furniture and room):

    work page

  8. [11]

    floor: on floor: Indicates that the furniture is placed on the floor

    work page

Show all 25 references
  1. [12]

    on floor

    wall: flush wall: Used to constrain objects that are mounted on the wall. against wall: An additional constraint to “on floor"; specifies that furniture on the floor is also placed against the wall. far wall: An additional constraint to “on floor"; specifies that furniture on ...

  2. [13]

    ceiling: hanging: used to constrain furniture that is mounted on the ceiling

  3. [14]

    side constraint (between furniture)

  4. [15]

    For example, a coffee table positioned in front of a sofa creates a convenient and accessible surface for those seated

    obj, in front of: This constraint specifies that one object is placed directly in front of another. For example, a coffee table positioned in front of a sofa creates a convenient and accessible surface for those seated

  5. [16]

    An example is two armchairs placed face to face across a coffee table in a living room

    obj, face to face : This involves two objects arranged facing each other, fostering interaction and conversation. An example is two armchairs placed face to face across a coffee table in a living room

  6. [17]

    obj, back to back : This means two objects are placed with their backs against each other, often to optimize space or define separate areas

  7. [21]

    on constraint between furniture obj, on top: Restrict objects to be placed on other items, such as on top of desk, tv stand, coffee table, island, dining table, or cocktail table. Format: The output format must be: object-x | global constraint | side constraint | on constraint...

  8. [22]

    For each object, there must be a global constraint or an on constraint, and there can be a side constraint, but a side constraint is not required

  9. [23]

    against wall

    Situations where side constraint cannot appear: The furniture itself has “against wall” constraint

  10. [24]

    Situations where on constraint appears: This furniture cannot have global constraint or side constraint

  11. [25]

    Please follow the desired format *strictly* (do not add any additional text at the beginning or end) to provide the constraints for each object . Example: Instruction: {room type : living room, furniture type list : sofa 1, side table 1, lamp 1, coffee table 1, vase 1, book st...

  12. [2011]

    ACM Trans

    Metropolis procedural modeling. ACM Trans. Graph. 30, 2 (2011), 11–1. Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner

    2011

  13. [2012]

    Computer Graphics Forum 31 (05 2012), 681–690

    Procedural Generation of Parcels in Urban Modeling. Computer Graphics Forum 31 (05 2012), 681–690. https://doi.org/10.1111/j.1467-8659.2012.03047.x Can Wang, Hongliang Zhong, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. 2024. Chat2Layout: Interactive 3D Furniture L...

    2024

  14. [2019]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision

    Hologan: Unsupervised learning of 3D representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 7588–7597. S. Ochmann, R. Vock, and R. Klein. 2019. Automatic reconstruction of fully volumetric 3d building models from orient...

    2019

  15. [2022]

    Mul- timedia Tools and Applications 81 (05 2022)

    Procedural modeling of plant ecosystems maximizing vegetation cover. Mul- timedia Tools and Applications 81 (05 2022). https://doi.org/10.1007/s11042-022- 12107-8 Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. 2019. Escaping Plato’s Cave: 3D shape from adversarial render...

    2022 arXiv

  16. [2023]

    arXiv preprint arXiv:2310.12945 (2023)

    3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945 (2023). Jerry O Talton, Yu Lou, Steve Lesser, Jared Duke, Radomír Mech, and Vladlen Koltun

    2023 arXiv

  17. [2024]

    Diffuscene: Denoising diffusion models for generative indoor scene synthesis. In Proc. CVPR. 20507–20518. Marcel Torne, Anthony Simeonov, Zechu Li, April Chan, Tao Chen, Abhishek Gupta, and Pulkit Agrawal. 2024. Reconciling reality through simulation: A real-to-sim-to- real ap...

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.