SceneCraft: Layout-Guided 3D Scene Generation

The creation of complex 3D scenes tailored to user specifications has been a tedious and challenging task with traditional 3D modeling tools. Although some pioneering methods have achieved automatic text-to-3D generation, they are generally limited to small-scale scenes with restricted control over the shape and texture. We introduce SceneCraft, a novel method for generating detailed indoor scenes that adhere to textual descriptions and spatial layout preferences provided by users. Central to our method is a rendering-based technique, which converts 3D semantic layouts into multi-view 2D proxy maps. Furthermore, we design a semantic and depth conditioned diffusion model to generate multi-view images, which are used to learn a neural radiance field (NeRF) as the final scene representation. Without the constraints of panorama image generation, we surpass previous methods in supporting complicated indoor space generation beyond a single room, even as complicated as a whole multi-bedroom apartment with irregular shapes and layouts. Through experimental analysis, we demonstrate that our method significantly outperforms existing approaches in complex indoor scene generation with diverse textures, consistent geometry, and realistic visual quality. Code and more results are available at: https://orangesodahub.github.io/SceneCraft

Paper

References (80)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer DHDm5/10 · confidence 4/52024-07-10

Summary

This paper proposes a layout-guided method for room generation. The proposed method uses 3D scene layout as a prior and leverages semantic and depth maps as 2D conditions. Specifically, the authors start by training a 2D conditional diffusion model named DreamScene2D, which utilizes ControlNet to incorporate projected semantic and depth maps and conditions. Then the 3D scene can be distillated from the conditional model. The proposed method achieves good rendering quality in complex room scenes.

Strengths

1. Using 3D layout as a prior for 3D scene generation is reasonable, which is a user-friendly format for scene control. 2. The proposed method achieves high rendering quality and good controllability. The generated results can be controlled using both text prompts and scene layouts. 3. The overall writing is easy to follow.

Weaknesses

1. The key insight of the proposed method is similar to recently released paper UrbanArchitect [1], which also uses layout as prior to generate urban scenes (more complex and larger scale). DreamScene uses a pipeline based on InstructNerf2Nerf and UrbanArchitect directly uses a layout-guided SDS. Although the two works can be seen as concurrent works, I hope the authors can include a discussion. 2. As a 3D generation model, it would be better to provide video results for better illustration. The authors can also provide 3D results for better visualization (e.g., generated 3D mesh like text2room). 3. Quantitative results are missing. The authors should at least perform User Study regarding the visual quality and the 3D consistency. Besides, since the ControlNet is trained on specific datasets, metrics like FID can also be used for better evaluation. 4. Some training details are missing, e.g., requirements for memory and time. 5. Since the layout prior is used, it would be desired to use decomposed 3D representations for further editing. 6. It would be better to include a method for automatic scene layout creation. For example, Ctrl-Room utilizes a scene code diffusion mode to generate scene layouts [2]. LLM can also be used to generate scene layouts [3]. [1] Lu, Fan, et al. "Urban Architect: Steerable 3D Urban Scene Generation with Layout Prior." arXiv preprint arXiv:2404.06780 (2024). [2] Fang, Chuan, et al. "Ctrl-room: Controllable text-to-3d room meshes generation with layout constraints." arXiv preprint arXiv:2310.03602 (2023). [3] Hu, Ziniu, et al. "SceneCraft: An LLM Agent for Synthesizing 3D Scenes as Blender Code." Forty-first International Conference on Machine Learning. 2024.

Questions

1. 3DGS has been widely used in 3D generation models due to its high speed and good quality. Have you tried to use 3DGS to replace NeRF? 2. Is there a limit to the complexity of the scenarios a method can handle? 3. Text2Room uses SD for missing parts inpainting. Have you tried to replace SD model in Text2room with your finetuned SD model for inpainting for controllability?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Please refer to weaknesses.

Reviewer Gzhh6/10 · confidence 3/52024-07-12

Summary

This work tackles the problem of 3D scene generation from bounding box scene, text prompt and camera trajectory. The core contribution of the work stems from a 3D scene generation method where the method is able to produce a complex scene, which the previous works could not generate due to using a panoramic representation. The method generates 2D images from multiview images using a diffusion model and then generates 3D. The results show that method is able to generate realistic 3D scenes from the user input.

Strengths

- The method makes sense. Conditioned on the semantic mask, depth map, and text from a viewpoint in the 3D scene, the work uses diffusion model to generate scenes and uses NeRF to reconstruct 3D scenes. To increase 3D consistency the authors use floc removal with periodical migration. - Using the bbox layout seems a nice way to control the 3D scene generation, since the users usually have a specific layout in mind. - The figures show a nice portrait of the method.

Weaknesses

- The scenes look more synthetic compared to the baselines such as MVDiffusion and the text2room. - More qualitative results are required to be convincing. What fascinates me about this work is that the method is able to produce a complex layout scene. However, there is only a single example in the appendix, which the panoramic scene methods cannot generate. The authors should provide more qualitative examples to show that the method can indeed generate complex scenes. Also, the authors claimed that the method can generate various scenes, but only few of the visual examples were provided. - In line 55, the authors mentioned that the method achieves SOTA performance quantitatively. Could the authors provide the quantitative results to back the claim? Also, I would like to see what the “qualitative” here refers to. - In Line 66, the authors mentioned “our DreamScene achieves high-quality generation in various fine-grained and complicated indoor scenes, which are never supported by previous work”. To me this sounds a little bit confusing. Previous works such as ControlRoom3D were able to generate various high-quality scenes, yet they were not able to generate complicated indoor scenes. I think the authors should clearly mention that the completed indoor scenes is the key difference between the previous works.

Questions

- I’m not fully understanding why camera trajectory is required to generate a 3D scene. If the users no the bounding box, couldn’t the method just use random camera trajectory to make the method work? - In line 134, the authors mentioned “.glet@tokeneonedot”, what does this mean? - What is the approximate run time of the method? Since the method generates 2D images and generate 3D, it maybe slow.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

2

Limitations

The method is limited to generation of a small scene. It would be nice to provide results on a large outdoor scene, which does not anyway diminish the work's contribution.

Authorsrebuttal2024-08-14

Reply to reviewer Gzhh’s feedback

We appreciate the reviewer's feedback on the need for more complex scene examples. - During the limited rebuttal period, we have tried our best to generate two new complex scenes, as shown in this anonymous link: https://github.com/DreamScene3D/DreamScene/blob/main/rebuttal-visualization.jpg?raw=true. The last two rows showcase these new results. - We would also like to point out that “complex scenes” are not just limited to non-square shaped ones – square scenes can also be complex so that the baselines will fail. In our Fig.3, the BBS-A/B/C contains many objects that are not adjacent to the walls and corners, which will also create occlusions in panoramas and make the depth map non-continuous and hard to predict. Therefore, these scenes should also be regarded as “complex scenes” where our DreamScene is the only work that can work well. We hope these results could establish our DreamScene’s ability to generate rooms based on various complicated layouts, and address your concern about this. We commit to including these new examples, along with additional complex scene generations, in our future revision.

Reviewer Nerf6/10 · confidence 3/52024-07-15

Summary

DreamScene proposes a method for 3D indoor scene generation with text and layout as input. For this, they propose using a 3D bounding box scene representation as the means to provide layout guidance. This is then used as an input, along with text prompt, to a 2D diffusion model, dreamscene2D, which is capable of generating high-fidelity and high-quality views of rooms following the rendered “bounding-box image” (containing 2D semantic maps and depth) from the bounding box representation. Dreamscene2D is fine-tuned from stable diffusion using LoRA adapters. This is then used for training a NeRF using Score Distillation Sampling. Additional measures, such as a depth constraint, floc removal and texture consolidation with backbone models helps generate the desired scene. In comparison with presented baselines, the results show the ability to model complex shapes in rooms, greater consistency across views, as well as the ability to generate objects of varying sizes. Ablation experiments are provided for the effect of the base prompt, the effect of the depth constraint and the effect of the texture consolidation using backbone models.

Strengths

+ The paper is structured in a manner that is easy to understand + The method is reasonable, and the various components are well motivated in the text and through ablations. + In terms of the qualitative comparison with the shown baseline methods, clear differences can be seen with the proposed method leading to more detailed and accurate scenes + The authors promise to open-source the code post acceptance

Weaknesses

- Lack of quantitative metrics: The paper claims superiority over baselines both qualitatively and quantitatively (L55). However, I am unable to find any quantitative metrics in the paper. - Generation quality: While the proposed method seems to perform better than the chosen baselines, the quality of the generated scenes seems to be relatively lower overall. For example, in the last row of Figure 4, the scene lacks texture and shows a degeneracy in color generation. - Lack of failure cases: In such works as scene generation, and especially since the scope of this work is limited to indoor scenes, I believe it is necessary to contextualize the kind of scenes where the proposed method is successful and where it fails.

Questions

- Runtime: I believe the time taken for 3D asset generation through such text-guided approaches is an essential metric for comparison. However I am unable to find this quantification in this work. - Quantitative metrics: I think it is important to quantify performance of the model (using maybe something like a CLIP score or FID), especially since this is mentioned in the introduction of the paper. - It would be good to get an understanding of failure cases of the proposed method, or settings where the proposed method is unable to generate a favorable output. For now, I would like to give a weak accept score and wait for the authors' responses to the mentioned questions and weaknesses.

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

The authors have addressed limitations and future directions in the appendix which is appreciated. I have some questions with regards to failure cases that I have addressed in previous sections.

Reviewer DRMD6/10 · confidence 4/52024-07-16

Summary

This paper addresses the task of 3D-consistent indoor scene generation using 2D diffusion models. As input the user provides a 3D bounding-box layout of the scene and a text prompt coarsely describing the scene. The proposed method contains a two-stage training phase. In the first stage, a 2D diffusion model is fine-tuned to synthesize a rendered view of a scene based on its rendered semantic layout and a depth map from the same view. In the second stage, the fine-tuned diffusion model is distilled into a NeRF representation with an SDS-based approach. The authors qualitative evaluate their method and compare with existing baselines, showcasing the effectiveness of their method.

Strengths

- The paper is well-written and easy to follow. - The proposed method is straightforward. - The provided results show promising results compared to the baselines. - The authors provide ablation study on their proposed components

Weaknesses

- The proposed method might not be very novel. The contribution of the paper seems mainly to be fine-tuning the diffusion model for semantic-guided image synthesis. The distillation stage is mostly similar to the existing methods. - The authors only evaluate their method qualitatively. Although the provided results seem very promising in terms of the 3D consistency of the scenes, a quantitate evaluation could provide more certainty about the advantage of the proposed method. - The authors could have provided video visualizations in the supplementary, as assessing the 3D consistency of the method from multi-view images is challenging

Questions

See the weaknesses

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Adequately discussed.

Authorsrebuttal2024-08-09

Video Link

Dear Reviewers, here is the **link to the anonymous video with our generated scenes**. https://github.com/DreamScene3D/DreamScene/raw/main/rebuttal_video.mp4

Reviewer Gzhh2024-08-12

Clarification required for the quantiative results

I appreciate the authors for the dedicated response. I have some questions regarding the provided quantiative metrics. Is the quantiative metrics presented here during the rebuttal with same datasets? In the rebuttal to the comments that I have wrote, the authors mentioned that > Our diffusion model is primarily fine-tuned on the synthetic Hypersim [45] dataset, chosen for its high-quality layout data. This choice aligns with other recent works like ControlRoom3D [48]. In contrast, MVDiffusion uses a version of Stable Diffusion fine-tuned on the real-world ScanNet dataset, while Text2Room uses the original Stable Diffusion model. which means that models to obtain the quantiative metrics might have been trained with different datasets, as far as I know. Could the authors clarify this? Because, the dataset matters and might be critical to the obtained quantitative results. Also, what were the choices for not using FID as an evaluation metric?

Authorsrebuttal2024-08-13

Reply to Reviewer Gzhh's feedback at General Response

We sincerely appreciate reviewer *Gzhh* for checking our rebuttal. We address this remaining concern as below. *** 1. **Different training datasets used for quantitative evaluation.** - Our DreamScene2D is designed to generate high-quality 3D scenes that match the input layout. To achieve this, we require a dataset with **high-quality 3D bounding box annotations** with accurate category labels for training. - We therefore chose to train our DreamScene2D using the Hypersim [45] dataset. This choice of the training dataset is also *consistent* with ControlRoom3D [48]. - Other baseline methods use ScanNet as their training dataset. However, the bounding box annotations in ScanNet are of low quality (noisy and inaccurate), and cannot satisfy our requirements. - While we understand the reviewer’s concern that using different training datasets might influence specific quantitative metrics, re-training the baseline models using the Hypersim dataset we employ is challenging and time-consuming. Additionally, we think that dataset selection is a critical design choice in developing these methods and should be considered during evaluation. For example, as mentioned above, we chose Hypersim for DreamScene2D specifically because of its high-quality 3D bounding box annotations. Therefore, using the exact same training data for comparison in our context may not be practical or necessary. - In fact, comparing methods trained on different datasets is a common practice in generative model research. For example, ControlRoom3D [48] also compares with methods (MVDiffusion and Text2Room) trained on different datasets. Similarly, the well-known 2D diffusion model Instruct-Pix2Pix is compared with T2L and SDEdit, which are trained on other datasets. - Importantly, our quantitative experiments focus on the CS, IS, and user study metrics, which evaluate the final generation results using either pre-trained models (CLIP and Inception) or participants’ judgments. These metrics do *not* require ground-truth scenes from the dataset and, therefore, are *agnostic* to the dataset. - Therefore, even when using a different dataset for training, our quantitative comparison remains fair and reasonable with the current metrics and adheres to established practices. 2. **Not using FID metrics.** - First, both our method and [48] focus on the overall quality of generated scenes and layout-following capability, unlike MVDiffusion which focuses on the quality of the diffusion image generation. Therefore, our method and [48] choose to report current metrics rather than FID. - On the other hand, the value of the FID metric is dependent on the dataset. Since we use different datasets for diffusion finetuning, applying FID would result in unfair and inaccurate comparison. *** If the reviewer has any follow-up questions, we are happy to discuss them.

Area Chair bT2w2024-08-11

Dear Reviewer, This is a gentle reminder to please review the rebuttal provided by the authors. Your feedback is crucial to the decision-making process. Please consider updating your score after reading the rebuttal. Thank you for your help with the NeurIPS! Best, Your AC

Reviewer DRMD2024-08-13

Post-Rebuttal Comment

I thank the authors for their answers, in particular for providing the quantitative evaluation and video visualisations. I raise my score to Weak Accept

Authorsrebuttal2024-08-13

Thank reviewer DRMD for your positive feedback!

We appreciate the reviewer for the positive feedback. Your constructive comments and suggestions are very helpful for improving the paper. Also, many thanks for raising the score. We will continue to improve our work and release the code. If the reviewer has any follow-up questions, we are happy to discuss!

Area Chair bT2w2024-08-11

Dear Reviewer, This is a gentle reminder to please review the rebuttal provided by the authors. Your feedback is crucial to the decision-making process. Please consider updating your score after reading the rebuttal. Thank you for your help with the NeurIPS! Best, Your AC

Area Chair bT2w2024-08-11

Dear Reviewer, This is a gentle reminder to please review the rebuttal provided by the authors. Your feedback is crucial to the decision-making process. Please consider updating your score after reading the rebuttal. Thank you for your help with the NeurIPS! Best, Your AC

Area Chair bT2w2024-08-11

Dear Reviewer, This is a gentle reminder to please review the rebuttal provided by the authors. Your feedback is crucial to the decision-making process. Please consider updating your score after reading the rebuttal. Thank you for your help with the NeurIPS! Best, Your AC

Reviewer Gzhh2024-08-12

Could the authors provide more qualititative results if possible?

Thank you for the fruitful response. I really appreciate the author's effort for the response. I know this is a last minute response (and I am sorry for this), but could the authors provide more qualitative results on complex scenes? I ask this because the main argument of the proposed method is that the method can generate complex scenes, however, the only complex scenes in the manuscript and the rebuttal is the 2 scenes in the rebuttal (the scene in the appendix has the same L shaped room as in the rebuttal, otherwise the scenes look like a bbox to me). It's not very convincing to believe that the proposed method can generate complex scenes with just 2 (complex scene) visualizations. I reallize that the authors have only one more day for the rebuttal, but could the authors provide more examples with complex scenes?

Reviewer DHDm2024-08-13

Response to rebuttal

The authors' rebuttal has addressed some of my concerns (training details, quantitative results, video results, etc.). Several experimental results are still missing (comparison with SDS-based pipeline, layout-enhanced Text2Room). Considering the weaknesses discussed above, I keep my score at Borderline Accept.

Authorsrebuttal2024-08-14

We appreciate the reviewer for checking our rebuttal. 1. **For SDS-based pipeline**, we believe that you are referring to the comparison with recent work UrbanArchitect. 1. We would like to point out that UrbanArchitect is an arXiv manuscript released just before the NeurIPS submission deadline, which, as the reviewer also noted, should be considered as a concurrent work of ours. Not comparing with this method should not be regarded as a weakness of ours, especially since this method focuses on a different task than ours (outdoor scene v.s. indoor scene). 2. Due to the limited time in the rebuttal phase, and the substantial difference between the outdoor and indoor scene generation settings (as we have discussed in Q1 of the rebuttal), we are unable to finish training of UrbanArchitect with any indoor dataset to use it for indoor scene generation. 3. As mentioned in the previous rebuttal, we will cite and discuss UrbanArchitect in the revision. 2. For **layout-enhanced Text2Room**: 1. We acknowledge that this can be seen as an interesting exploration. However, we believe this is weaker than the baselines used in our submission, as Text2Room relies on a depth-based naive 3D reconstruction approach. 1. This fundamental aspect of its architecture would likely still result in challenges related to inappropriate or incomplete geometry, regardless of the diffusion model used for inpainting. 2. Simply replacing the diffusion model of Text2Room to our DreamScene2D could only enable the control of the appearance of generated images, but cannot at all help with the geometry issues resulting from the naive reconstruction, as such reconstruction is still only based on a weak monocular depth estimator. 3. Hence, we anticipate that such a model is unlikely to achieve the same level of quality as our current method. Importantly, lacking this ablation study does not demolish the contribution of our work, which was already demonstrated by our state-of-the-art results. 2. For the reviewer’s reference, we would also like to point out that previous published work Ctrl-Room [16] and ControlRoom3D [48], while performing layout-guided scene generation, do not include this additional experiment. 3. Per the reviewer’s request, we have already implemented this baseline and the model is training. However, due to the limited time in the rebuttal phase, we are not able to get the final results before the end of the discussion phase. The current preliminary results indicate that this method is significantly worse than ours. We will add this experiment in the final revision. We hope that our clarifications and additional results will address the reviewer’s remaining concerns. We look forward to a potentially more favorable evaluation of our work.

Reviewer Nerf2024-08-13

Post-Rebuttal Comment

I thank the reviewers for their careful response to the questions that were raised. I appreciate the presented quantitative results, the discussion on limitations and training details. Given the scope and visual quality of the results, and taking into account the author responses, I would like to retain my score of weak accept.

Authorsrebuttal2024-08-13

Thank reviewer Nerf for your positive feedback!

We appreciate the reviewer for the positive feedback. Your constructive comments and suggestions are indeed helpful for improving the paper. We will continue to improve our work and release the code. If the reviewer has any follow-up questions, we are happy to discuss!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC