LRM-Zero: Training Large Reconstruction Models with Synthesized Data

We present LRM-Zero, a Large Reconstruction Model (LRM) trained entirely on synthesized 3D data, achieving high-quality sparse-view 3D reconstruction. The core of LRM-Zero is our procedural 3D dataset, Zeroverse, which is automatically synthesized from simple primitive shapes with random texturing and augmentations (e.g., height fields, boolean differences, and wireframes). Unlike previous 3D datasets (e.g., Objaverse) which are often captured or crafted by humans to approximate real 3D data, Zeroverse completely ignores realistic global semantics but is rich in complex geometric and texture details that are locally similar to or even more intricate than real objects. We demonstrate that our LRM-Zero, trained with our fully synthesized Zeroverse, can achieve high visual quality in the reconstruction of real-world objects, competitive with models trained on Objaverse. We also analyze several critical design choices of Zeroverse that contribute to LRM-Zero's capability and training stability. Our work demonstrates that 3D reconstruction, one of the core tasks in 3D vision, can potentially be addressed without the semantics of real-world objects. The Zeroverse's procedural synthesis code and interactive visualization are available at: https://desaixie.github.io/lrm-zero/.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer CePT7/10 · confidence 4/52024-07-19

Summary

This paper proposes a pure synthetic training dataset named Zeroverse which is composed of synthetic data generated by simple shape primitives and textures without any real-world semantics. With the Zeroverse dataset, the authors trained a GS-LRM 3D construction model called LRM-Zero and showed that LRM-Zero achieved comparable results to GS-LRM trained on real-world data. The extensive evaluations result to prove the feasibility of using pure synthetic data for 3D reconstruction as well as the effectiveness of the proposed data augmentation techniques.

Strengths

- This paper is well-written and has a clear presentation, comprehensive evaluations, and insightful discussions. - The proposed dataset, Zeroverse, has been shown to deliver great 3D reconstruction results in a sim-to-real generalization manner. This is a very valuable contribution to the community since it proposes a novel aspect of generating synthetic data for 3D reconstruction: the dataset doesn't have to contain real-world semantic objects at all to achieve good 3D reconstruction results. Also, I agree with the argument that 3D reconstruction relies heavily on local cues. - The evaluation results are comprehensive and show that LRM-Zero achieves comparable results to GS-LRM on all the real-world dataset being evaluated.

Weaknesses

- Although Zeroverse and LRM-Zero show very impressive results, it is unclear to me how to improve the result further. LRM-Zero still falls behind GS-LRM, though very closely, and the last bit may be on the semantics. I don't see a clear path to further improve Zeroverse, so the real-world applicability is questionable. - Although the idea of generating a pure synthetic dataset for 3D reconstruction is novel, the dataset generation method and rendering method themselves are mainly taken from the prior work, *Deep image-based relighting from optimal sparse samples*. So there is not much novelty in the dataset generation algorithm. - Training on Zeroverse seems to introduce more instability and one needs to spend more effort on the training tuning. Also, I don't quite understand why training on perfect synthetic data can introduce instability and it could mean that there is something did not work as expected on the data generation side.

Questions

- It would be great if the authors could show a few representative examples where GS-LRM works much better than LRM-Zero, or vice versa. This way we can understand better where the gap is from. - Does it help to use Zeroverse as pretraining and then finetuning on Objectverse? Or how about mixing them during training?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors have discussed the limitations in supplementary.

Reviewer NfXg6/10 · confidence 4/52024-07-22

Summary

The paper explores the feasibility of using procedurally synthesized datasets of 3D shapes to optimize existing large reconstruction models (LRMs), that procure 3D shapes in the form of NeRFs. The data is constructed by 1) sampling several shapes from a set of parametric primitives (cubes, spheres, tori, etc.) with random positions, scales, and rotations; 2) applying randomly sampled textures (from a preselected set of textures); 3) possible application of additional augmentations (changing the curvature of objects, and allowing concave and thin structures). Numerous experiments on existing Objaverse and proposed Zeroverse datasets show that given the correct degrees of the proposed augmentations, existing LRMs can be trained with the proposed synthetic data to produce comparable although slightly worse results, which suggests that existing scarcity of semantically rich 3D data might not be the main limiting factor for the development of reconstruction models since this data is not necessarily needed.

Strengths

* Since the experiments are very computationally demanding, the presented results are valuable (reproducing them is computationally prohibitive for most academic labs). * A lot of ablation experiments explore the effects of the proposed augmentations suggesting the existence of an optimal amount of 3D shape augmentation for proper training of LRMs. * The paper is well-written and easy to follow.

Weaknesses

* In terms of the metrics, models trained with the proposed synthetic data are worse but the authors claim that qualitatively it is hard to see the difference. At the same time, there are almost no qualitative comparisons of reconstructed objects from models trained on real/synthetic data in the main text, supplementary, or on the provided anonymous webpage. It would be better to show these and let the readers conclude if they are close. * One of the main ideas of the paper that the LRMs do not require semantically rich realistic data for training is somewhat successfully challenged in the same paper, given that the LRMs trained for a longer time improve even further, increasing the gap between the proposed synthetic and conventional data. * Another issue, mentioned by the authors is that the amount of the used augmentations for optimal performance may depend on the test dataset, so the domain gap between the synthetic data and considered test sets may differ, which limits, to some extent, the generalization ability of the proposed models.

Questions

Have the authors tried to perform any similar experiments for significantly different data synthesis pipelines? For example, the space of primitives can be extended with existing non-primitive shapes or with shapes generated by existing shape generation methods. If the semantics do not matter, the primitive set should not be crucially important, but at the same time non-primitive shapes used as base shapes will contain some relevant data patterns needed to capture high-frequency details in other non-primitive shapes from the test sets. Maybe this way, the authors would not need to find a set of perfect parameters for the augmentations to make the method work. See limitations.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

While I do not think any ethics review is necessary for this paper, I think it may make sense to disclose the full computational budget for this project (for example in GPU hours spent) to warn any potential readers of the costs of such works.

Reviewer CePT2024-08-09

Thanks for the authors' reply. After reading the responses, I decided to keep my rating as it would be insightful work for the community.

Authorsrebuttal2024-08-12

Webpage is removed

Dear Reviewers, Per the request of the Area Chairs, we have taken town the webpage for GS-LRM and LRM-Zero comparisons. We will add these comparisons to our final version. Thanks, Submission49 Authors

Reviewer 3hvP2024-08-12

Thanks for the authors' reply. Some of my questions are addressed. However, I think my following questions are not answered: (1) L132: In public release version, seems the texture dataset is different from that is used in the paper. Will the performance become worse with the released texture dataset? (2) Would it be interesting to finetune LRM-Zero on a small set of high-quality captured scenes to learn semantics and improve the performance? For question (1), authors can answer it if they have any results on the public texture dataset. For question (2), though authors do not answer explicitly, I think the third row in Tab.1 of the rebuttal pdf shows that such finetuning on the Objaverse dataset with semantics may not help to improve performance on object reconstruction.

Authorsrebuttal2024-08-13

Thanks for the reply and below are our response: 1. **Texture dataset**: We have not tested with the open material dataset due to the limited time and computation cost (need to re-render all objects with new materials and run the training). In our observations, LRM relies more on low-level correspondence instead of high-level semantics for reconstruction, and therefore switching to the public material dataset is not expected to have a big impact on the quality. We will also release our original training shapes and images upon approval for better reproducibility. 2. **Finetune LRM-Zero on data with semantics**: Thanks for the suggestion. We did a preliminary experiment where we fine-tune the Zeroverse model with Objaverse data. We observe that the fine-tuning leads to better results (PSNR increased from 31.2 to 32.3 on GSO) than Zeroverse-only training but can not beat training on Objaverse data. It’s possible that a more careful filtering to better bridge the distribution gap would lead to better results, and a more careful tuning of the fine-tuning parameters such as learning rate would also further boost the results. We also want to clarify the results in rebuttal Table 1: we do Zeroverse training in row 1, Objaverse training in row 2, and Zeroverse+Objaverse joint training in row 3 (i.e., not fine-tuning). We found that join training (row 3\) can improve from Zeroverse-only training (row 1\) results. However, since joint-training merges two data distributions, which is more challenging than training on either Zeroverse or Objaverse, it requires tuning our model configurations (such as a larger network capacity) and training hyperparameters that we did not perform. This explains why joint training (row 3\) performs worse than training only on Objaverse (row 2).

Reviewer xM782024-08-12

I thank the authors for the rebuttal. After reading the rebuttal and other comments by the reviewers I would like to keep my original rating.

Authorsrebuttal2024-08-12

Webpage is removed

Dear Reviewers, Per the request of the Area Chairs, we have taken town the webpage for GS-LRM and LRM-Zero comparisons. We will add these comparisons to our final version. Thanks, Submission49 Authors

Reviewer 3hvP2024-08-14

Thanks for the response. It would be nice to add the results on public texture dataset in the final version to help the readers understand the performance difference.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC