Meta 3D AssetGen: Text-to-Mesh Generation with High-Quality Geometry, Texture, and PBR Materials

We present Meta 3D AssetGen (AssetGen), a significant advancement in text-to-3D generation which produces faithful, high-quality meshes with texture and material control. Compared to works that bake shading in the 3D object's appearance, AssetGen outputs physically-based rendering (PBR) materials, supporting realistic relighting. AssetGen generates first several views of the object with factored shaded and albedo appearance channels, and then reconstructs colours, metalness and roughness in 3D, using a deferred shading loss for efficient supervision. It also uses a sign-distance function to represent 3D shape more reliably and introduces a corresponding loss for direct shape supervision. This is implemented using fused kernels for high memory efficiency. After mesh extraction, a texture refinement transformer operating in UV space significantly improves sharpness and details. AssetGen achieves 17% improvement in Chamfer Distance and 40% in LPIPS over the best concurrent work for few-view reconstruction, and a human preference of 72% over the best industry competitors of comparable speed, including those that support PBR. Project page with generated assets: https://assetgen.github.io

Paper

Similar papers

Peer review

Reviewer Xevg6/10 · confidence 4/52024-07-12

Summary

Emu3D is a novel approach in text-conditioned 3D generation, producing high-quality 3D meshes and materials using Physically-Based Rendering (PBR). It uses a two-stage process: first generating images from standard viewpoints, then reconstructing the 3D shape and appearance. This approach is faster and more reliable than earlier methods and creates realistic 3D graphics by modeling light interaction with surfaces. Emu3D also improves mesh generation by predicting a signed-distance field (SDF) for better quality and refining textures for greater detail. It surpasses existing methods in both image-to-3D and text-to-3D tasks, offering superior visual quality and text alignment.

Strengths

The quality of the generated geometry, texture as well as the PBR materials offers an efficient and high quality generation workflow. The paper is well-presented, with sufficient experiments to validate the quality of the results.

Weaknesses

- The generated PBR material includes only limited components: albedo, metallicity, and roughness. PBR materials typically also consist of other components, such as index of reflection, scattering, coat, sheen etc. Therefore, the generated material can only represent surface reflectance from diffuse to specular.

Questions

- What’s the performance of the proposed method over anisotropy materials? - What's the performance of the extracted mesh over transparent object like a class bottle?

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

- The generated material has significant representational limitations, as it primarily models surface reflection using only albedo, metalness, and roughness.

Reviewer cPQZ6/10 · confidence 4/52024-07-12

Summary

The paper adopts a two-stage PBR information generation method. In the first stage, the text-to-image model is used to predict the PBR channel. In the second stage, it uses SDF to reconstruct the geometry and PBR information. Then, a cross-view attention texture reflector network is used to improve the coarse UV of the input.

Strengths

Methods used are quite effective. The writing is relatively easy to understand. The experiments are comprehensive. The performance is indeed quite good.

Weaknesses

The technical novelty is weak. I noticed that neither the dataset nor the code are intended to be open sourced. I understand that this technology involves commercial use, but I believe that at least all the test sets used in the experiments and the 3D (PBR) assets generated on the test sets should be open sourced to facilitate comparison in future work. The PBR resolution of 512x512 with 2x2 grid for 4-view is still relative low.

Questions

Recently, Unique3d [1] generated mesh with higher-resolution texture. As Emu3d and Unique3d [1] are concurrent works, it is not necessary to directly compare with it. However, theoretically speaking, compared to Unique3d, the texture resolution of Emu3d is still relatively low. It is suggested to add a paragraph in the paper to discuss the advantages and disadvantages of the two methods relative to each other. [1] Wu, Kailu, et al. "Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image." arXiv preprint arXiv:2405.20343 (2024).

Rating

6

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

No. But they have discussed in the limitation part. They use a triplane representation, which cannot represent large-scale scenes.

Reviewer Zgd26/10 · confidence 4/52024-07-13

Summary

This paper introduces a two-stage 3D asset generation pipeline that outputs high-quality 3D meshes with PBR materials in approximately 30 seconds. The technical novelties include: - In text-to-image stage, generating multiple views of both shaded and unshaded images, and predicting PBR materials during the image-to-3D stage. - Using SDF instead of opacity field which yields higher-quality meshes. - Introducing a new texture refiner designed to recover and enhance details in texture maps.

Strengths

The proposed pipeline introduces several components to enhance and extend the existing pipeline [39]. Most design choices are intuitive and well-justified through ablation studies, which include both qualitative and quantitative analyses. Additionally, the paper compares the proposed pipeline with SOTA methods, demonstrating more favorable results through a user study.

Weaknesses

I found certain technical novelties challenging to assess: - The use of SDF instead of opacity fields in the reconstruction pipeline is not a novel idea but more of an extension to Instant3D. This approach has also been explored in 3D generation works such as MeshLRM and Latte3D. - The texture refiner, which I found the most interesting and novel part of the pipeline, is difficult to evaluate from the provided qualitative examples. Looking at the texture maps in Figures 2 and 13, which contain many small components and seams from xAtlas outputs, one concern is whether these artifacts might affect the UNet's predictions in UV space, e.g. causing color inconsistencies. The ablation in Figure 14 only shows textures improved in flat regions, without the corresponding UV maps. Showing texture maps before and after refinement around seams would help readers better understand the model's effectiveness, and justify refinement in UV space as opposed to image space.

Questions

Regarding the blurriness of the pre-refinement texture map: Is this a result of the image-to-3D reconstruction process or an issue with sampling from the texture field? If it is due to the sampling process, why not optimizing with respect to volSDF renderings to recover the details?

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors discuss the limitations of their method and broader impact in the paper.

Reviewer z8qF7/10 · confidence 5/52024-07-15

Summary

The paper mainly works on text-to-3D with PBR materials. It follows the diffusion-based multiview generation and LRM-based reconstruction paradigm. It is a fast feed-forward solution for PBR generation: The diffusion model predicts both shaded and albedo, and the LRM predicts PBR via differentiable rendering with an efficient pixel-wise deferred shading loss. Moreover, it designs a UNet-based texture refiner to project input images to mesh, bringing sharp textures. Also, it adapts LightplaneLRM with the VolSDF renderer and SDF loss to enhance geometry. The generation only takes 30 seconds. It outperforms state-of-the-art feed-forward baselines and performs comparably to top optimization-based industry solutions.

Strengths

1. Good performance. a. It compares with and outperforms some very recent SOTA approaches like InstantMesh and MeshLRM as well as some commercial products like Meshy and Luma Genie. b. It showcases that it can predict well on objects with mixed materials, which is very impressive. Its PBR prediction is SOTA performance. Also, it is not optimization-based and thus fast. c. Its texture refinement module is robust and preserves many details (Figure 4, 8), which helps win out in terms of visual appearance. 2. It conducts many quantitative and qualitative ablation studies to validate each component’s effectiveness. 3. The writing is very clear and easy to follow. Notations are used properly and are well explained. Each design choice is well-motivated and elaborated. It also includes a nice preliminary reading of the used BRDF model.

Weaknesses

1. Image-to-3D may have a degraded material decomposition. Since the proposed approach does not predict albedo for input images, PBR prediction correctness will drop significantly as pointed out in the ablation study (Table 1, Fig. 3). 2. The diffusion model is inherently stochastic, which leads to variance in albedo prediction. How is this variance? How will this variance impact the material decomposition afterward? This point is not quantitatively evaluated. 3. The paper evaluates PSNR for metallic. It is great. But metallic is usually binary. A simple 2x2 confusion matrix on metallic prediction on single-material objects can show the performance more intuitively. 4. If input images are taken under non-uniform lighting conditions, I doubt if albedo can still be correctly decomposed. 5. Lack of failure case analysis. Following the previous point, I wonder if there are some typical failure cases for PBR prediction.

Questions

- Fewer than 10% meshes are selected as high-quality samples to finetune Emu (L226). What is the filtering standard? In terms of appearance quality, material correctness, and composition complexity? Please elaborate more. - L221: What dataset is used? Is it publicly available, like Objaverse? Is the dataset used under proper license? - How is the text-to-image model (3-channel output) adapted to output 6 channels? Please provide more details on that. - For results in Figure 7., if I understand correctly, they are deterministic? What is their PBR PSNR? (How cherry-picked are they?) - May consider adding some qualitative comparisons to optimization-based PBR generation baselines like Richdreamer. Although they are time-costly and may not be as competitive as commercial solutions, such a comparison can strengthen the claim. - In the Image-to-3D part (excluding texture refinement), Are albedo predictions only used as inputs to MLP? - Some qualitative examples of Luma Genie 1.0, stage 2 vs the proposed approach (A.5). It would help to show where optimization helps to get an edge. - 30-sec Runtime breakdown to each step.

Rating

7

Confidence

5

Soundness

4

Presentation

4

Contribution

3

Limitations

Some suggestions are included in the Questions above.

Authorsrebuttal2024-08-12

With the reviewer-author discussion period nearing its end, we want to ensure that our responses have thoroughly addressed your primary concerns. We recognize the importance of this dialogue in resolving the issues raised by the reviewers and are committed to providing any further clarifications you might need. Please let us know if there are any additional questions or if more explanation is required. Warm regards, Emu3D Authors

Authorsrebuttal2024-08-12

With the reviewer-author discussion period nearing its end, we want to ensure that our responses have thoroughly addressed your primary concerns. We recognize the importance of this dialogue in resolving the issues raised by the reviewers and are committed to providing any further clarifications you might need. Please let us know if there are any additional questions or if more explanation is required. Warm regards, Emu3D Authors

Authorsrebuttal2024-08-12

With the reviewer-author discussion period nearing its end, we want to ensure that our responses have thoroughly addressed your primary concerns. We recognize the importance of this dialogue in resolving the issues raised by the reviewers and are committed to providing any further clarifications you might need. Please let us know if there are any additional questions or if more explanation is required. Warm regards, Emu3D Authors

Authorsrebuttal2024-08-12

With the reviewer-author discussion period nearing its end, we want to ensure that our responses have thoroughly addressed your primary concerns. We recognize the importance of this dialogue in resolving the issues raised by the reviewers and are committed to providing any further clarifications you might need. Please let us know if there are any additional questions or if more explanation is required. Warm regards, Emu3D Authors

Authorsrebuttal2024-08-12

We thank the reviewer for raising this important point. We understand these concerns and take ethics and IP rights extremely seriously. The data used in this paper was **not** obtained by scraping the internet. The data was purchased from a well-respected and widely-known vendor of 3D graphic assets. We acquired a data license that **explicitly allows** use of the models in machine-learning applications, including all applications in this paper. We follow the terms and conditions of this license scrupulously, also based on internal legal advice. The exact commercial details of the license are, as it may be expected, confidential.

Reviewer Zgd22024-08-13

Thank the authors for the detailed rebuttal and the additional results provided. It’s surprising to see that the texture refinement is effective, even with UV texture maps that contain numerous isolated pieces. Regarding the third point, I meant to ask about optimizing with respect to the renderings of volSDF, rather than the input images. Nevertheless, I agree with the authors’ point regarding the time cost. The rebuttal has addressed my major concerns, and I have adjusted my scores accordingly.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC