CAT3D: Create Anything in 3D with Multi-View Diffusion Models

Advances in 3D reconstruction have enabled high-quality 3D capture, but require a user to collect hundreds to thousands of images to create a 3D scene. We present CAT3D, a method for creating anything in 3D by simulating this real-world capture process with a multi-view diffusion model. Given any number of input images and a set of target novel viewpoints, our model generates highly consistent novel views of a scene. These generated views can be used as input to robust 3D reconstruction techniques to produce 3D representations that can be rendered from any viewpoint in real-time. CAT3D can create entire 3D scenes in as little as one minute, and outperforms existing methods for single image and few-view 3D scene creation. See our project page for results and interactive demos at https://cat3d.github.io .

Paper

Similar papers

Peer review

Reviewer Twzm7/10 · confidence 5/52024-06-30

Summary

This paper proposed CAT3D, a pipeline that enabled the production of 3D representations from one or a few input views. CAT3D comprises a multi-view generation model to synthesize novel images from different viewpoints and a Zip-Nerf to achieve 3D reconstruction based on generated views. The 3D results shown in this paper are very impressive with high quality, and the authors provided sufficient experiments to verify the effectiveness of CAT3D.

Strengths

1. CAT3D achieves impressive results of 3D generation. 2. The overall pipeline of multi-view generation is straightforward, yet effective with good robustness as verified in the experiments. 3. Although CAT3D is trained with constrained multi-view images (no background: objaverse; object-centric: CO3D, MVImgNet, most indoor scenes: RealEstate10k), it still enjoys good generalization. 4. This paper can be seen as evidence to verify that scaling up 3D generation through multi-view synthesis is feasible. 5. The ablation study is convincing and sufficient.

Weaknesses

1. Most techniques have been proposed by previous works, including the raymap, Nerf with LPIPS loss (IM-3D). But I think this point is not the main issue of this paper, while CAT3D proves the scalability of combining all these techniques together. 2. Since ZIP-Nerf is utilized in CAT3D for dense 3D reconstruction, it is unclear whether the results shown in Figures 1, 2, 4, and 5 in the paper are derived from ZIP-Nerf or generated from the multi-view diffusion process. Are all these results exclusively from ZIP-Nerf? If not, it would be beneficial to provide more results directly from the diffusion model. This distinction is important, especially considering the potential inconsistencies noted in the limitations section. 3. It is not inappropriate to provide the inference efficiency with 16 A100 GPUs (Line573) only, which is not a regular setting for most users. The authors should provide the inference time with one GPU. More importantly, it is unclear whether the efficiency results from Table 2 are all fairly compared with one GPU or the same hardware condition. 4. CAT3D seems only being trained with 1cond+7target and 3cond+5target. Could it address the combination of arbitrary conditions and targets without fine-tuning? For example, 4cond+4target or 2cond+6target. 5. CAT3D enjoys good generalization, which is just trained on objaverse, CO3D, MVImgNet, and RealEstate10k. However, these datasets are all constrained (no background: objaverse; object-centric: CO3D, MVImgNet, most indoor scenes: RealEstate10k). How to confirm the generalization of full-model trainable StableDiffusion, especially for some text-to-image samples shown in the supplementary? This point is not mentioned in the main paper. 6. Unfortunately, the authors did not promise to release the code as shown in the checklist. Therefore, some implementation details should be further clarified as mentioned in the "Questions".

Questions

Some unclear implementation: 1. Lines 190-191 are unclear. How to deal with images with different aspect ratios during training and inference? 2. Missing details about the shift of the noise scheduler. 3. What is meant by "drop the text embedding"? Does it imply using an empty string "" as the input text for all samples, or does it completely remove all cross-attention layers? Additionally, how is stability maintained when all cross-attention layers are removed at the beginning of training? Some other minor questions: 1. As mentioned in the paper, a mask channel is concated to the inputs to distinguish conditions and targets. Why not try to use SD-inpainting as the initialization to cover this task? 2. Could 3DGS be converged with fewer generated images? Generating 80/720 views is too costly in my opinion.

Rating

7

Confidence

5

Soundness

4

Presentation

3

Contribution

4

Limitations

The authors discuss the limitations. However, no qualitative limitations are shown in the paper.

Authorsrebuttal2024-08-09

Thanks for the reply

Thanks for the reply. Please find our response below: > It is interesting to see that CAT3D can be generalized to various condition and target views. Is this capability attributable to that no positional encoding is used in CAT3D? To my knowledge, video models with positional encoding fail to be extended to inference with arbitrary lengths. While CAT3D does not use an embedding of time (e.g. positional encoding of the time index for each frame), it does use an embedding of camera pose (via the raymap). Unlike time embeddings where during training you only see a small but finite number of time embeddings, we see a much larger continuous set of pose embeddings which may aid in generalization. In video and language models, one still can get generalization depending on how you structure and interpolate the time embeddings (see e.g. [1, 2]). [1] Chen, Shouyuan, et al. "Extending context window of large language models via positional interpolation." [2] Kazemnejad, Amirhossein, et al. "The impact of positional encoding on length generalization in transformers." > For the shift of the noise scheduler, I think the conclusion of using log(N) to shift the scheduler (N is the number of target images) is not convincing. Because when CAT3D is trained with 5 and 7 target views, it is just evaluated with N=5 in the scheduler. This only confirms that using large N helps multi-view training. The specific relationship between the number of views and the scheduler has not been thoroughly assessed in this paper, leaving the log(N) shifting here somewhat ambiguous. Furthermore, it's important to note that multi-view images should not simply be equated with high-resolution images, given that the overlap among multi-view images can vary greatly and is stochastic. It is common practice in training video diffusion models to shift the noise schedule based on the number of target frames, to compensate for the amount of redundant information which may exist across pixels. This is similar to what is done when increasing a models spatial resolution. In reality, the amount of redundant information in a video is a function of the amount of camera and scene motion, i.e., how many pixels are similar or the same across frames, but the number of frames serves as a reasonable approximation. In practice, our model is adapted from a single image diffusion model, and therefore modifying the noise schedule from that base model is necessary, since supervising and predicting multiple frames strictly has more redundant information. And indeed, in practice, we found shifting noise (while training and sampling) improves the quality of results. It is true that we don’t dynamically adjust the schedule based on the number of target frames, we just use the same shift of log(5) for both 5 and 7 target frames, but we found that the difference between that shift in practice is small (e.g. the average LPIPS on the in-domain diffusion samples is 0.235 for shifting log(5) vs. 0.240 for shifting log(7)). The precise formula for log(N), while not explicitly defined in prior work, was something we derived approximately from the numerical noise schedule information provided in the video model literature [3,4]. We will add this detail to the paper. Future work on multi-view models may want to instead condition the model on logSNR and use different shifts when training and sampling with different numbers of frames. [3] Blattmann, Andreas, et al. "Align your latents: High-resolution video synthesis with latent diffusion models." [4] Blattmann, Andreas, et al. "Stable video diffusion: Scaling latent video diffusion models to large datasets." > I think that providing multi-view generation from diffusion before Nerf learning is very important to confirm the capacity limit of stablediffusion. Even if the model itself cannot be publicly released, sharing samples—such as 80 views generated from Stable Diffusion before NeRF optimization (in the gallery.html)—would greatly benefit the community. This would enable researchers to reproduce performance based on these multi-view images in their own NeRF optimizations. Great point, we will include more samples alongside our NeRF results in the project page (and will also need to combine those videos with a serialized form of the camera pose trajectories).

Reviewer Twzm2024-08-10

Thanks for the reply

Thanks for the reply. My concerns are all addressed. So I raise my score to 7.

Reviewer xucG8/10 · confidence 5/52024-07-11

Summary

In this paper, the authors propose a two-stage method for 3D creation. Specifically, they introduce a multi-view diffusion model to generate novel views given observed input views. Then using these views, they perform a robust 3D reconstruction using a Zip-NeRF variant. To generate consistent views, they also design a data-dependent sampling strategy. In addition, they also conduct extensive experiments to validate the effectiveness of the proposed method.

Strengths

+ The proposed method is simple but effective. It also obtains a superior 3D creation on scenes. + The experiments are exhaustive and validate each components comprehensively, making this paper solid. + This paper is easy to follow and provides much details. Thus, it is easy to reproduce it.

Weaknesses

- It is certain that using larger diffusion models can boost the performance. But it is interesting to showcase the improvement trend with increased diffusion models.

Questions

See the Weaknesses.

Rating

8

Confidence

5

Soundness

4

Presentation

3

Contribution

4

Limitations

Yes, the authors have adequately addressed the limitations and potential negative societal impact of their work.

Reviewer d7U97/10 · confidence 4/52024-07-11

Summary

This paper introduces CAT3D, a novel approach for generating 3D representations from a flexible number of input images. The authors tackle the challenge of limited input data, a common bottleneck for 3D reconstruction, by leveraging the power of multi-view diffusion models. Their method generates a collection of novel viewpoints consistent with the input, effectively transforming a sparse-view reconstruction problem into a more manageable dense-view scenario. These generated views are then fed into a robust 3D reconstruction pipeline based on a modified Zip-NeRF to produce the final 3D model. The authors demonstrate impressive results on various benchmarks, showcasing CAT3D's ability to handle single images, sparse multi-view captures, and even text prompts as input. The method exhibits state-of-the-art performance on few-view 3D reconstruction tasks, outperforming existing methods in terms of both speed and accuracy on established datasets. While single-image 3D generation shows promise, the authors acknowledge the performance is not yet on par with leading methods specifically designed for that task, particularly for single objects. The paper presents a compelling advancement in 3D content creation by unifying different input modalities within a single framework and showcasing significant efficiency gains.

Strengths

- The paper presents a novel approach to 3D content creation by reframing the challenge of sparse input as a view generation problem. This core idea of generating the data needed for robust reconstruction is a valuable contribution to the field. - The authors demonstrate the effectiveness of CAT3D through comprehensive experiments on established benchmarks. Their results on few-view 3D reconstruction tasks are particularly impressive, showcasing state-of-the-art performance on standard metrics and surpassing existing techniques in terms of both speed and accuracy. The ablation study is well-executed, providing valuable insights into the contribution of different components of their method. - The paper is generally well-written and easy to follow. The authors clearly motivate their work, provide sufficient background information, and describe their methodology in a structured manner. The figures are informative and complement the textual descriptions well. - The ability to generate high-quality 3D content from a flexible number of input images, as CAT3D aims to achieve, has substantial practical significance. This flexibility is highly desirable for various applications. The demonstrated speed improvements over existing iterative optimization methods further add to its potential impact by enabling more efficient workflows.

Weaknesses

- While CAT3D aims to handle sparse inputs, its dependence on calibrated camera poses, presents a significant limitation. How does performance degrade with increasing sparsity and decreasing pose accuracy? - The paper relies on manually designed camera trajectories for novel view synthesis, which limits practicality and scalability. The authors briefly mention adapting trajectories based on scene characteristics but provide no concrete details. Developing an automated trajectory selection or optimization procedure, potentially guided by learned priors or scene understanding techniques, would significantly enhance the method's value and broader applicability. - The paper, at times, overstates its accomplishments (e.g., "achieving state-of-the-art performance across nearly all settings") and does not adequately address its limitations. A more nuanced and critical self-assessment would strengthen the work.

Questions

- Could the authors provide a more quantitative assessment of how performance degrades with increasing pose noise or sparsity? - Have the authors considered incorporating an automatic trajectory optimization scheme within CAT3D? - While open-sourcing the code and trained models is ideal, could the authors at least elaborate on their plans for sharing their work and facilitating reproducibility? Providing more details about the training procedure and hyperparameters would also be beneficial.

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors list several limitations, including the reliance on constant camera intrinsics during training, the expressiveness limits of the base text-to-image model, the small number of output views, and the need for manual camera trajectories. However, the discussion lacks concrete examples or quantifiable measures of the limitations.

Reviewer 1Rgc6/10 · confidence 5/52024-07-13

Summary

The objective of this paper is to achieve single-view or few-view to 3D. The core of their method lies in a multi-image-based diffusion model that leverages 3D attention and raymap encoding for the camera poses. This setup is different from concurrent work, IM-3D, which repurposes video generation model to achieve 3D, or ReconFusion, which iteratively refine novel view with diffusion model conditioned on PixelNeRF embeddings. Experiments are conducted against several competitive Baselines, like IM-3D and ReconFusion. Result-wise, their proposed method showcases a nice balance between quality and efficiency.

Strengths

1. The quality of the generated/reconstructed 3D scene is state-of-the-art, and it is more efficient compared with ReconFusion and other iterative methods. 2. The proposed multi-view diffusion model is effective. In LRM related literature, image-based diffusion is tricky to generate multi-view consistent. The proposed CAT3D certainly gives better multi-view image generation quality without explicit maintaining a 3D representation, which can benefit many related tasks. 3. The paper is clearly motivated, and the experiments are carefully designed and reported.

Weaknesses

1. It seems cumbersome to generate a large number of viewpoints(Line:174) by doing anchor first, and enriching frames in between by repeated running CAT3D. Indeed, more view generated together from the multi-view diffusion is good. But running it iteratively will still produce inconsistencies over runs over different set of camera viewpoints, as is can be seen from Fig.6. 2. The trajectory shape seems to influence the quality of the multi-view diffusion. This limits its generalizability to arbitrary/scattered image viewpoints towards the same scene target, which are common in daily life.

Questions

1. What happens if we only use anchor views to run recon in step2? How the quality compared with the current setup? 2. What if we densify view through some existing video interpolation model (instead of re-run CAT3D)? Would be more efficient and better quality?

Rating

6

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes.

Reviewer Twzm2024-08-09

Thanks for the rebuttal

Thanks for the rebuttal. Most concerns about implementation details have been addressed; however, I still have some questions that warrant further discussion. 1) It is interesting to see that CAT3D can be generalized to various condition and target views. Is this capability attributable to that no positional encoding is used in CAT3D? To my knowledge, video models with positional encoding fail to be extended to inference with arbitrary lengths. 2) For the shift of the noise scheduler, I think the conclusion of using log(N) to shift the scheduler (N is the number of target images) is not convincing. Because when CAT3D is trained with 5 and 7 target views, it is just evaluated with N=5 in the scheduler. This only confirms that using large N helps multi-view training. The specific relationship between the number of views and the scheduler has not been thoroughly assessed in this paper, leaving the log(N) shifting here somewhat ambiguous. Furthermore, it's important to note that multi-view images should not simply be equated with high-resolution images, given that the overlap among multi-view images can vary greatly and is stochastic. 3) I think that providing multi-view generation from diffusion before Nerf learning is very important to confirm the capacity limit of stablediffusion. Even if the model itself cannot be publicly released, sharing samples—such as 80 views generated from Stable Diffusion before NeRF optimization (in the gallery.html)—would greatly benefit the community. This would enable researchers to reproduce performance based on these multi-view images in their own NeRF optimizations.

Reviewer xucG2024-08-10

Thanks for the rebuttal. My concerns have been addressed. So I retain my score.

Reviewer 1Rgc2024-08-13

Thanks for the detailed clarification! It is one solid work. I retained my score

Reviewer d7U92024-08-13

Thanks for the rebuttal

Thanks for the clarification. After reading the rebuttal and other reviews, I decide to retain my score (accept). I believe the work deserves to be presented at NeurIPS.

Program Chairsdecision2024-09-25

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC