L4GM: Large 4D Gaussian Reconstruction Model

We present L4GM, the first 4D Large Reconstruction Model that produces animated objects from a single-view video input -- in a single feed-forward pass that takes only a second. Key to our success is a novel dataset of multiview videos containing curated, rendered animated objects from Objaverse. This dataset depicts 44K diverse objects with 110K animations rendered in 48 viewpoints, resulting in 12M videos with a total of 300M frames. We keep our L4GM simple for scalability and build directly on top of LGM, a pretrained 3D Large Reconstruction Model that outputs 3D Gaussian ellipsoids from multiview image input. L4GM outputs a per-frame 3D Gaussian Splatting representation from video frames sampled at a low fps and then upsamples the representation to a higher fps to achieve temporal smoothness. We add temporal self-attention layers to the base LGM to help it learn consistency across time, and utilize a per-timestep multiview rendering loss to train the model. The representation is upsampled to a higher framerate by training an interpolation model which produces intermediate 3D Gaussian representations. We showcase that L4GM that is only trained on synthetic data generalizes extremely well on in-the-wild videos, producing high quality animated 3D assets.

Paper

Similar papers

Peer review

Reviewer 2k2z5/10 · confidence 4/52024-07-08

Summary

This paper proposed L4GM, an efficient large 4D Gaussian reconstruction model to produce animated objects from videos by a single feed-forward. L4GM leverages ImageDream and LGM to achieve multiview images of the first frame as the input. The overall model is built upon the pre-trained LGM with cross-view and temporal attention blocks, which could be easily modified by rearranging the feature dimension. Besides, this paper proposed autoregressive reconstruction and 4D interpolation, largely improving the overall smoothness of the generated 4D. To train L4GM, the authors rendered animated objects from Objaverse, and conducted sufficient experiments to verify the effectiveness.

Strengths

1. As a feed-forward technique, L4GM enjoys good efficiency and performance, especially for the high-resolution 4D generation. 2. The overall pipeline is convincing while the paper is also clearly written. 3. This paper contributes to large-scale 4D data with 12M videos rendered from Objaverse.

Weaknesses

1. The main concern is the novelty. Although the overall pipeline is convincing, this paper is more like an extension of LGM for 4D generation. Most techniques are very straight-forward, such as temporal and cross-view attention, and multi-view synthesis by ImageDream+LGM. 2. Another concern is the setting of repeating the multiview images from the initial timestep as inputs for other frames. It works more like seeking temporary relief rather than a solid solution. For example, this repeating suffers from solving 4D generation with the motion of turning around, whose multiview images would contain conflict to the inputs. Why not use ImageDream+LGM to achieve more multiview inputs? Moreover, more quantitative and qualitative ablation studies should be considered for this setting, such as no-repeating views, repeating views, and multi-timestep-multi-view inputs respectively.

Questions

1. Maybe there is a typo of temp. embed. in Figure 6(b), which should be time embed? 2. Why there are 12M videos at all? The authors said that 110k animations are captured with 48 views. So the overall videos should be about 5M. 3. Since L4GM is trained with animated objects from Objaverse, the comparison of Consistent4D in Table 1 is not convincing enough. To the best of my knowledge, most objects in the Consistent4D test set are from Objaverse too. Please refer to https://github.com/yanqinJiang/Consistent4D/blob/main/test_dataset_uid.txt for more details.

Rating

5

Confidence

4

Soundness

3

Presentation

4

Contribution

2

Limitations

The authors have discussed limitations in the supplementary.

Reviewer hQSp6/10 · confidence 4/52024-07-09

Summary

This paper utilizes rendering of animated objects from objverse(-xl) to extend lgm into 4D generation. Specifically, L4GM uses four orthogonal images of an object and the object's monocular dynamic video to obtain 3D Gaussians at each moment, enhancing the consistency between different moments through temporal self-attention layers. Subsequently, the smoothness of generated actions is further improved through a 4D Interpolation Model. L4GM leads in generation speed and evaluation metrics in the task of video-to-4D.

Strengths

1. This paper significantly improves the generation speed of text-to-4D by utilizing a large-scale, object-centered dynamic dataset to extend lgm into 4D generation, and can model 4D objects in wild videos, achieving highly generalized results. 2. The paper proposes a 4D interpolation Model that can increase the frame rate of generated 4D objects, making the motion smoother. This will alleviate the problem of insufficient frame rates in video generation models.

Weaknesses

The paper discusses extensively how to use dynamic datasets for pre-training, which is also a very important part of this work and will have a significant impact on the community. Whether this dataset is open source is also extremely important for evaluating this work, but the paper does not mention this point.

Questions

1. In section 5, it is mentioned that some objects come with animations. Does this mean the objects already have predefined animations? 2. In the supplementary materials, I noticed that some generated 4D objects have limited motion range and motion rationality, which is a common issue in other 4D generation works. Since evaluating these is a resource-intensive task, are there any possible evaluation methods that could be discussed? This could be part of future work, and once an evaluation method is established, the motion quality of 4D generation could potentially be further improved.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The main issue lies in the amplitude and controllability of the animations. Is there a potential direction for this in the future?

Reviewer hQSp2024-08-08

The author addressed my concerns. I agree that the animation is determined by the video. Thanks a lot. I will raise my score from borderline accept to weak accept.

Authorsrebuttal2024-08-08

Thanks to Reviewer hQSp

We are glad to hear that our response addressed your concerns. Thank you for raising the score.

Reviewer UkTx6/10 · confidence 3/52024-07-13

Summary

This work introduces a novel framework for generating animated 3D objects from single-view videos. The proposed framework employs a feed-forward approach, thus eliminating the need for computationally expensive optimization. The core idea is to create a large-scale synthetic multi-view video dataset and train a corresponding 4D LGM model, which is initialized from the 3D LGM [49] and enhanced with temporal self-attentions. Additionally, an interpolation network is introduced to improve the frame rate further. Overall, this method achieves superior performance in producing 4D assets, both in terms of efficiency and reconstruction quality.

Strengths

- To the best of my knowledge, this is the first work to achieve feed-forward generation of 4D assets from a given video. The successful utilization of a large-scale dataset to tackle this challenging task may inspire further research in the community on related topics. - The proposed framework demonstrates superior performance compared to existing methods across all evaluation metrics, as evidenced by the current evaluation (Figure 5, Tables 1). Additionally, the user study indicates a general preference for the results produced by this framework. - Moreover, several significant ablation studies have been conducted on different components of the framework. Notably, the justification for using a pre-trained LGM (Figure 6(b)) is particularly convincing. - Furthermore, the autoregressive generation method can produce animated objects with longer intervals, beyond the training time intervals.

Weaknesses

Technical contributions - Despite its impressive performance, this work's technical contributions are somewhat limited. The two main contributions are temporal self-attention and the 4D interpolation network. - The temporal self-attention can be seen as a direct extension of the multi-view self-attention proposed in [49]. - While the 4D interpolation network can speed up inference, its overall performance does not seem to have a substantial impact, as demonstrated in the supplementary material video. Having said that, I agree that a straightforward solution to a new problem should be recognized, and the above weakness is relatively minor.

Questions

Some Questions: - Regarding the temporal self-attention layer, why is only the time dimension considered in the self-attention mechanism rather than incorporating both multi-view and temporal information (full attention)? An analysis focusing on efficiency and performance would be valuable to understand the rationale behind this design choice. - Why is the ablation study performed using PSNR (a reconstruction-based metric) instead of the metrics provided in Table 1? - It appears that only qualitative comparisons are offered for the interpolation network. Could the performance be evaluated using the metrics shown in Table 1? - Given that this work is an original contribution to the feed-forward generation of 4D assets, the community would greatly benefit from the release of the dataset and code to enable the reproduction of the work. Are there any plans to release these resources?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

- This work also necessitates a "synthetic" multi-view video dataset for training. According to the implementation details, generating this dataset takes approximately 3 days using 200 GPUs. While this is easier to obtain than a real dataset, it still demands significant computational resources. Including a more comprehensive discussion on this in the limitations section would be beneficial. - The proposed framework is claimed to generalize extremely well to in-the-wild videos (Abstract). However, there are only a few samples provided (4 samples in Figures 1 & 4), and many of these are in an animated style, with the exception of the left example in Figure 4. This left example displays various artifacts on the produced objects, such as blurred arms. Given the current evidence, this claim appears to be inadequately supported, and it is recommended to either include more samples or lower the claim accordingly.

Reviewer 4G9y6/10 · confidence 5/52024-07-13

Summary

This paper proposes a model for 4D reconstruction from a single video, building upon dynamic 3D gaussians and LGM architecture [49] previously applied to static 3D scenes. By processing generated multi-view images (derived from the first frame using prior method) alongside the video, the model outputs 4D Gaussians to reconstruct the video dynamics. The model was trained on a synthetic multi-view video dataset (Objaverse) and shows qualitative generalization to real-world images. Based on the good performance and extensive ablation studies, I recommend accepting this paper.

Strengths

- The paper is clearly written, with good figures, structure, and well-explained motivation of design decisions - The task of 4D reconstruction is both timely and of significant interest to the community. - Though the paper contains strong-worded claims (see below), the demonstrated results show strong 4D reconstruction capabilities - Ablation studies highlight the benefits of pretraining and the chosen representation.

Weaknesses

- Claiming that model generalises “extremely well” to in-the-wild lacks empirical support (apart from cherry-picked qualitative results) and likely not true due to training assumptions (e.g. masks, static camera at 0 degree elevation). One possible evaluation to substantiate the claim would be a comparison with other works such as HyperNeRF. - Since the code is not available, the details about the architecture and training are not sufficient for reproducing experiments and should be expanded. E.g. I could not find how exactly 32 heldout views are sampled neither in supplementary nor in the main paper. - Minor: The architecture and training are not new and relies heavily on prior techniques, including a multi-view image generator. This perhaps aligns work more closely with the computer vision and engineering community (e.g., CVPR)

Questions

- Could you provide more information about various axis of failure cases and the generalization limits to in-the-wild videos? - Could you provide more details about "grid distortion" mentioned on line 594? The referenced paper also lacks clarity on this. - Given the use of 128 80GB A100 GPUs, could you provide details on training the model with fewer resources? (particularly memory use)

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Authors have acknowledged the limitations, however, the claim about generalizing "extremely well" on in-the-wild videos is unsubstantated and likely not true due to training assumptions (e.g. masks, static camera at 0 degree elevation).

Reviewer 2k2z2024-08-10

Thanks for the rebuttal

Thanks for the rebuttal. The rebuttal addressed most of my concerns. I raise my score to 5. Importantly, the limitation and related discussion about repeating views should be included in the revision.

Authorsrebuttal2024-08-12

Thanks to Reviewer 2k2z

Thank you for raising the score. We appreciate your suggestions and will include them in the revision.

Reviewer UkTx2024-08-12

Thank you very much for the additional experiments and comments. These have addressed my concerns. I am maintaining my positive stance towards the current manuscript.

Authorsrebuttal2024-08-12

Thanks to Reviewer UkTx

Thank you for the response. We are glad to have your concerns clarified.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC