Neural Assets: 3D-Aware Multi-Object Scene Synthesis with Image Diffusion Models

We address the problem of multi-object 3D pose control in image diffusion models. Instead of conditioning on a sequence of text tokens, we propose to use a set of per-object representations, Neural Assets, to control the 3D pose of individual objects in a scene. Neural Assets are obtained by pooling visual representations of objects from a reference image, such as a frame in a video, and are trained to reconstruct the respective objects in a different image, e.g., a later frame in the video. Importantly, we encode object visuals from the reference image while conditioning on object poses from the target frame. This enables learning disentangled appearance and pose features. Combining visual and 3D pose representations in a sequence-of-tokens format allows us to keep the text-to-image architecture of existing models, with Neural Assets in place of text tokens. By fine-tuning a pre-trained text-to-image diffusion model with this information, our approach enables fine-grained 3D pose and placement control of individual objects in a scene. We further demonstrate that Neural Assets can be transferred and recomposed across different scenes. Our model achieves state-of-the-art multi-object editing results on both synthetic 3D scene datasets, as well as two real-world video datasets (Objectron, Waymo Open).

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer jimW7/10 · confidence 4/52024-06-30

Summary

This paper presents a novel method for 3D-aware editing of multi-object images. Given the original image and 2D bounding boxes to specify the objects to be edited, and the 3D pose as the target of the editing, the proposed method encodes the object appearance feature and their pose features, which are then taken as conditions to generate the edited images. The proposed approach disentangles the appearances and poses of the objects, and enables flexible and better control of the 3D pose of each object in the image.

Strengths

The strengths of this paper include: (1)It proposes a novel approach to disentangle the appearance feature and pose feature of each object in the multi-object image. (2)The proposed approach outperforms the STOA 3D-aware image editing methods, especially in the multi-object setting. (3)It enables object editing and controllable scene generation applications. Overall, this paper presents a practical solution for the 3D-aware multi-object image editing and achieves the best performance. The network design is quite straightforward, but it indeed makes the best use of the pre-trained models and the large training set.

Weaknesses

Basically, this paper provides an excellent 3D-aware image editing model for multi-object images benefiting from the pre-trained image generation model and the large training dataset. The idea of disentangling the appearance and pose features of objects is the key for multi-object image editing. However, since there have been existing works on this topic, such as BlobGAN-3D, it is essential to add an in-depth discussion to validate the superiority of the proposed network architecture. [1] BlobGAN-3D: A Spatially-Disentangled 3D-Aware Generative Model for Indoor Scenes

Questions

Line 244 claims that the proposed method is not pre-trained on multi-view rendering of 3D assets. It is actually trained on the different frames of videos, which is more appropriate for the multi-object image setting. This claim is true but I would not take it as an advantage of the proposed method.

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

This paper includes the discussions of limitations and broader impacts. However, it's better to present some failure cases of the proposed approach to better understand its performance.

Reviewer W8vg7/10 · confidence 4/52024-07-08

Summary

This paper considers controlling the 3d poses of different objects in an image generated by a diffusion model. By conditioning the diffusion model on a sequence of per-object appearance and pose tokens instead of text tokens it is possible to finely control the object poses, and furthermore to e.g. transfer objects between scenes. The diffusion model is trained using videos with 3d bounding box annotations and correspondences between different frames, and it is shown experimentally to generate realistic images and to outperform related methods.

Strengths

- The approach to replace text tokens with a sequence of per-object appearance and pose tokens is very clean and elegant. - The results are impressive and the method achieves state-of-the-art results for the problem. The closest related work is 3DIT and the authors compare to and outperform that method on the tested datasets. - While previous work mainly relied on synthetic data (e.g. 3DIT which introduces the OBJect dataset/task from objaverse assets) the authors propose to use 3d bounding boxes from real-world monocular videos as supervision, which has consistent appearance but different poses, which fits well for the task and can be used to generate training data. - Qualitatively, the generated images generally look very good, especially for Waymo Open. E.g. in fig. 8 there is realistic lighting dependent on the background, so the generated objects adapt well to the rest of the scene. Also, for the scaling examples one can see that the shadows around the object mostly scale accordingly.

Weaknesses

- There seems to be some issues with the modelling of the background for the Objectron dataset, though this is not an issue with Waymo Open. Failure cases are e.g. that rotations and scaling of objects also rotate and scale the background around the object, even outside the object bounding boxes. Did the authors try e.g. any alternative ways of modelling the background to fix this? - This is not a major weakness, but there is no measure of 3d consistency, which is common in previous work, e.g. zero-1-to-3 and follow up works often measure 3d reconstruction errors on e.g. Google Scanned Objects (GSO). It would be straight-forward to evaluate the 3d consistency of the proposed method on this dataset, and compare to e.g. zero-1-to-3. It should be mentioned though that this would not reflect the full strength of the proposed method since it can handle multiple objects and diverse backgrounds which is not captured when evaluating on that dataset.

Questions

Please see weaknesses Can the authors clarify if code and models will be released? It is mentioned that the authors “will consider” releasing code upon acceptance (line 844). Will any trained models be released? Will code to replicate the training be released?

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

This is adequately addressed

Reviewer bNz47/10 · confidence 5/52024-07-12

Summary

The paper introduces an object-centric representation for multi-object 3D pose control in image diffusion models. Instead of text token sequences, it utilizes Neural Assets, which are per-object representations learned by pooling visual features from reference images and reconstructing objects in target frames. This disentangles appearance and pose features by fine-tuning a pre-trained text-to-image diffusion model while conditioning on target frame poses, enabling fine-grained 3D pose and placement control in scenes. The results are demonstrated on both synthetic and real-world datasets in multi-object editing tasks.

Strengths

1. Object-centric representation: the proposed approach introduces a novel object representation to tackle conditional image generation, which disentangles 3D appearance and 3D pose. This representation provides a new solution to tackle similar tasks in controllable 2D image / 3D scene editing tasks. 2. Experimental Results: The experiments demonstrate the effectiveness of the proposed Neural Assets in achieving fine-grained control over individual object appearances and poses within a scene. Its application is shown in both synthetic and real-world scenes, supporting tasks like background swapping and object transfers across different scenes. 3. The paper is overall well written, providing good motivations and comparison with related work. Some concerns in the method section could be further clarified (see below).

Weaknesses

I value the performance the proposed model presented in manipulating objects while keeping the rest unchanged. I spent quite some time trying to comprehend how the current design leads to its performance, and still have the following questions, which might be helpful if they are discussed in the paper: [-] The paper utilizes projection and depth of the absolute object poses in the pose embedding. This is kind of confusing as the it requires the model to encode how the object should look like in the canonical space. If this is the case, the model should be able to perform per-object reconstruction with the object-centric representation through images under different poses and even maybe object pose estimation. A common practice is to use the relative pose change, as the authors discussed in the background modeling. [-] The authors leverage SD as base model, but one significant difference is that the proposed model put the learning bottleneck to the conditional module, compared to Zero-1-to-3 or ControlNet. All the useful features are embedded in the conditions now, including how the images should look like. In original SD, the conditional module is relatively light-weight. I'm wondering if there are limitations for the current design to capture the complex conditions and recover the input images. [-] One related concern is that from the results, it seems the reconstruction results are also decent. I'm wondering what role is the randomness of the noise play in the current model. Will the model output different images if the seed changes? [-] How should the model comprehend the objects in the target images that do not exist in the src images, especially during training? Some other concerns: [-] The model may lack a total understanding of the scene, especially the interactions between objects, e.g., when one object is supported by another or when the 3D bounding boxes of two objects collide with each other. How will the model react in such scenarios? [-] The proposed model faces limitations in fully manipulating 3D object and their relationships as naturally as humans do. The requirement of 2D/3D bounding box pose challenges to the detection models and the manipulation as projection points is not intuitive as humans who commonly use language.

Questions

See above. Overall, I think the paper offers a new perspective for object-centric learning, especially under the image generation settings. I still have some concerns and will consider raising my rating if the rebuttal resolves them.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer h1dc6/10 · confidence 4/52024-07-14

Summary

This paper addresses the task of multi-object pose and scale control in image generation and editing using diffusion models. The authors introduce an object-centric representation with a disentangled pose and appearance called Neural Asset. The neural asset for each object in an image is estimated using pose and appearance encoders. The authors exploit the naturally occurring pose and appearance variations in training video datasets to train such encoders. To generate an image based on neural assets, the authors fine-tune an image diffusion model conditioned on the sequence of neural assets as its condition. The authors evaluate and compare their method with existing baselines, showcasing the ability of their method for pose and scale control, object removal, and background change.

Strengths

- The paper is well-written and easy to follow - The proposed method is reasonable and of sufficient novelty - The provided results show the efficacy of the proposed approach for modeling multiple objects in images and controlling their pose and scale - The experiments are sufficient and contain ablation study on different design choices

Weaknesses

- Based on the provided results, the background has small changes when an object in the image is edited even in the areas far from the edited object - The proposed method is mainly limited to the domain of the training dataset as opposed to the recent training-free diffusion-based image editing methods

Questions

See the weaknesses

Rating

6

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

sufficiently discussed

Reviewer bNz42024-08-08

Post Rebuttal

I thank the authors for their responses. They address my questions about the model design and qualitative results, and thus I have raised my score accordingly.

Reviewer jimW2024-08-09

Thank you for the response. It addresses my questions well. I would maintain my rating as "accept".

Reviewer W8vg2024-08-12

I thank the authors for their answers. I keep my score as 7 (accept).

Area Chair 43uf2024-08-12

Authors' Response

Dear Reveiwer h1dc, The authors have provided a response to the questions that you raised in your review. It would be great if you could read and acknowledge their response before 11:59 pm AOE August 13, 2024. Best, AC

Reviewer h1dc2024-08-13

Final Comment

I thank the authors for their response. I keep my score.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC