MVInpainter: Learning Multi-View Consistent Inpainting to Bridge 2D and 3D Editing

Novel View Synthesis (NVS) and 3D generation have recently achieved prominent improvements. However, these works mainly focus on confined categories or synthetic 3D assets, which are discouraged from generalizing to challenging in-the-wild scenes and fail to be employed with 2D synthesis directly. Moreover, these methods heavily depended on camera poses, limiting their real-world applications. To overcome these issues, we propose MVInpainter, re-formulating the 3D editing as a multi-view 2D inpainting task. Specifically, MVInpainter partially inpaints multi-view images with the reference guidance rather than intractably generating an entirely novel view from scratch, which largely simplifies the difficulty of in-the-wild NVS and leverages unmasked clues instead of explicit pose conditions. To ensure cross-view consistency, MVInpainter is enhanced by video priors from motion components and appearance guidance from concatenated reference key&value attention. Furthermore, MVInpainter incorporates slot attention to aggregate high-level optical flow features from unmasked regions to control the camera movement with pose-free training and inference. Sufficient scene-level experiments on both object-centric and forward-facing datasets verify the effectiveness of MVInpainter, including diverse tasks, such as multi-view object removal, synthesis, insertion, and replacement. The project page is https://ewrfcas.github.io/MVInpainter/.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer NjjA4/10 · confidence 3/52024-07-09

Summary

This paper proposes a new method for multi-view consistent inpainting. The input includes a video and a sequence of masks. The first image could be manipulated by any 2D editing approach, while the remaining images need to be inpainted using the proposed method. The method is built upon the SD1.5-inpainting model and fine-tuned with two LoRAs: one for object-centric and one for scene-level, using different datasets. It employs motion priors by fine-tuning the domain-adapted LoRA and temporal transformers from AnimateDiff, pre-trained on a video dataset. The technique of reference key & value concatenation from previous works is used to enhance consistency with the reference image. Flow features are extracted and added to the UNet as additional conditions. During inference, a heuristic strategy is proposed to warp and propagate the mask from the first frame to all subsequent frames.

Strengths

1. The task of propagating a single 2D edited image to generate a consistent video is interesting. 2. The overall pipeline is well-motivated, built upon the SD1.5-inpainting base model, fine-tuning two LoRAs on different datasets, and including temporal modules, flow features, and reference attention techniques. 3. The proposed method doesn't require camera pose as input and has a short inference time. 4. The paper evaluates the method on several datasets with various tasks and also provides some ablation studies. 5. It's interesting to know that the dense feature is less effective than the slot embeddings.

Weaknesses

1. While the title contains "to Bridge 2D and 3D Editing," the proposed techniques are primarily designed for 2D images or videos, with no explicit 3D representation used or 3D output generated. It would be interesting to explore whether 3D output can be extracted from the resulting image sequence. Otherwise, it might be better to change the title "3D editing" (and other mentions in the text) to something less misleading. 2. The utilization and fine-tuning of domain-adapted LoRA and temporal transformers from video models are intuitive and straightforward, but there do not appear to be any insightful design choices, limiting the novelty. 3. The technique of "Reference Key & Value Concatenation" has been used in many previous works. We fail to see much novelty here. 4. While the section introducing the techniques is generally clear, the concrete settings and details for training, inference, and evaluation are very hard to follow. I tried my best to understand the setup but still feel confused about many details, which is quite frustrating. The paper needs to be rewritten in a more well-organized way, introducing all necessary settings and details. Specifically, the following points need clarification: a) During inference, you employ 2D models to generate reference images for different tasks (e.g., object removal/editing/replacement). However, what is the input-output pair during training? For example, do you have specific ad-hoc training pairs for the object removal task? b) The experiment contains various setups, and the concrete settings are confusing. For each setup, it would be better to explicitly mention the task (e.g., inpainting with object/random masks, object generation, object removal, object replacement), dataset, the input, ground truth output, and how the **mask** are generated. Currently, it only says "Object-Centric Results," "Forward-Facing Results," and "Real-World 3D Scene Editing." But the concrete tasks are unclear. For example, what is the difference between "Object Removal" and "Scene-Level Inpainting"? I assume "object removal" is also achieved by inpainting? If the setup for quantitative and qualitative experiments is different, please also mention it. For example, there are some qualitative results for "object removal," but it should not be evaluated quantitatively (no ground truth)? 5. In equation (1), you concatenated both the masked latents and unmasked noised latents. Some related questions: a) The 9-channel input differs from the original input of SD1.5-inpainting. How do you achieve that? b) The noise-free latent is currently masked after VAE encoding. Will this cause leakage? It might be better to mask before VAE encoding. 6. What is the difference between the "Masking strategy" (line 163) and "masking adaptation" (line 228)? Is one for training and one for inference? 7. The method extracts flows before masking. This may be acceptable for the tasks of removal and replacement, but will it cause leakage for the task of general NVS and inpainting, considering the flow model may have seen the ground truth image?

Questions

1. Have you considered directly fine-tuning a video inpainting model instead of using SD1.5-inpainting? 2. Line 186: The sentence "which always captures the unmasked reference latent without noise (Eq. 1), thus it is unnecessary to re-scale the latent before adding noise from another U-Net" is unclear to me. Could you please clarify? 3. Line 207: The sentence "except that the former should be normalized in the query dimension, while the latter is normalized in the key dimension" is unclear. Could you elaborate? 4. Line 212: The phrase "with comparable performance and fewer trainable weights" needs clarification. Which two setups are being compared? 5. Could you provide more details about the 3D attention mechanism? 6. Line 309: In the context of not having object-level tracking masks, how do you obtain the mask?

Rating

4

Confidence

3

Soundness

2

Presentation

2

Contribution

2

Limitations

Yes, discussed in supplementary.

Reviewer n43z7/10 · confidence 4/52024-07-12

Summary

The paper formulates the 3D object editing task as a multi-view 2D in-painting task. Firstly, MVInpainter-F is employed to remove the object and obtain the background scene. Then, MVInpainter-O generates multi-view images based on the reference view.

Strengths

1. Solving the 3D object editing task as multi-view generation is interesting and innovative. 2. The thoughtful selection of different training datasets (indoor scene/object-centric) for training MVInpainter-F and MVInpainter-O effectively decomposes the object replacement task. 3. The extensive experiments, including the generation of long sequences and the 3D scene reconstruction using the generated multi-view images, further demonstrate the 3D consistency of the resulting outputs.

Weaknesses

It is uncertain whether the design of mask adaption is robust enough.

Questions

1. During training, the flow is obtained using RAFT and masked. In the inference stage of MVInpainter-F, is the flow only obtained from the masked image? Does the utilization of this low-quality flow negatively impact the results? 2. How can we ensure that mask adaption remains robust when dealing with different masks and backgrounds?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The author provides a detailed discussion of limitations and broader impacts in the supplementary material.

Reviewer VJEq5/10 · confidence 3/52024-07-14

Summary

This paper proposes a new 3D editing method by regarding it as a multi-view 2D inpainting task. It ensures cross-view consistency through video priors and concatenated reference key/value attention and controls camera movement without explicit poses using slot attention. It shows effectiveness in object removal, insertion, and replacement.

Strengths

- The slot-attention-based flow grouping has a novel and reasonable design. As demonstrated in various video/multi-view models, self-attention-based key and value concatenation is effective in ensuring view consistency. Overall, this paper proposes a well-structured recipe for multi-view editing. - The strategies used in the proposed method and their justifications are well explained in detail. - The video results and quantitative results demonstrate good performance.

Weaknesses

- Some assumptions are heuristic, such as in L232. While these assumptions may be reasonable in most cases, they might not be robust across diverse user cases. - Overall, the figures and flow of the paper are complicated. It would be beneficial if the content of the figures were made more comprehensive.

Questions

How is the generalizability? What is the performance in completely OOD cases?

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

the authors adequately addressed the limitations.

Reviewer jssh6/10 · confidence 3/52024-07-15

Summary

The paper studies the task of multi-view consistent real-world object removal and insertion, enabled by learning a model, MVInpainter, trained to perform multi-view 2D inpainting. The paper demonstrates its effectiveness on both object-centric and scene-level datasets with the task of object removal and object insertion.

Strengths

The task that the paper addresses is important yet less explored. Learning strong priors for scene level multi-view consistent editing can serve as a basis model for various 3D applications. The method introduces motion priors and optical flow to guide the generation, which seems to be interesting and effective with various quantitative and qualitative evaluations. The presentation is good and the paper is overall easy to read.

Weaknesses

1.Though the paper tackles the setting of video input (consecutive frames to obtain the motion prior), I am curious how the method performs with just multi-view inputs? Does the performance significantly drop? If the model can be applied to those as well, it could potentially lead to bigger impact and be applied to more real-world scenarios. 2.How are the priors different from AnimateDiff and Flow Grouping? Why are they complementary? 3.Could the authors please explain why the proposed method cannot be compared to SPIn-NeRF-type methods? Is it because of camera poses? If so, I am still curious to see how large the gap is? If the method is close to the methods with camera poses, it would make the paper stronger. 4.One limitation is that the method still works on images with a relative simple background, as pictured in the supplementary materials Figure 19.

Questions

I think the paper is overall of good quality and tackles an important task. Some questions (as listed) could be explained for a better understanding of the method.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Authorsrebuttal2024-08-12

Feel free to discuss any remaining questions or concerns

Thanks for your valuable feedback. We have carefully addressed all your concerns and provided the details in our rebuttal. Our models and codes will be open-released, including all details to reproduce our results. As the rebuttal deadline approaches, please feel free to discuss any remaining questions or concerns. We will try our best to answer your questions.

Authorsrebuttal2024-08-12

Feel free to discuss any remaining questions or concerns

Thanks for the valuable feedback from all reviewers. We have carefully addressed all concerns in our rebuttal. Our models and codes will be open-released, including all details to reproduce our results. As the rebuttal deadline approaches, please feel free to discuss any remaining questions or concerns. We will try our best to answer your questions.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC