MVDiffusion: Enabling Holistic Multi-view Image Generation with Correspondence-Aware Diffusion

This paper introduces MVDiffusion, a simple yet effective method for generating consistent multi-view images from text prompts given pixel-to-pixel correspondences (e.g., perspective crops from a panorama or multi-view images given depth maps and poses). Unlike prior methods that rely on iterative image warping and inpainting, MVDiffusion simultaneously generates all images with a global awareness, effectively addressing the prevalent error accumulation issue. At its core, MVDiffusion processes perspective images in parallel with a pre-trained text-to-image diffusion model, while integrating novel correspondence-aware attention layers to facilitate cross-view interactions. For panorama generation, while only trained with 10k panoramas, MVDiffusion is able to generate high-resolution photorealistic images for arbitrary texts or extrapolate one perspective image to a 360-degree view. For multi-view depth-to-image generation, MVDiffusion demonstrates state-of-the-art performance for texturing a scene mesh.

Paper

Similar papers

Peer review

Reviewer VE1c7/10 · confidence 4/52023-07-04

Summary

This paper proposes a new method to generate multi-view images, ensuring pixel-to-pixel multi-view consistency. The multi-view consistency is guaranteed by the correspondence aware module, which is utilized in the latent diffusion multi-image generation process. The proposed method outperforms previous works for tasks related with multi-view image generation, especially in the multi-view consistency.

Strengths

1. The idea to guarantee pixel correspondence in the diffusion process is novel and reasonable. 2. The supplementary material provides codes, which show some details of the implementation. 3. The proposed method may improve the results for text to texure given meshes. 4. For different view images, the user can input different prompt.

Weaknesses

1. To obtain the pixel correspondence for multi-view images, the proposed method requires to input the depth maps. 2. The authors do not explain how to deal with the depth occlusion and how to deal with the situation where the pixels in target images are not in the view of the source image.

Questions

1. How does the proposed method perform for objects? Could we use the proposed method to generate textures given object meshes? 2. How does the authors deal with the depth occlusion and the situation where the pixels in target images are not in the view of the source image?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors addressed the limitations and the work does not have negative societal impact.

Reviewer xhAT6/10 · confidence 5/52023-07-06

Summary

This paper introduces MVSDiffusion, a diffusion framework that generates multi-view images with content consistency. The problem setting is interesting and essential in practice. The proposed correspondence-aware attention mechanism provides cues for multi-view consistency.

Strengths

1. The paper proposes a novel problem setting: generating images for multiple camera poses with consistent contents. 2. The paper proposes a multi-view diffusion technique to obtain content consistency: extract features from other views guided by depth maps. 3. Panorama generation can also be improved via the multi-view diffusion framework.

Weaknesses

1. The MVDiffusion relies on given high-quality multi-view depth maps. However, the depth map obtained for leaves and strings may lack details, which may impact the generated image quality. 2. The correspondences are derived from wrapped depth so that it can only support rigid static scenes. 3. The interpolation module is time-consuming but the time consumption of each module is not provided. Please show them. 4. The tailored super-resolution module is not novel.

Questions

1. The presented panorama is blurry compared with SD (pano). Why? 2. From my point of view, MVDiffusion is able to generate multi-view images and thus can synthesize images in 3D novel views instead of pure rotations (panorama). But all the presented image sequences are almost pure rotation. What's the problem?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

1. The multi-view image generation relies on given geometries. 2. Regions that are not captured by the pre-defined camera poses cannot be recovered. 3. As the framework needs to interpolate camera poses between keyframes. The time consumption may be quite large.

Reviewer x9w47/10 · confidence 3/52023-07-06

Summary

This paper presents MVDiffusion, a new diffusion model to generate consistent multi-view images, e.g., panorama. The authors propose a novel correspondence-aware attention mechanism in order to enforce pixel-level correspondence and cross-view consistency. More specifically this mechanism is used in three modules: a generation module that generates consistent low-resolution multi-view images; an interpolation module that generates images in between; a super-resolution module. After injecting such module into Stable Diffusion Unet layers and fine-tuning the model on multi-view image dataset, the model can synthesize consistent multi-view images based on text or depth.

Strengths

1. The main idea (improving multi-view consistency through a correspondence-aware attention mechanism) is novel, simple, effective and easy to understand. The generated results are very impressive, see figure 1, 4, 5, 6 and all the images in the supplementary. 2. The paper is well-written especially the method part where the main design is demonstrated clearly, see section 4.1 and figure 3. The authors also provide the code in the supplementary. 3. The comparison with previous methods and some straightforward baselines are comprehensive and the improvements are convincing.

Weaknesses

1. The pipeline figure (Figure 2) could be further improved: this figure is supposed to make the audience understand the whole workflow without looking at the method text, however, for now, in this figure, it's a bit unclear to me how these modules are associated with each other at first glance. 2. Adding a brief section on failure cases would be great: the results in the main paper and supplementary are truly impressive, however, I'm also curious what are the common failure cases of MVDiffusion. Adding a short paragraph discussion this would be beneficial to the whole community. 3. I feel a bit confused about the panorama generation during inference time: in Line 145, it says "The generation module generates eight 256 × 256 images, and the super-resolution module upscales to eight 1024 × 1024 images". I am just wondering where's the interpolation module then? Or let me put this way, how many images are generated by the generation module and how many are generated by the interpolation module? I assume the key frames are generated by the generation module and more images in between are generated by the interpolation module? 4. How to handle the conflicts in the overlap region between two generated images? If I understand correctly, although correspondence attention works very well, it still cannot guarantee the overlap region between two generated images are *exactly* matched, so I'm just curious how does this mismatch is handled? I asked this because in the stitched 8 perspective images, it looks very consistent. 5. Three modules are proposed in this paper, it would be great if an ablation study could be conducted such that the audience could understand the influence of each individual module better, e.g., how much difference it would make if removing the interpolation module? 6. In addition to the stitch image, it would be great if showing more *densely* (much more than 8) generated multi-view images and make it an animation, such that the multi-view consistency could be better evaluated qualitatively by the audience.

Questions

It would be great if the authors could respond the points mentioned above. Thanks!

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer 2y4s5/10 · confidence 3/52023-07-10

Summary

The paper proposes a multiview latent diffusion model that is aware of the correspondence between views. Equipped with correspondence-aware attention blocks, the proposed generation module, interpolation module and the super-resolution module help MVDiffusion outperforms existing works.

Strengths

The proposed Correspondence-Aware Attention block connects different views. Also, keyframes are first generated and then upsampled both spatially and temporally.

Weaknesses

My biggest concern of the paper is the evaluation of measuring multi-view consistency. On L217 the authors mentioned to use pixel-level similarity for multi-view consistency. Is the PSNR computed between generated image and ground truth image? Does that mean the generation of depth-to-image is deterministic? It might be a good idea to conduct some evaluation in 3D. For example, train a NeRF with generated images and evaluate the difference between rendered image and the MVDiffusion generated image. Another way might be calculating some reprojection error since we know the camera poses of the generated images. Also, for panorama generation, the author only compares with Stable Diffusion. Would be nice if DiffCollage or MultiDiffusion is also considered for comparison.

Questions

Please refer to weakness.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Please refer to weakness.

Reviewer x9w42023-08-17

thanks for the detailed response

Thanks a lot to the authors for the detailed response! I don't have more questions but just curious where I am able to see the video? (I think this is supposed to be a question for the area chair instead of the authors then).

Authorsrebuttal2023-08-17

I have informed AC to send the link.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC