acceptodds
Under review as a conference paper at ICLR 2027

OPERA: Object Perception Enhances Single-view 3D Reconstruction

Abstract

Single-view 3D reconstruction is a challenging task in computer vision due to information missing from the single input image. Generative model-based approaches can produce plausible 3D objects from a single image, thanks to data-driven priors learnt from rich and large-scale datasets. However, plausible generation does not guarantee fidelity to the geometry and appearance of the particular input object. Inspired by object perception in human vision, we propose OPERA, a framework that guides multi-view diffusion sampling with pretrained perception models through lightweight alignment modules. These modules are trained independently while both the generative and perception models remain frozen, allowing multiple signals to be combined at inference without joint fusion training. We apply OPERA to two state-of-the-art single-view 3D reconstruction baselines on two standard datasets. We also compare our method with recent image-to-3D models. We provide in-depth analyses of the design choices and their effects across datasets and backbones. Our project page is at https://opera-3d.github.io.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
single-view 3D reconstruction, diffusion-based multi-view generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_upPhOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces OPERA, a framework designed to improve single-view 3D reconstruction by integrating object-specific perceptual knowledge into the multi-view diffusion sampling process. The core idea is to employ Generation-Perception Alignment Modules (GPAMs) to map intermediate generative features from a diffusion model into the target feature spaces of pre-trained 2D perception foundation models. The gradients derived from these perception models are then utilized to guide the sampling process at test time. The authors demonstrate their method on top of existing baselines.

Critical strengths
  1. The paper is well-written. The architecture of the GPAMs and the overall guidance pipeline are clearly explained. The implementation provided in the appendix ensure its reproducibility.

  2. The idea of leveraging the generalized 2D priors from perception foundation models to constrain and improve 3D generation is well-motivated. Align disparate feature spaces using GPAMs is technically sound and well designed.

Critical weaknesses
  1. The framework is essentially a test-time "classifier guidance" mechanism. It relies on external 2D priors to patch existing baselines rather than solving root 3D reconstruction challenges.

  2. The gradient-based optimization at each diffusion step causes an impractical inference time (~4 minutes per object), making it uncompetitive with modern feed-forward models for real-world applications.

  3. The quantitative improvements, particularly on the Wonder3D baseline, are minor and fail to justify the massive computational overhead introduced by the guidance mechanism.

  4. The objects shown in the current figures lack geometric complexity. How does the framework perform on inputs with intricate structures or severe occlusions? Providing such examples would better validate the necessity of the perceptual guidance.

Questions for authors
  1. Could you provide a deeper comparison between OPERA and existing standard "classifier guidance" techniques? What makes the GPAM approach fundamentally different or superior to standard feature-matching guidance in diffusion models?

  2. The gains on Wonder3D are somewhat marginal. How does OPERA perform when applied to more recent, inherently more consistent 3D generation models? Does the guidance still provide meaningful improvements, or do the base model's native capabilities saturate the benefits of 2D perceptual guidance?

2 comments
AuthorsOpenReview ↗

Responses to Reviewer upPh's comments

W1. OPERA as a test-time classifier guidance mechanism

We respectfully clarify that OPERA is not simply an application of conventional test-time classifier guidance or an attempt to patch existing 3D reconstruction methods using external 2D priors.

A fundamental challenge in generative single-view 3D reconstruction is that the input image provides only partial observations of an object. Consequently, existing generative approaches can produce plausible but semantically or perceptually inconsistent views, leading to hallucinated structures and errors in the subsequent 3D reconstruction.

Our motivation is to address this limitation by leveraging the rich perceptual knowledge encoded in pretrained foundation models to improve the understanding of the object being reconstructed. Specifically, OPERA introduces GPAM to predict perceptual representations from intermediate diffusion features and uses these predictions to guide multi-view sampling.

Importantly, unlike conventional classifier guidance, which typically relies on gradients from a pretrained classifier with respect to a class-label objective, OPERA exploits foundation-model perceptual representations as guidance signals. Furthermore, GPAM learns to predict these representations during training, eliminating the need to run the external perception models during inference.

Therefore, OPERA is better understood as a perception-guided multi-view generation framework that enhances existing diffusion-based single-view 3D reconstruction pipelines, rather than a conventional classifier-guidance mechanism.

W2. Computational overhead

We would like to clarify that the reported inference time of approximately 4 minutes refers to the entire reconstruction pipeline, rather than the additional computational overhead introduced by OPERA.

Specifically, the baseline multi-view generation process using Era3D/Wonder3D takes approximately 1.8 minutes, while the subsequent 3D reconstruction using Wang et al. (2021) takes approximately 2 minutes. The additional perception-guided sampling introduced by OPERA takes only approximately 10 seconds.

Therefore, OPERA introduces a relatively small computational overhead compared with the overall reconstruction time. Importantly, the pretrained perception models are not required during inference, as GPAM directly predicts the perceptual representations from intermediate diffusion features.

We will clarify this breakdown in the revised manuscript to avoid confusion between the total reconstruction time and the additional overhead of our method.

W3. Improvements on Wonder3D

We acknowledge that the improvement margins vary across the two backbones. As discussed in our response to Reviewer 1, the effectiveness of perceptual guidance is influenced by the capabilities of the underlying generative backbone.

Era3D generally produces higher-quality and higher-resolution multi-view images than Wonder3D, providing richer visual details and more reliable geometric cues for subsequent reconstruction. Consequently, the two backbones have different baseline performance levels and different potential for improvement.

Nevertheless, OPERA is designed as a plug-in framework that can be applied to different diffusion-based multi-view generation backbones without fine-tuning the underlying generative models. The different improvement margins should therefore be interpreted in the context of the respective backbone capabilities, rather than as evidence that OPERA is inherently dependent on Era3D.

W4. Geometric complexity of evaluated objects

We would like to clarify that our evaluation is not limited to the relatively simple objects shown in the main paper. In particular, the 100 objects from OmniObject3D used in our evaluation include objects with more intricate geometric structures and varying levels of shape complexity.

To provide a more comprehensive assessment, we will include additional qualitative visualisations from the OmniObject3D evaluation, showing reconstructed objects from multiple viewpoints. These examples will help demonstrate the behaviour of OPERA on objects with more complex geometric structures and provide a broader view of its reconstruction capabilities.

AuthorsOpenReview ↗

Answers to Reviewer upPh's questions - updated

Q1. Differences from conventional classifier guidance

Conventional classifier guidance typically employs a pretrained classifier to provide gradients associated with a target class label during diffusion sampling. However, our task is fundamentally different: we aim to generate multiple perceptually consistent views from a single input image for subsequent 3D reconstruction, rather than generate a 2D image conditioned on a class label.

In particular, a class label provides only coarse semantic information and does not capture the detailed perceptual characteristics of the specific object being reconstructed. OPERA instead exploits richer perceptual representations extracted from pretrained foundation models.

Furthermore, extending perceptual guidance to multi-view generation requires addressing the relationships between representations across different views. As discussed in our response to Reviewer 1, direct feature matching using an L1/L2 objective is suitable for front-view guidance, but does not explicitly capture cross-view representational relationships. We therefore introduce CKA-based alignment to support multi-view perceptual guidance.

Another important distinction is the role of GPAM. Rather than requiring the external perception models to provide guidance gradients at every sampling step, GPAM learns to predict the corresponding perceptual representations from intermediate diffusion features. The perceptual targets are extracted once from the input image and remain fixed throughout sampling, avoiding repeated evaluation of the pretrained perception models.

Thus, the contribution of OPERA is not merely the use of an external classifier or feature-matching loss, but the formulation of a foundation-model-driven perceptual alignment and guidance framework for multi-view generation.

Q2. Generalisation to more recent 3D generation models

Please refer to our response to Reviewer 1 regarding the differences in improvement margins between Wonder3D and Era3D.

We appreciate the reviewer's question about whether perceptual guidance can benefit more recent 3D generation models with stronger native capabilities. To investigate this, we have conducted an additional experiment by integrating our proposed perceptual guidance into the structure flow-matching sampling process of TRELLIS.2. We evaluate the improvement on an evaluation set of 1,250 objects from Toys4K, the benchmark dataset used in the original TRELLIS.2 paper.

The results demonstrate consistent improvements across all four evaluation metrics. The table reports paired comparison p-values against TRELLIS.2.

Setting[email protected] ↑pPSNR ↑pSSIM ↑pLPIPS ↓p
TRELLIS.20.416—23.89—0.9436—0.0728—
TRELLIS.2 + Ours0.4270.000124.06<0.00010.94410.01320.0702<0.0001

In particular, our method improves geometric F1 from 0.416 to 0.427, while also improving PSNR, SSIM, and LPIPS. These results are important because they demonstrate that the benefits of perceptual guidance are not restricted to the diffusion-based multi-view generation pipelines originally investigated in OPERA. Instead, the proposed approach can also be extended to flow-matching-based 3D generation, improving both geometric reconstruction and rendering quality.

This experiment provides additional evidence that foundation-model perceptual guidance is complementary to the native generative capabilities of more recent 3D generation models, rather than having its benefits saturated by stronger generative backbones.

Reviewer_u2UyOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper aims at the challenge that 3D reconstruction from a single image usually do not take actual image cues into account and might not match the image (but focus on being a good reconstruction). The paper proposes to overcome this by monocular depth cues and semantic cues. The core idea of the reconstruction fitting the image better was previously explored in Pixal3D which is not mentioned in the manuscript https://github.com/TencentARC/Pixal3D be feasible

Critical strengths

I like the idea of semantic perception, I can however not judge how novel it is (perhaps the language model part, but not the Dino part).

Critical weaknesses

I had a very hard time trying to understand what the paper is actually doing, a core issue from my perspective is the conciseness of the writing and the quality of the figures. The figure s do look very nice, but I do not understand them. In the pipeline figure it is not clear what the semantic perception is, if it is a segmentation or something else? the depth perception also does not look like a monocular depth map and the 3D reconstruction is rendered from the same view as the input. This figure looks very nice but it does not really transport a message. Figure 2 shows training and inference, the method part starts with X from y and refers to this figure. X however does not show up in the figure. The image success that GPAM has not seen the Unet output during training. And there is no flow from the perception module down to GPAM during inference. I understood that differently from the paper.

The method part is written very technical and jumps right into detail - for me hard to get into what is actually performed.

There are heavy logical jumps in the method part (e.g. "This background parameterisation is distinct from the DDIM sampling updates used below. The generated normal maps provide geometric supervision for surface reconstruction, while colour images supervise appearance.") which makes it unreadable. A separate figure for what can be a GPAM might be helpful.

The experiments are mainly demonstrated in the appendix Qualitative results of the ablations are missing Quantitative results against trellis in table 10 might be an unfair comparison as trellis does not fit to image space

The experimental evaluation is limited to very few objects. Also on the project page the same few objects are shown. Visual results in Figure 9 are the only qualitative comparison against state of the art and I would judge that trellis is superior for most cases. The book example also seems cherry picked, as for the proposed method the view from the side looks actually very bad (see project page).

Questions for authors

I would propose to not only render the evaluation set completely in the appendix, but also the reconstructions for each of the datapoints (since the evaluation set takes less than a page) as one concern on my side is the very few shown reconstruction examples that do not look that strong on the project page on close inspection

2 comments
AuthorsOpenReview ↗

Responses to Reviewer u2Uy's comments

We thank the reviewer for the detailed feedback. We acknowledge that some aspects of the presentation, particularly the pipeline illustrations and the organisation of the methodological description, could be improved to make the proposed approach more accessible. However, we believe that some of the concerns arise from misunderstandings of the proposed framework, and we would like to clarify several important aspects below.

1. Methodological understanding and figures

  • Semantic and depth perception: The semantic perception represents high-level semantic features extracted using the pretrained vision-language model, rather than a semantic segmentation map. Similarly, the depth perception represents monocular depth information estimated by the pretrained depth model. We will revise Figure 1 to explicitly identify these representations and improve their visualisation so that their roles are more apparent.

  • 3D reconstruction visualisation: The reconstruction shown from the input viewpoint is intended to illustrate the reconstruction pipeline and its consistency with the input image, rather than demonstrate novel-view rendering. We will improve the figure by including additional viewpoints to better communicate the reconstructed 3D geometry.

  • Training and inference pipeline: We agree that Figure 2 could better illustrate the information flow in GPAM. In particular, the intermediate features from the diffusion U-Net are provided to GPAM, which is trained to predict the corresponding perceptual representations using supervision from the pretrained perception models during training. At inference time, all models are frozen; the alignment between GPAM's output and the input perceptual signal is used for perception-guided sampling. X is the 3D reconstructed object. We will revise Figure 2 to improve its clarity. We will also correct the missing notation in the figure.

2. Methodological presentation and readability

We appreciate the suggestion to improve the accessibility of the method section. Our intention was to provide sufficient technical details to ensure reproducibility, but we agree that a clearer introduction to the overall framework before presenting the mathematical formulation would improve readability.

We will revise the method section to introduce the motivation and overall operation of GPAM first, followed by its architecture, training objectives, and integration into diffusion sampling. We will also improve the transitions between the different technical components, particularly the background parameterisation, DDIM sampling, and the subsequent 3D surface reconstruction. These components serve different purposes in the overall pipeline, and we will make their relationships more explicit.

3. Experimental evaluation and comparisons

We would like to clarify that our evaluation against recent 3D generation methods is not restricted to the 30 GSO objects. As reported in Table 10, we additionally evaluate the methods on another set of 100 objects from OmniObject3D.

Regarding the comparison with TRELLIS.2, we acknowledge that the two approaches follow different reconstruction pipelines. OPERA improves diffusion-based multi-view generation, whereas TRELLIS.2 follows a different 3D generation pipeline. Therefore, the comparison should be interpreted in the context of their respective objectives and reconstruction pipelines, rather than as a direct comparison of identical optimisation procedures. We report both rendering-based metrics and geometric F1 to provide complementary perspectives on reconstruction quality. In particular, OPERA achieves better PSNR, SSIM, and LPIPS, while TRELLIS.2 achieves better geometric F1. We do not claim that OPERA produces superior geometry in all aspects.

We will also improve the presentation of the experimental results by bringing more of the relevant findings into the main paper and providing additional qualitative ablation results in the supplementary material.

4. Qualitative results and potential cherry-picking

We clarify that the qualitative examples currently presented in the paper and on the project page are selected for visualisation purposes and are not intended to represent the entire evaluation set. We agree that providing a more comprehensive visual evaluation would help readers better assess the strengths and limitations of the proposed method.

To address this concern, we will include a comprehensive visualisation of the reconstructions for the evaluation objects, including different viewpoints, in the supplementary material. This will allow readers to inspect the results across the evaluation set rather than relying on a small number of selected examples. We will also include additional qualitative comparisons and ablation results to provide a more complete assessment of OPERA.

Thank you for adressing all my comments. The suggested changes are from my perspective a major revision of the paper and the paper should go through review again after such substantial changes. I hope the feedback helped to improve the work, but I would still recommend to reject the paper from ICLR.

On a different note: the discussion phase is almost 10 days long. Releasing a response to the reviewers after over 90% of the time passed is from my perspective bad practice and I would suggest to consider releasing your response for future venues earlier. That way you could reduce impact on the private life of reviewers that want to do their reviewing duties properly. I also would like to mention that it did not change my assessment of your rebuttal based on this aspect.

Reviewer_fH5YOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents OPERA, which is a perception-guided framework for single-view 3D reconstruction. The key idea is to leverage pretrained perception models, including DINO, Depth Anything, and vision-language models, to improve the guidance for multi-view diffusion sampling. The authors train lightweight alignment modules (GPAMs) to predict perceptual features from intermediate diffusion features, and use the alignment losses to guide the sampling process. The generated multi-view images and normal maps are then used for 3D reconstruction. Experiments on Era3D and Wonder3D using GSO and OmniObject3D show substantial improvements on Era3D, while the gains on Wonder3D are relatively limited.

Critical strengths
  • The proposed method shows that it can be integrated into existing multi-view diffusion pipelines without modifying the pretrained generative models. The independent training of alignment modules makes it relatively easy to incorporate different perceptual signals.
  • The method achieves noticeable improvements on Era3D, particularly in terms of geometric reconstruction accuracy. The qualitative results also show some improvements in object shapes and appearance details.
Critical weaknesses
  • Limited methodological novelty. The main idea of predicting perceptual features from intermediate diffusion features and using them to guide sampling is quite similar to Readout Guidance. The main differences seem to be the extension to multi-view reconstruction and the use of different perceptual targets and alignment losses. Overall, this feels more like an application of an existing approach rather than a substantial methodological contribution.
  • Limited improvements across different backbones. While OPERA shows clear improvements on Era3D, the gains on Wonder3D are much less convincing. This makes me wonder how well the method generalizes across different generative backbones, and whether its effectiveness depends heavily on the underlying model.
  • Insufficient experimental comparisons. I would like to see comparisons with simpler guidance strategies, such as directly applying perceptual losses to predicted images or adapting Readout Guidance to multi-view generation. Although the authors compare with recent image-to-3D methods in Appendix F, OPERA still performs worse in terms of geometric F1. The evaluation is also somewhat limited, especially with only 30 objects from GSO.
Questions for authors
  1. How does OPERA compare with a straightforward adaptation of Readout Guidance to multi-view diffusion? It would be helpful to better understand what the proposed GPAM and alignment objectives bring beyond existing guidance methods.
  2. Why does OPERA show much smaller improvements on Wonder3D than on Era3D? Does the effectiveness of the method depend on the choice of backbone? Some discussion of these differences would be helpful.
  3. How much does the proposed view-specific alignment actually help? Appendix C.2 shows that removing the auxiliary CKA terms gives very similar reconstruction results. Since the pretrained backbone already models cross-view interactions, it is unclear how much additional benefit the multi-view alignment provides over front-view guidance alone.
  4. Table 10 shows that OPERA falls behind TRELLIS.2 and SAM 3D Objects in geometric F1, despite achieving better rendering metrics. Could the authors clarify whether this advantage in rendering quality mainly comes from the image-based reconstruction pipeline rather than improved 3D geometry? Given these results, what are the main advantages of OPERA over more recent 3D generation methods?
3 comments
AuthorsOpenReview ↗

Responses to Reviewer fH5Y's comments

W1. Limited methodological novelty

We clarify that the contribution of OPERA is not simply the application of Readout Guidance to multi-view generation. Rather, our work investigates how perceptual knowledge learnt by pretrained foundation models can be integrated into diffusion-based single-view 3D reconstruction to improve the quality and consistency of generated views.

While Readout Guidance provides a mechanism for extracting and exploiting intermediate diffusion representations, extending this idea to multi-view generation introduces additional challenges. In particular, perceptual guidance must account for the different characteristics of individual views while maintaining cross-view consistency, rather than simply guiding the generation process towards a perceptual target extracted from a single image. To address these challenges, we introduce the Generation–Perception Alignment Module (GPAM), which learns to predict foundation-model perceptual representations from intermediate diffusion features. We further investigate appropriate perceptual targets, alignment objectives, and the integration of perceptual guidance into multi-view sampling.

Therefore, the contribution of OPERA lies in the formulation and implementation of a perception-guided multi-view generation framework for single-view 3D reconstruction, rather than the introduction of Readout itself. To the best of our knowledge, OPERA is the first work to explore this direction in diffusion-based single-view 3D reconstruction.

W2. Limited improvements across different backbones

We agree that the improvement varies between the two backbones. However, the effectiveness of perception guidance is naturally influenced by the capabilities of the underlying generative backbone.

As also demonstrated in the original papers, Era3D generally produces higher-quality multi-view images than Wonder3D. In particular, its ability to synthesise higher-resolution views provides richer visual details and more reliable geometric cues for subsequent 3D reconstruction. Consequently, the two backbones have different baseline performance levels and potentially different room for improvement.

Importantly, OPERA is designed as a plug-in perception-guidance framework rather than a backbone-specific architecture. Its GPAM can be trained while keeping the generative backbone frozen, allowing the approach to be applied to different diffusion-based multi-view generation models. The different improvement margins should therefore be interpreted in the context of the underlying backbone capabilities, rather than as evidence that OPERA is inherently dependent on Era3D.

W3. Insufficient experimental comparisons

We would like to clarify that the experiments in Table 8 of the Appendix include alternative perceptual guidance configurations, where perceptual features from Qwen2.5-VL + T5 or DINOv3 are fused using front-view guidance. These experiments provide comparisons against simpler guidance configurations that exploit perceptual information without the proposed view-specific alignment strategy. The results demonstrate the importance of incorporating perceptual information into the multi-view generation process and the benefits of our proposed alignment strategy over these simpler configurations.

Regarding comparisons with recent image-to-3D methods, we acknowledge that OPERA does not achieve the best geometric F1 score among all compared approaches. Nevertheless, OPERA achieves substantially better multi-view rendering metrics (PSNR, SSIM, and LPIPS) than TRELLIS.2, as reported in Table 10. This distinction is important because OPERA primarily improves the quality of intermediate multi-view image generation, which subsequently benefits image-based 3D reconstruction.

Regarding the evaluation scope, we would like to clarify that our comparison with TRELLIS.2 is not limited to the 30 objects from GSO. As reported in Table 10, we additionally evaluate the methods on 100 objects from OmniObject3D, providing a broader evaluation across two different object datasets. We believe this additional evaluation provides a more comprehensive assessment of OPERA against recent 3D generation methods.

AuthorsOpenReview ↗

Answers to Reviewer fH5Y's questions

Q1. Comparison with a straightforward adaptation of Readout Guidance

Please refer to our response to W3. In particular, the alternative configurations in Table 8, using Qwen2.5-VL + T5 or DINOv3 with front-view guidance, provide comparisons with simpler ways of incorporating perceptual information into multi-view generation.

The key difference is that these configurations primarily rely on perceptual guidance from the input/front view, whereas OPERA introduces GPAM and view-specific alignment to exploit perceptual information in the multi-view generation process. This allows the guidance mechanism to account for the perceptual characteristics of individual generated views, rather than relying solely on the front-view representation. The improvements reported in Table 8 demonstrate the benefits of the proposed design over these simpler configurations.

Q2. Why does OPERA show different improvements on Wonder3D and Era3D?

Please refer to our response to W2. The improvement margins are influenced by the underlying generative backbone, particularly its multi-view synthesis capabilities and baseline reconstruction quality. Era3D generally produces higher-resolution and higher-quality multi-view images than Wonder3D, which also affects the effectiveness of subsequent perception-guided reconstruction.

Importantly, OPERA is designed to be applicable to different diffusion-based multi-view generation backbones, without requiring the backbone itself to be fine-tuned.

Q3. Contribution of view-specific alignment

We would like to clarify that extending perceptual guidance from a single front view to multiple views introduces an important challenge in feature alignment. For front-view guidance, a direct L2-std loss can be used to align the predicted perceptual features with the corresponding target features (see Eq. (5)). However, when multiple views are considered during sampling, directly applying L2-std loss to individual view features is not appropriate, as it does not explicitly account for the relationships between different views.

To address this challenge, we employ CKA-based alignment to capture the relationships between perceptual representations across multiple views, enabling more effective multi-view perceptual guidance. This is an important distinction between the proposed multi-view alignment and straightforward front-view guidance.

Regarding the ablation in Appendix C.2, we clarify that the row "Spatial, front only" implies the application of perceptual guidance to the front-view generation only. It does not indicate that multi-view alignment itself is unnecessary.

Q4. Comparison with TRELLIS.2 and SAM 3D Objects

We agree that OPERA does not outperform TRELLIS.2 and SAM 3D Objects in geometric F1. It is important, however, to distinguish the objectives and reconstruction pipelines of these approaches.

Our approach builds upon diffusion-based single-view reconstruction pipelines such as Era3D and Wonder3D. These methods first synthesise multiple views conditioned on the input image and subsequently reconstruct a 3D object using the reconstruction method of Wang et al. (2021). OPERA improves the perceptual quality of the generated multi-view images through foundation-model guidance, thereby improving the quality of the subsequent reconstruction and rendering.

The results in Table 10 reflect this distinction: OPERA achieves better multi-view rendering metrics (PSNR, SSIM, and LPIPS) than TRELLIS.2, although its geometric F1 remains lower. We therefore do not claim that OPERA produces superior geometric reconstruction to these recent 3D generation methods.

Instead, the main advantages of OPERA are its ability to improve existing diffusion-based single-view reconstruction pipelines through foundation-model perceptual guidance, its flexibility in incorporating different pretrained perception models, and its training efficiency. Specifically, only GPAM needs to be trained, while the underlying diffusion backbone can remain frozen. This makes OPERA a flexible and relatively lightweight approach for improving existing multi-view generation pipelines.

Would you mind comnenting with your response to the review you are adressing?

There is very limited time left for your reviewers to read your responses. In half of the world the workday ended. So please make sure it is clear which reviewer you adress

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.