acceptodds
Under review as a conference paper at ICLR 2027

JointEdit3D: Feed-Forward 3D Scene Editing in a Unified Latent Space

Abstract

Existing 3D scene editing methods typically rely on per-scene optimization over explicit 3D representations or cascaded edit-and-reconstruct pipelines, resulting in high test-time cost, limited 3D awareness, and structural inconsistencies. To couple appearance synthesis with geometry prediction, we adapt a pretrained unified RGB-geometry latent space to feed-forward scene editing. Given a source video and an edited reference image, performs asymmetric latent inpainting: it observes only the edited RGB reference latent and jointly generates the remaining RGB latents and the entire geometry latent along the source trajectory. JointEdit3D introduces a dedicated SceneAnchor Branch to inject source-scene structure without forcing direct copying, and adopts edit/background-aware losses to balance edited-region fidelity with unedited-content preservation. To address the lack of paired resources for standardized 3D scene editing evaluation, we introduce , a dataset with 15K paired editing samples and renderer-provided 3D annotations, together with , a curated 100-sample benchmark. Experiments show that JointEdit3D improves edited-region quality and 3D structural completeness over prior baselines while maintaining competitive background preservation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Scene Editing, Feed-Forward Generation, Latent Diffusion Models, 3D Reconstruction.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qXVaOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces JointEdit3D, a 3D editing method which jointly generates edited appearence and geometry rather than editing views first and reconstructing afterward. The input of this method is a source video and a single edited reference frame. Then, it uses an asymmetric RGB-geometry latent inpainting process to obtain edited appearance and geometry of the scene. This paper also introduces SceneEdit3D-15K, a dataset of 15K paired scene edits with renderer-provided 3D annotations, and a benchmark SceneEdit3D-Bench for evaluation. Results show that the proposed method achieves better results than exisiting methods, in terms of PSNR and LPIPS.

Critical strengths
  1. The idea of editing the appearence and geometry in one single feed-forward pass is much better then previous optimization-based methods.

  2. This paper features a video for introduction, the results are clearly demonstrated, good results.

  3. The contribution of SceneEdit3D-15K dataset could be beneficial to the community.

Critical weaknesses
  1. All of the editing shown in the paper focuses on object deletion, add or move, which makes the paper more similar to a 3D inpainting paper rather than an editing paper. Results involve color-changing, like day to night, spring to fall, or changing from one object to another, would be interested to see in the paper.

  2. The results only provides PSNR and LPIPS. More metrics, including CLIP metrics to see if the editing is aligned with text prompt, FID, time, etc, are expected. Meanwhile, qualitative results are only on the self-constructed SceneEdit3D-Bench, with some inpanting benchmark in the supplementary.

  3. Given that many of test cases in this paper are inpainting, the following paper would be interesting to compared with:

You, Junqi, et al. "Instainpaint: Instant 3d-scene inpainting with masked large reconstruction model." Advances in Neural Information Processing Systems 38 (2026): 153574-153594.

4.It's not critical, but the reference annotation in this paper is black, the same color as text, which makes it difficult to read.

Questions for authors
  1. I would be interested to know the (1) time (2) results on large appearance related editing for this work.

  2. The pipeline includes the editing with one reference image and a source unedited video. What if we use a video editing model to edit the entire video?

  3. Since the input of the model is source video, it doesn't perfectly aligned with the idea of "3D editing". What if the given images are not in a time order?

5 comments
AuthorsOpenReview ↗

Response to Reviewer qXVa (1/2): Editing Scope, Metrics, and Runtime

We thank the reviewer for the constructive feedback and for recognizing our joint appearance-geometry editing and SceneEdit3D-15K. All new results have been added to Appendix F of the revised PDF.

W1 & Q1. Editing scope and large appearance changes.

We respectfully clarify that JointEdit3D targets object-level 3D scene editing rather than 3D inpainting, which typically refers to filling the region of a removed object, as in SPIn-NeRF and AuraFusion360. Beyond removal, JointEdit3D handles addition, relocation, appearance change, and multi-operation edits, which require generating new content with consistent geometry and occlusion. This scope follows the closest feed-forward 3D editing methods, MVInpainter and Omni-3DEdit, which target the same family of object-level operations. Accordingly, appearance editing in our paper refers to object-level changes such as recoloring or retexturing a specific object (Table 3, Figure 9).

Following the reviewer's suggestion, we added object replacement and more object-level appearance edits (Figures R1 and R2). JointEdit3D keeps the new content consistent across views while preserving the surrounding scene.

Global changes such as day-to-night or seasonal transitions belong to scene stylization, a separate line of work represented by Instruct-NeRF2NeRF and EditCast3D. Our training data contains no such edits, yet zero-shot tests already transfer these changes across views, although coverage and lighting are not yet fully consistent (Figure R2(b)). Our Blender pipeline can render paired stylization data, and adding it to training is a natural direction for future work.

W2 & Q1. Metrics, runtime, and real data.

Several of these results are already in the paper. They include SSIM and depth metrics (Table 7), point-cloud and region-wise Chamfer metrics (Table 4), and runtime (Table 10). JointEdit3D takes 11.94 s per scene, compared with 43 s for MVInpainter, 237 s for Omni-3DEdit, and 285 to 1619 s for optimization-based methods. Every edit, including global changes such as the stylization in Figure R2(b), runs through the same feed-forward pass with a fixed number of denoising steps, so the runtime is essentially the same regardless of the edit type or how much of the scene changes.

Following Omni-3DEdit, we further report CLIP metrics and FID (Table R5). JointEdit3D performs best on all four metrics, with a much lower FID (14.27 vs. at least 29.45).

MethodCLIP-T↑CLIP-dir↑CLIP-I↑FID↓
SEVA0.22770.06080.935546.24
MVInpainter0.22060.08260.938229.45
Omni-3DEdit0.22990.05410.929747.75
InstaInpaint0.21030.04620.930548.15
JointEdit3D0.23550.08790.961914.27

Table R5. CLIP metrics and FID on SceneEdit3D-Bench.

For real data, the paper already reports quantitative results on 360-USID (Table 2) and Inpaint360GS (Table 12), and we now add SPIn-NeRF (Table R1). Qualitative real-scene results are also part of the paper, on Instruct-NeRF2NeRF and DL3DV captures (Figure 4), real videos (Figure 5), and DAVIS, ScanNet++, and casual captures (Figures 9–12). We have added further real-scene examples in Figures R1, R2, R4, R5, and R6.

AuthorsOpenReview ↗

Response to Reviewer qXVa (2/2): InstaInpaint, Video Editing Models, and Input Format

W3. Comparison with InstaInpaint.

We evaluated InstaInpaint on SPIn-NeRF, giving all methods the same real object-free reference (Table R1, in our response to Reviewer rarN). JointEdit3D is better on all metrics except SSIM (e.g., M-PSNR 23.12 vs. 21.73, M-LPIPS 0.266 vs. 0.285), while requiring neither masks nor camera poses. InstaInpaint is designed mainly for object removal. On the more complex addition, relocation, appearance, and multi-operation edits of SceneEdit3D-Bench, it fails on a large number of edits, with Edit PSNR of 13.27 to 16.91 compared with 23.19 to 27.67 for JointEdit3D (Table R7), and it does not place new content in other views (Figure R4). Changing or adding InstaInpaint's input views does not close the gap (Edit PSNR at most 25.87 vs. 31.92, Table R6).

W4. Citation color.

Thank you for pointing this out. The default hyperref link boxes may not be displayed in some PDF viewers, so we have switched to colored citation text in the revision.

Q2. Editing the entire video with a video editing model.

Editing the video first and reconstructing it afterwards is the cascaded paradigm that JointEdit3D is designed to avoid. We added two first-frame-guided video editors, I2VEdit and AnyV2V, followed by VGGT reconstruction (Table R3, in our response to Reviewer rarN). Errors from the video editing stage carry into the geometry and raise CD to 0.1254 or higher (vs. 0.0976). Since these editors have no training mechanism for preserving unedited content, the overall color and style of the edited videos also drift from the source as the video progresses, and local errors such as a ghost of the removed object appear (Figure R6).

Even without its 3D output, JointEdit3D is a strong reference-guided video editor. Its RGB outputs in Table R3 are directly comparable with those of the video editors, and JointEdit3D achieves higher edit and background quality than the better of the two (Edit PSNR 26.54 vs. 22.99, Bg. PSNR 32.73 vs. 23.99), without the per-clip LoRA training that I2VEdit requires. The Table 5 variant "w/o GEO in Wan Flow (+VGGT)" is an RGB-only video editor trained on the same data, and it is worse in appearance as well as geometry (Edit PSNR 24.75 vs. 26.82, Bg. PSNR 27.85 vs. 30.69, CD 0.1506 vs. 0.1053), suggesting that joint geometry generation also benefits appearance editing. JointEdit3D is also more flexible in how the edit is given. First-frame-guided video editors such as I2VEdit and AnyV2V take the edit only at the first frame, so a reference elsewhere requires reordering the video (Figure R6), whereas JointEdit3D accepts the edited reference at any frame of the clip (Appendix B.1) and multiple edited references at inference (Figure 11), without fine-tuning.

Q3. Video input and unordered images.

The "video" here denotes ordered multi-view observations of a static scene, and the output is 3D geometry. JointEdit3D is designed for such input, which lets it build on a video generation prior and operate without camera poses. Our current scope is views organized along a camera trajectory. Fully unordered image collections would need to be ordered first, and we have not evaluated such reordering or very sparse, wide-baseline inputs.

We hope these clarifications address the reviewer's concerns. If so, we would be grateful if the reviewer would consider updating the score, and we are happy to discuss any remaining questions.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

We would like to follow up on our response above. We clarified the distinction between object-level editing and global stylization, added zero-shot stylization results and further editing examples, and provided CLIP/FID evaluations and comparisons with InstaInpaint and video-edit-then-reconstruct baselines (Appendix F). We also highlighted the geometry, depth, runtime, and real-scene evaluations already included in the original manuscript.

We would appreciate your feedback on whether these address your concerns and would be happy to clarify any remaining issues. Thank you for your time and consideration.

Best regards,

The authors

Reviewer_FV7AOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces JointEdit3D, a novel approach to 3D scene editing within a feed-forward framework. To address the high per-scene optimization cost and the inconsistencies arising from appearance–geometry disentanglement in existing methods, the authors formulate reference-image-guided 3D editing as asymmetric latent inpainting in a unified RGB–geometry latent space. The model observes only the RGB latent of a single edited reference frame and simultaneously generates the RGB latents of the remaining frames along the source video trajectory as well as all geometry latents. To preserve structural consistency in unedited regions, the paper designs a SceneAnchor Branch that injects source-scene image features into the main generative network in a residual manner. It further introduces Edit-Aware Losses that explicitly separate edited from background regions in the latent space and assign them different weights. In addition, the authors construct SceneEdit3D-15K, the first dataset comprising 15K paired editing samples with renderer-annotated ground truth, together with an evaluation benchmark, SceneEdit3D-Bench. Experiments show that the proposed method outperforms existing baselines in edited-region quality, 3D geometric integrity, and inference speed.

Critical strengths

S1. Novel Formulation and Architecture:

The paper unifies 3D scene editing in a joint RGB–geometry latent space, thereby avoiding the cascaded errors of the conventional two-stage pipeline. The proposed SceneAnchor Branch adopts an interleaved residual fusion mechanism that enables source-scene conditioning to adapt dynamically to the current denoising state, achieving high-quality background preservation without requiring input masks.

S2. High-Quality Data Infrastructure and Benchmark:

SceneEdit3D-15K fills the gap in the 3D scene editing community for paired editing datasets with 3D ground-truth annotations. The test split covers deletion, addition, movement, appearance modification, and multi-operation compositions, providing comprehensive evaluation dimensions.

S3. Extensive Experiments and Strong Results:

The method demonstrates clear advantages in both quantitative metrics and qualitative results, with highly efficient feed-forward inference. The ablation studies are rigorously designed and substantiate the necessity of each core module.

Critical weaknesses

W1. Domain Shift / Synthetic-to-Real Generalization:

SceneEdit3D-15K is built on synthetic scenes from Blender. Although Table 2 and Figure 4 demonstrate transfer to real-world datasets (e.g., 360-USID and Instruct-NeRF2NeRF), the stability of the geometry latent space may be challenged under realistic conditions involving complex lighting, depth measurement noise, and inaccurate camera poses.

W2. Sensitivity to Reference Frame Quality:

Inference relies heavily on the single input edited reference frame. If this reference frame, when generated by a 2D diffusion model, exhibits view inconsistency, artifacts, or lighting disharmony, to what extent would JointEdit3D's geometry prediction mechanism be affected?

W3. More Visualizations for Appearance Editing Are Expected:

Compared with add, move, and delete operations, the visualizations for appearance editing are relatively scarce. It would be desirable to see more visualizations of appearance editing results.

Questions for authors

Please see weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer FV7A: Real-Scene Generalization, Reference Quality, and Appearance Editing

We thank the reviewer for the positive assessment of our formulation, the SceneAnchor design, the dataset, and the experiments. All new results have been added to Appendix F of the revised PDF.

W1. Synthetic-to-real generalization.

JointEdit3D keeps all pretrained components frozen, including the geometry latent space and its decoder, which were pretrained on large-scale, mostly real-world data. Only the 70M-parameter SceneAnchor Branch is trained, so the synthetic data teaches the editing behavior while the real-world priors are largely preserved. JointEdit3D also takes neither depth nor camera poses as input, so depth noise and pose errors do not enter the pipeline.

On real scenes, JointEdit3D outperforms all baselines on 360-USID and Inpaint360GS (Tables 2 and 12) and on SPIn-NeRF on all metrics except SSIM (Table R1), without masks or camera poses. It also performs object replacement and appearance editing on real captures, and transfers global stylization zero-shot (Figures R1 and R2). In qualitative comparisons on many real scenes, JointEdit3D also produces cleaner edits than the baselines (Figures 4, 5, and 9 of the paper and Figures R4–R6).

W2. Sensitivity to reference frame quality.

We agree that reference quality matters. Relying on a single edited reference is shared by the reference-guided paradigm, including Omni-3DEdit, MVInpainter, EditCast3D, and AuraFusion360. To quantify this sensitivity, we perturbed the edited reference on SceneEdit3D-Bench with a view misalignment, where the reference is placed 5 frames away from the frame it was edited from, a global lighting shift, underexposure, a soft one-sided shadow, and slight defocus. These mimic the view inconsistency, lighting disharmony, and artifacts of 2D editors (Table R4).

Reference perturbationΔ Edit PSNRΔ Edit CDΔ Bg. PSNRΔ Bg. CD
View misalignment−1.39+0.0342−0.43+0.0073
Global lighting shift−7.27+0.0165−7.07+0.0020
Underexposure−1.19+0.0604−2.76−0.0029
Soft one-sided shadow−0.55+0.0722−1.72−0.0056
Slight defocus+1.07+0.0286−0.31−0.0036

Table R4. Change relative to the clean reference. Lower PSNR and higher CD are worse.

In the background, large 2D changes barely affect the geometric error. The global lighting shift lowers background PSNR by 7 dB because the output follows the reference appearance, yet background CD changes by only 0.002, and it changes by less than 0.008 under all perturbations, since the background geometry is anchored by the source video. In the edited region, the global lighting shift also changes CD only by 0.017 despite a 7 dB PSNR drop, but the geometry is more sensitive to local lighting changes on the edited object, which raise its CD by up to 0.07, since the edited content is specified only by the reference. View misalignment changes PSNR only slightly (−1.39 and −0.43) and mainly affects the edited region (CD +0.034), as the reference no longer matches the source viewpoint at its position. Supplying multiple edited references helps in practice (Appendix B.3), and we discuss this limitation in Appendix E.

W3. More appearance editing results.

In addition to the backpack recoloring in Figure 9 of the paper, we added appearance edits on a real capture, changing a sofa to red and changing its material to leather (Figure R2(a)). Both edits propagate consistently to all views, while the sofa shape and the surrounding scene are preserved. We also added zero-shot global stylization (Figure R2(b)) and object replacement, where the generated point cloud changes together with the replaced object (Figure R1).

We hope these results address the reviewer's concerns, and we are happy to discuss any remaining questions.

Thank you for your detailed response! I decide to keep my score.

AuthorsOpenReview ↗

Thank You for the Continued Support

We are glad that our response addressed your concerns. Thank you for your time and for your continued support of our work.

Reviewer_rarNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents JointEdit3D, a feed-forward 3D scene editing framework that jointly predicts edited appearance and geometry in a unified RGB-geometry latent space through asymmetric latent inpainting. Given a source video, a single edited reference frame, and an optional language instruction, JointEdit3D jointly generates the remaining RGB latents and the entire geometry latent. The method further introduces a SceneAnchor Branch and edit-aware training objectives to propagate edits across views while preserving unedited scene content. In addition, the authors introduce SceneEdit3D-15K, a paired 3D scene editing dataset, and evaluate the proposed method on SceneEdit3D-Bench, 360-USID, and additional real-world scenes.

Critical strengths
  1. The paper presents a unified 3D scene editing framework that jointly generates edited appearance and geometry in latent space. The experimental results demonstrate promising performance across various editing tasks.
  2. The qualitative results show convincing improvements in editing quality.
  3. The paper is well-written.
Critical weaknesses
  1. JointEdit3D is largely built upon the pretrained Gen3R framework. However, the paper does not include direct comparisons with Gen3R using either a single edited reference image or edited multi-view images as input. Such comparisons would help clarify the benefits of the proposed editing framework over the underlying reconstruction model.
  2. The quantitative evaluation of object removal is mainly conducted on SceneEdit3D-Bench and 360-USID. Although the paper includes qualitative comparisons on additional real-world scenes, further quantitative evaluation on widely used benchmarks such as SPIn-NeRF [1] dataset would help establish the method's generalization. In addition, comparisons with recent 3D inpainting approaches, such as AuraFusion360 [2] and InstaInpaint [3], would strengthen the evaluation.
  3. In some demo videos in the supplementary material, noticeable changes in color tone or illumination appear in regions that should remain unchanged. Since background preservation is an important objective of this work, further analysis of these appearance inconsistencies would be helpful.

[1] Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. CVPR 2023. [2] Aurafusion360: Augmented unseen region alignment for reference-based 360 unbounded scene inpainting. CVPR 2025. [3] Instainpaint: Instant 3d-scene inpainting with masked large reconstruction model. NeurIPS 2025.

Questions for authors

Please refer to the concerns raised in the Critical Weaknesses section.

2 comments
AuthorsOpenReview ↗

Response to Reviewer rarN (1/2): Comparison with Gen3R

We thank the reviewer for the positive assessment of our framework, editing quality, and writing. All new results have been added to Appendix F of the revised PDF.

W1. Direct comparison with Gen3R.

We compare with Gen3R using a single edited reference, and evaluate edited multi-view input through video-edit-then-reconstruct pipelines (Table R3).

(a) Single edited reference. Gen3R's single-image mode conditions on the first frame, so all methods use the edited first frame as the reference.

(b) Edited multi-view input. Since independent per-frame 2D edits are inconsistent across views, we propagate the reference with two first-frame-guided video editors, I2VEdit and AnyV2V, and reconstruct the edited videos with VGGT.

MethodInputEditBg.NearFarCD↓F-score↑
Gen3Rfirst-frame ref. + cameras21.13 / 0.34520.06 / 0.33022.53 / 0.24418.13 / 0.3930.14580.9386
I2VEdit → VGGTedited video22.99 / 0.34223.95 / 0.30726.83 / 0.27021.67 / 0.3390.12540.9447
AnyV2V → VGGTedited video20.25 / 0.42923.99 / 0.27924.79 / 0.24722.47 / 0.3160.15240.9225
JointEdit3Dsource video + first-frame ref.26.54 / 0.19332.73 / 0.07832.75 / 0.08330.93 / 0.0870.09760.9636

Table R3. Comparison with an edited first-frame reference on the 94 SceneEdit3D-Bench samples in which the edited object covers at least 0.5% of the first frame. RGB cells are PSNR↑ / LPIPS↓. Near and far denote frames close to and far from the reference.

Without the source video, Gen3R cannot preserve unedited content (Bg. PSNR 20.06 vs. 32.73). It hallucinates scene content as the camera moves away from the reference, and its imprecise camera control lets the views drift from the given trajectory (Figure R3). Its PSNR therefore drops from 22.53 on near frames to 18.13 on far frames, while JointEdit3D stays above 30 dB. The outputs of the video editors drift from the source in color and style (Figure R6), and VGGT carries these errors into the geometry (CD 0.1254 and 0.1524 vs. 0.0976). JointEdit3D performs best on all metrics.

Benefits over the underlying Gen3R. Our comparisons form a progression from Gen3R to JointEdit3D. Table R3 evaluates Gen3R as is, and the ablation study in Table 5 of the paper includes two straightforward ways of turning Gen3R into an editor with the same training data. Since the SceneAnchor outputs are zero-initialized, JointEdit3D also starts training exactly from Gen3R, so these comparisons indicate the benefit of our design. (i) Fine-tuning Gen3R with LoRA to condition on the source video tends to copy the source and edits poorly (Edit PSNR 23.53 vs. 26.82), which shows the benefit of injecting the source through the SceneAnchor Branch. (ii) Generating RGB only and reconstructing it with VGGT degrades geometry (CD 0.1506 vs. 0.1053), which shows the benefit of generating appearance and geometry jointly in the unified latent space. These results suggest that the improvements over Gen3R come from these two designs rather than from the base model alone.

AuthorsOpenReview ↗

Response to Reviewer rarN (2/2): Real-Scene Evaluation, InstaInpaint, AuraFusion360, and Color Tone

W2. Quantitative evaluation on real scenes and recent 3D inpainting methods.

We added SPIn-NeRF (Table R1), where all methods receive the same real object-free photograph as the reference.

MethodMask-freePose-freeM-PSNR↑BG-PSNR↑SSIM↑LPIPS↓M-LPIPS↓
MVInpainter–✓19.4824.730.7420.3180.387
Omni-3DEdit✓–16.0818.240.5800.4750.428
InstaInpaint––21.7323.330.7880.2940.285
JointEdit3D✓✓23.1226.400.7730.2910.266

Table R1. Object removal on the 10 SPIn-NeRF scenes. M- and BG- denote the object region and the background.

JointEdit3D achieves the best scores on all metrics except SSIM, without using masks or poses. SPIn-NeRF favors few-view methods such as InstaInpaint, since its forward-facing scenes keep the object near the image center and the test views close to the input views. InstaInpaint fills only the masked object region, so it keeps the book's reflection on the table, whereas JointEdit3D follows the reference without a mask and removes it as well (Figure R5).

To compare with InstaInpaint further, we also evaluated it quantitatively on SceneEdit3D-Bench and added qualitative comparisons on deletion, addition, and relocation, including real scenes such as ScanNet++ (Figure R4). Since InstaInpaint uses four views with masks and poses, whereas JointEdit3D uses the source video without either, we first varied InstaInpaint's input views on deletion (Table R6).

Method (input views)Edit PSNR↑Edit LPIPS↓
InstaInpaint (4 views, near the reference)24.650.295
InstaInpaint (4 views, uniform)25.870.240
InstaInpaint (8 views, uniform)23.470.326
JointEdit3D (49-frame source video)31.920.151

Table R6. InstaInpaint with different input views on deletion (SceneEdit3D-Bench).

Spreading the four views over the trajectory improves InstaInpaint by 1.2 dB, while eight views lower its PSNR to 23.47. Its best setting remains 6 dB below JointEdit3D (25.87 vs. 31.92), so the gap persists across the tested view settings. Unlike SPIn-NeRF, SceneEdit3D-Bench contains complex scenes in which the edited object is often not prominent in the view, where InstaInpaint does not perform well. InstaInpaint is designed mainly for object removal, and on the more complex addition, relocation, appearance, and multi-operation edits it fails on a large number of edits (Edit PSNR at most 16.91 vs. at least 23.19 for JointEdit3D, Table R7 in Appendix F).

AuraFusion360.

We added AuraFusion360 to the deletion comparison of Table 2 (Table R2). JointEdit3D outperforms it on both benchmarks, without masks, camera poses, or per-scene optimization.

MethodSceneEdit3D-Bench360-USID
Omni-3DEdit24.86 / 0.30717.78 / 0.3655
Omni-3DEdit + SceneEdit3D FT28.99 / 0.23018.02 / 0.3637
AuraFusion36021.06 / 0.38517.53 / 0.3764
JointEdit3D31.92 / 0.15118.57 / 0.3426

Table R2. Deletion edit-region PSNR↑ / LPIPS↓, following Table 2, where the other baselines are listed.

W3. Color tone changes in the demo videos.

JointEdit3D is trained with a region-decomposed objective that explicitly guides background reconstruction. It still regenerates the whole frame rather than copying unedited pixels, so small changes in fine details and tone cannot be fully avoided. Compared with other methods that regenerate the scene, however, JointEdit3D is markedly more stable. On SceneEdit3D-Bench, its background PSNR is 32.33, compared with 24.65 for Omni-3DEdit, and close to the mask-guided MVInpainter (32.76) without using masks (Table 3). Even dedicated video editing models preserve the background much worse. In Table R3, I2VEdit and AnyV2V reach a background PSNR of at most 23.99 (vs. 32.73), and the overall color and style of their outputs drift from the source as the video progresses (Figure R6). JointEdit3D stays above 30 dB on both near and far frames.

We hope these results address the reviewer's concerns, and we are happy to discuss any remaining questions.

Comments on the paper

1 comment

Dear reviewers,

Thanks for your continued efforts. The authors have provided a rebuttal to the initial reviews. Please review it carefully, along with all other reviews and the corresponding author responses, and consider whether you have any further questions for the authors or whether the rebuttal has addressed your concerns. There are only a few days remaining before the discussion period closes, so please use this time to seek any further clarifications or raise any additional questions about this paper.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.