acceptodds
Under review as a conference paper at ICLR 2027

OmniShowcase: Omnidirectional Subject-Consistent Video Generation From Multi-View Images

Abstract

Generating videos of a specific subject from a few reference images remains challenging when the camera reveals surfaces that are not visible in the input. Early subject-to-video (S2V) methods condition on a single reference image and must therefore infer the subject's unseen appearance and geometry as the viewpoint changes. To address this limitation, recent methods extend S2V to multiple reference views, providing more information about the subject from different viewpoints. However, they do not explicitly specify the 3D structure that connects these observations. In addition, existing training data do not guarantee that viewpoint changes preserve the same underlying 3D subject. We present \method, built on two components: (i) a fully automated data synthesis pipeline that produces training videos along diverse camera trajectories, with viewpoint variation driven by underlying subject geometry, even spanning full 360-degree orbits; and (ii) a geometry-conditioning method that provides the video model with a turntable of surface-normal maps reconstructed from the reference views without requiring temporal alignment with the generated frames. We construct Showcase-4K}, a corpus of 3D-asset-based videos covering twelve camera motions across trajectory configurations with reconstruction-derived surface-normal conditioning, and ShowcaseBench, a held-out evaluation set with the source meshes and camera parameters used to render each video. Experiments show state-of-the-art subject identity and 3D consistency while maintaining competitive video quality and motion.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Generation, Subject-to-Video, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kFNmOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposed OmniShowcase, a system that generates an orbiting video of an object from the object's multi-view renderings and a text description. To train such a model, the author proposed to synthesize data by inpaint the object and use pretrained VACE to generate the orbiting video. Then the model is trained using both reference images and rendered geometry (depth and surface normal) information. Results show improvement upon existing methods.

Critical strengths

None

Critical weaknesses
  1. The need for generative video synthesis is insufficiently justified. The paper only considers static subjects under orbiting camera motion and already reconstructs a 3D representation from the input views. A natural alternative is to render that reconstruction within a 3D scene, which offers explicit camera control and geometric consistency. Obtaining such 3D scene is very simple, e.g., either by download or purchase from Internet sources or use any existing scene generation systems to create them. The paper does not prove that the proposed generative approach is a good trade-off. Thus, the central design choice insufficiently motivated. Moreover, the experiments do not resolve this design question. Comparisons with other subject-to-video methods demonstrate improvements within the chosen approach, but omit reconstruction-and-rendering alternatives using the same four reference images.

  2. Following the first point, the proposed data pipeline introduces uncertainty into the scene consistency it aims to promote. Rendering the subject from a fixed mesh provides a sound basis for consistency. However, generating the surrounding scene with video inpainting does not enforce a consistent background, especially when the base model is poor, and the authors acknowledge related limitations in Appendix. As a result, the training data might be 3D-inconsistent and harm the overall performance.

  3. The methodological novelty appears incremental. The method adapts a pretrained video DiT through LoRA and conditions it on reference-image and reconstructed-normal latents concatenated along the temporal axis. This is a very standard design. Given that the dataset part is also not well-motivated, I do not find this paper delivering meaning contribution to the community.

  4. Minor issues

  • Section 4.1 characterizes the four-view baseline extensions as stronger on almost every metric, whereas Tables 6–7 show worse GT-CD for both Phantom and MAGREF on both datasets; on NAVI, MAGREF also worsens on MEt3R and RotCoverage.
  • Table 8’s caption does not correctly distinguish the four orbit variants from the other motions, while Table 9 states that arc motions take no suffix despite the framing suffixes specified in Section G.2 and Table 10.
  • The discussion of Table 3 introduces the undefined labels “RGB DC” and “Sparse DC,” rather than using the table’s terminology.
Questions for authors

Please address my concerns mentioned in the weakness section.

7 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer kFNm (1/3) - W1a

Thank you for checking the appendix tables against the main text closely enough to find a factual error in Section 4.1 and two caption errors, and for pointing out the comparison we had left out, reconstruct-and-render from the same four references. The revision now includes that comparison and all three corrections. We answer the first part of W1 here, and the remaining points in parts (2/3) and (3/3).

Our description of the setup was compact, so we restate three details in brief to avoid any misunderstanding. (1) The NAVI references are real photographs of real-world subjects. (2) Each training clip renders the subject from its 3D asset and generates only the background around it (Section 3.2). (3) The geometry condition is surface normals only, which are converted from rendered depth maps (Section 3.2).

Reviewer kFNm - W1a: Comparison with reconstruct-and-render

"omit reconstruction-and-rendering alternatives" (W1)

We implement this as a new baseline to complete the comparison and evaluation. The revision adds it to the main comparison (Table 2), with implementation details in Appendix G.1.

Reconstruct-and-Render (R&R). The subject is reconstructed from the same four reference images with MV-SAM3D, the same reconstruction our pipeline runs, placed into a 3D scene that Marble generates from a scene description taken from the prompt the generative methods receive, and rendered along the commanded trajectory. R&R is provided with the ground-truth camera path of each sample and with colors matched to the ground truth, while OmniShowcase receives neither and is given only the trajectory sentence in its prompt. Its identity and 3D metrics are computed on the subject rendered alone, so they judge the reconstruction rather than the scene around it, while the video quality and prompt alignment metrics use the whole video.

Table 2 (partially excerpted from the updated PDF). Reconstruct-and-render against OmniShowcase.

SetMethodDINO↑CLIP↑MEt3R↓GT-CD↓RotCov↓ViCLIP↑Imaging↑Aesthetic↑
Held-outR&R0.7140.8980.3120.10150.6740.19054.900.498
Ours0.7990.9250.1900.08240.6650.25974.290.574
NAVIR&R0.7620.9160.3170.09881.0130.20957.840.522
Ours0.8280.9180.2490.09751.0740.26074.580.551

Analysis. Because it follows the ground-truth path, R&R leads every generative baseline on GT-CD and RotCov. OmniShowcase, given only the trajectory sentence, is still ahead of it on DINO and MEt3R on both benchmarks and on held-out GT-CD (0.0824 versus 0.1015), and matches it on NAVI GT-CD (0.0975 versus 0.0988) without the ground-truth path. R&R also records the lowest ViCLIP, imaging and aesthetic scores in Table 2. R&R keeps a small edge in RotCov on NAVI (1.013 versus 1.074), which follows from executing the path directly. The identity gap is visible in Figure 5, where R&R is shown at the upper bound of our implementation, with the ground-truth camera path and colors matched to the ground truth. The fine print on a label is still blurred in four-view reconstruction by MV-SAM3D, and careful composition of the scene cannot fix it. OmniShowcase uses the same reconstruction only for its geometry, as a normal turntable, and takes the appearance from the reference images, so this limit does not apply to it.

This is the trade-off the Reviewer kFNm asked us to demonstrate. R&R gives explicit camera control by construction, but it can only render the subject as reconstructed inside a scene that was fixed when it was generated, and the two have to be fitted together sample by sample. This fitting has to meet three requirements that work against each other, following the commanded trajectory, keeping the scene elements described in the prompt readable, and keeping the camera out of the scene geometry, as detailed in part (2/3) and Appendix G.1. OmniShowcase instead leaves scene creation to the generative prior and draws the scene that the prompt describes around the subject in a single video. It adds only the part the prior lacks, keeping the subject's identity and 3D structure consistent along the commanded trajectory, without per-sample human intervention. Measured on the same subjects with the same protocol, OmniShowcase yields better identity and MEt3R, lower GT-CD on the held-out set and comparable GT-CD on NAVI, and higher imaging and aesthetic scores with high visual quality.

We continue our rebuttal in the next comment.

AuthorsOpenReview ↗

First Rebuttal for Reviewer kFNm (2/3) - W1b, minor issues

Reviewer kFNm - W1b: Obtaining a 3D scene

"Obtaining such 3D scene is very simple" (W1)

We agree that obtaining 3D scenes itself is relatively easy. However, scenes that match each prompt exactly are very hard to obtain off the shelf, because our prompts describe eighty-five distinct environments (50 for ShowcaseBench and 35 for NAVI) that no asset library covers. We therefore took the other route you suggested, an existing scene generation system, and generated each scene with Marble from World Labs, a commercial text-to-world system, on a pay-as-you-go basis at $1.26 per scene. Marble is prompted with a scene description taken from the same prompt our model receives, by dropping the camera-trajectory clause and the subject description and keeping the remaining scene as a noun phrase. Since R&R's identity and 3D scores are measured on the subject rendered alone, the scene does not lower them. The hard part turned out to be composing the subject into the scene, not obtaining it.

To implement R&R faithfully at benchmark scale, we fit a composition instruction rather than placing each subject by hand (Appendix G.1, Figure 12). For a small set of representative failure modes we prepared handcrafted targets, and an LLM agent (Claude Opus 5.5) revised a shared instruction through a render-and-review loop until its compositions matched them. The fitted instruction was then applied to all 85 samples. Per sample, reconstructing the subject takes 2.6 minutes, generating the scene about five, the automated render-and-review loop about one, and our own inspection and corrections about five, so a sample costs on the order of fifteen minutes end to end. This is somewhat longer than the 11.8 minutes OmniShowcase needs per video (Appendix H.4), and part of that time is human effort that does not go away with more compute. Building it this way showed two problems.

  • The scene is fixed once it is generated. Running the commanded orbit at a scale where the background is legible drives the camera through a wall or a piece of furniture. Tightening the orbit clears the geometry but exposes the scene's finite resolution and leaves almost no background in frame, and enlarging the scene pushes the elements described in the prompt out of readable range. The composition finally used is a compromise rather than a solution.
  • Placement is per sample. The fitted instruction carries the strategy but not the surface, scale and heading each scene calls for, and for a few samples the camera path itself had to be adjusted slightly.

R&R keeps explicit camera control and a rigid subject. OmniShowcase produces the commanded orbit from a text description alone, without per-sample scene acquisition, placement or relighting, and on the same subjects it gives better identity and visual quality (Table 2).

Reviewer kFNm - Minor issues

We are glad that you point them out. All three are correct and now they are fixed. In the revision, Tables 6 to 10 of the submission are now Tables 7, 8, 12, 13 and 14, and Section G.2 is now Section I.2.

  1. Section 4.1 now states that the four-view extension improves identity for both baselines but degrades GT-CD in every case, and on NAVI also MAGREF-MV's MEt3R and RotCoverage, that we report the four-view numbers because they use the same four reference images as our method, and that OmniShowcase outperforms both settings on every metric of Tables 7 and 8.
  2. The Table 12 caption now distinguishes the four orbit variants from the other motions, and the Table 13 statement now matches the framing suffixes in Section I.2 and Table 14.
  3. The Table 3 discussion now uses the table terminology, Textured and Sparse, in place of RGB DC and Sparse DC.

We continue our rebuttal in the next comment.

AuthorsOpenReview ↗

First Rebuttal for Reviewer kFNm (3/3) - W2, W3

Reviewer kFNm - W2: Background consistency of the training data

"the training data might be 3D-inconsistent" (W2)

We are grateful that you raised it, which has led to clarification of the scope of OmniShowcase. Our work targets subject-level 3D consistency, meaning that the subject's shape and appearance should stay coherent as the camera moves around it. Background consistency was not among the goals we set for the method, and the evaluation reflects this, since every identity and 3D metric is computed on the masked subject (Section 4.1). We recognize, however, that the submission does not state this scope where it matters most. Section 3.2 was titled Automated 3D-Consistent Data Synthesis without a qualifier, and the compositing paragraph described the inpainted background as coherent, so it is natural to read the claim as covering the whole scene.

We have retitled the subsection to Automated Data Synthesis with 3D-Consistent Subjects to make this scope explicit, and the compositing paragraph now describes a realistic background that follows the scene prompt rather than a coherent one.

During curation, gate G2 rejects composited clips whose background does not behave like one rigid 3D place (Appendices I.1 and K). We have also measured the background directly. Appendix H.2 (Table 10) compares Showcase-4K clips with clips that Uni3C, a camera-controllable image-to-video model, generates for the same subjects along the same camera paths. With a similar amount of background motion, Showcase-4K is more consistent on both background MEt3R (0.138 versus 0.186) and VBench background consistency (0.903 versus 0.889). So the inpainting step leaves the training background more consistent than generating the whole clip would. In the generated videos, our background MEt3R matches that of the ground-truth render (0.182 for both) at a similar camera rotation, and the model trained on these clips leads every method on DINO, CLIP, MEt3R and GT-CD (Table 2).

Reviewer kFNm - W3: Methodological novelty

"This is a very standard design." (W3)

We agree that LoRA adaptation and temporal-axis concatenation are established components, and we do not claim them as contributions. The design is deliberately simple, and each of its choices comes from the analyses in the paper. Our key observation is that, when the camera reveals sides that no reference shows, the difficulty is geometric rather than purely photometric (Section 1). Therefore, the model needs training data with exact subject geometry and explicit dense geometry at inference.

Our data synthesis pipeline provides the first. It renders the subject from its 3D asset and inpaints only the background, so the subject and its 3D consistency come from the asset itself. Vision-language gates with fixed rubrics replace human review, so the corpus scales with compute rather than annotator time (Appendix I.1). Against clips of the same subjects and camera paths generated by Uni3C, a camera-controllable image-to-video model, Showcase-4K clips score better on DINO (0.8030 versus 0.7167) and MEt3R (0.1871 versus 0.3052) (Appendix E).

The normal turntable provides the second, and the ablations show that it is essential to our model. Without it, held-out GT-CD rises from 0.0824 to 0.2092 and NAVI GT-CD from 0.0975 to 0.1987. A textured turntable of the same geometry gives 0.1698 and a sparser one 0.1777, and even a turntable from a single view helps (Appendix H.1). Since it needs no frame-by-frame alignment to the target, a single render of the reconstruction serves multiple camera motions.

With these two parts, our model requires neither large-scale training data nor per-subject optimization (Table 1). On ShowcaseBench held-out DINO, the best generative baseline scores 0.707 and R&R 0.714, our model without geometry conditioning 0.752, and the full model 0.799, and NAVI shows the same pattern (Tables 2 and 3). The R&R alternative, which we newly added to address W1, also falls behind this simple design on DINO and MEt3R on both benchmarks (Table 2). Showcase-4K and ShowcaseBench will be released with their source meshes and ground-truth cameras, and we hope they will support future research on this task.


Thank you again for the close reading, which has made the paper more accurate and its claims more carefully scoped. We hope our rebuttal addresses the reviewer's concerns and would appreciate it if the reviewer could reconsider the assessment, taking our rebuttal into consideration. We are happy to answer any further questions.

Best Regards,

The Authors

Reviewer_n2TgOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents OmniShowcase, a framework for omnidirectional subject-consistent video generation from multi-view reference images. The authors propose an automated data synthesis pipeline that renders subjects from 3D assets and uses VACE to generate realistic backgrounds, producing a geometrically consistent training dataset, Showcase-4K. In addition, the method incorporates reconstructed surface-normal sequences as geometry-aware conditioning to improve subject consistency under large viewpoint changes.

The experimental results demonstrate promising improvements in subject identity preservation and 3D consistency compared with existing subject-to-video generation methods.

Critical strengths

1. Meaningful problem and interesting data synthesis pipeline.

The paper addresses an important problem in subject-to-video generation, particularly maintaining subject identity and geometric consistency under large camera movements. The proposed data synthesis pipeline is interesting and practical. Combining 3D asset rendering with video inpainting provides a reasonable way to construct geometrically consistent foreground subjects in diverse scenes.

2. Strong experimental results.

The experiments are generally comprehensive, including comparisons with existing methods, geometry-conditioning ablations, and evaluations on both synthetic and real-world subjects. The reported improvements in identity preservation and 3D consistency are encouraging and demonstrate the effectiveness of the proposed framework.

Critical weaknesses

1. Background consistency under camera motion.

My first concern is whether the backgrounds generated by VACE can maintain reliable 3D and temporal consistency under large camera movements, especially full 360-degree rotations.

Although the authors employ VLM-based filtering to check scene consistency during data construction, it remains unclear whether such filtering is sufficient to ensure consistent background geometry in the final generated videos.

In particular, the main identity and geometric consistency metrics, such as DINO, MEt3R, and GT-CD, are evaluated on segmented foreground subjects. Consequently, the paper lacks a direct quantitative evaluation of background consistency.

I would appreciate further clarification or additional experiments evaluating the geometric and temporal consistency of the generated backgrounds under camera motion.

2. Effectiveness and scalability of the 4K training dataset.

The proposed Showcase-4K dataset contains only 4,000 training videos, which is relatively small for video generation tasks. Nevertheless, the method achieves substantial improvements over existing approaches.

While this is an encouraging result, I would like the authors to provide more discussion or analysis explaining why such a relatively small dataset is sufficient to achieve these gains.

For example, how much of the improvement can be attributed to the geometric consistency and quality of the synthesized training data? How does performance change with different amounts of training data?

A data-scaling analysis, or further discussion of the effectiveness and diversity of Showcase-4K, would help clarify the contribution and scalability of the proposed data synthesis approach.

3. End-to-end inference efficiency.

The proposed method requires additional geometric processing for each new subject, including multi-view 3D reconstruction and rendering a surface-normal turntable before video generation.

Although these steps provide useful geometric information, they also introduce additional inference overhead compared with methods that directly condition on reference images.

Could the authors report the end-to-end inference time and computational cost, including a breakdown of 3D reconstruction, normal rendering, and video generation? A runtime comparison with existing baselines would help assess the practical efficiency of the proposed approach.

Questions for authors

Overall, I find this paper interesting and practically valuable. The proposed data synthesis pipeline addresses a meaningful problem, and the experimental results are generally convincing, particularly for foreground subject identity preservation and 3D consistency.

My main concerns are the lack of direct evaluation of background consistency, the limited analysis of why 4K training videos are sufficient, and the additional computational overhead during inference.

I am generally positive about this work and would appreciate further clarification on these points during the rebuttal.

5 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer n2Tg (1/2) - W1

We thank the Reviewer n2Tg for the positive assessment and the constructive questions. We are glad the reviewer found our data synthesis with 3D-consistent subjects pipeline interesting and practical and the experiments generally comprehensive. All three requests were ones that we could work on, and we have measured each of them. We answer W1 here, and W2 and W3 in part (2/2).

Reviewer n2Tg - W1: Background consistency under camera motion

"the paper lacks a direct quantitative evaluation of background consistency" (W1)

We thank you for pointing this out. Our identity and 3D metrics are computed on segmented foreground subjects, so the submission did not measure the background directly. VLM filtering during data construction is a filter on training data and is not a measurement of the generated output.

We have added a direct measurement. Appendix H.2 (Table 10) reports background MEt3R for 3D consistency and VBench background consistency for temporal consistency, computed after removing the subject, for our method and the baselines on the same videos as Table 2. Both measures worsen when the frames of a ground-truth video are shaken or shuffled, so they do detect an inconsistent background. However, they also improve when the camera barely moves, so the table also reports the actual camera rotation of each video as a ratio to the commanded rotation.

Table 10(b) (partially excerpted from the updated PDF). Background consistency of the generated videos on ShowcaseBench.

MethodBg MEt3R↓Bg VBench↑Rotation ratio↑
Ground-truth render0.1820.8950.62
Phantom-MV0.0390.9560.05
MAGREF-MV0.0520.9400.06
MV-S2V0.0910.9240.21
3DreamBooth0.1090.9170.24
Ours0.1820.9000.58

For this reason, the baselines obtain the best background scores on ShowcaseBench. They turn the camera far less than the ground-truth render does, and a background that hardly moves is easy to match. The fair reference is the ground-truth render, measured in the same way. OmniShowcase turns the camera more than twice as far as any baseline and about as far as the ground-truth render, and matches its background MEt3R, so its background is as consistent as the data it learns from. This is the measurement you asked for. On ShowcaseBench, whose four trajectories are all 360-degree orbits, our background scores stay close to those of the ground-truth render while the camera turns nearly as far. On the training side, the same appendix shows that our training clips are more consistent than Uni3C clips of the same subjects and camera paths (details in our response to Reviewer kFNm, W2).

We have also made the scope explicit. Section 3.2 is now titled Automated Data Synthesis with 3D-Consistent Subjects, and Appendix H.2 states that the 3D consistency we claim concerns the subject (see also our response to Reviewer kFNm, W2).

We continue our rebuttal in the next comment.

AuthorsOpenReview ↗

First Rebuttal for Reviewer n2Tg (2/2) - W2, W3

Reviewer n2Tg - W2: Why 4,000 videos suffice

"why such a relatively small dataset is sufficient" (W2)

The submission contains the decomposition you are asking for but does not present it as one. The reference-only variant is trained on Showcase-4K with no geometry conditioning. Comparing it to the best baseline therefore indicates what the training data contributes, and comparing our full model to it isolates the contribution of the conditioning.

Tables 2 and 3 (compiled from the updated PDF). Contribution of the data and of the conditioning, measured by DINO - the higher the better. The best generative baseline is 3DreamBooth on ShowcaseBench and MV-S2V on NAVI.

held-out DINONAVI DINO
Best generative baseline0.7070.770
Reconstruct-and-render0.7140.762
Ours without normal conditioning (None in Table 3)0.7520.793
Ours0.7990.828

On ShowcaseBench DINO, the margin over the strongest baseline splits about evenly between the training data and the geometry conditioning, and on NAVI about 40 percent comes with the training data. Appendix E compares the clips themselves with Uni3C clips.

We believe a relatively small dataset is sufficient because each clip provides a very clean training signal. Every clip is rendered from a single mesh with known cameras, and the subject pixels stay unchanged when the background is added. Any change in how the subject looks within a clip therefore comes only from the change in viewpoint. The pretrained backbone already knows the appearance of objects and scenes. It still needs to learn the appearance of a specific subject from viewpoints the references do not show, and Showcase-4K provides exactly this. Showcase-4K is also diverse, covering 2,593 assets in 620 categories from three asset sources and all twelve camera motions with eighty-eight configurations we define (Appendix I).

To support this, section 4.3 now includes a data-scaling analysis at 500, 2,000 and 4,000 videos, each trained for the same number of epochs (Table 4, Appendix H.3). Every metric improves as the data grows on both benchmarks, with aesthetic quality on NAVI nearly flat. The largest gains are in 3D consistency, where held-out GT-CD falls from 0.1260 to 0.1076 to 0.0824 and NAVI GT-CD from 0.1609 to 0.1359 to 0.0975. Even the 500-video model surpasses every baseline in Table 2 on DINO and MEt3R on both benchmarks. So 4,000 videos are enough to place the model ahead of existing methods but are not an upper limit, and more data only requires running the same pipeline on more assets.

Reviewer n2Tg - W3: End-to-end inference cost

"report the end-to-end inference time and computational cost" (W3)

We thank the reviewer for asking about efficiency, which the submission did not report. Appendix H.4 (Table 11) breaks down inference cost into 3D reconstruction, normal rendering, and video generation. All runs use one B200 at 832x480, and times are in minutes, averaged over five ShowcaseBench subjects.

Table 11 (from the updated PDF, with each stage's share added). Inference cost per subject by stage (minutes).

StageOursPhantom-MVMAGREF-MVMV-S2V3DreamBooth
3D reconstruction2.6 (22%)----
Normal rendering0.2 (2%)----
Per-subject fine-tuning----19.2 (73%)
Video generation9.0 (76%)10.8 (100%)4.6 (100%)10.8 (100%)7.0 (27%)
Total per video11.810.84.610.826.2
Another video of the same subject9.010.84.610.87.0
Peak GPU memory (GiB)57.3120.388.294.7110.4
DiT evaluations (sampling steps x evaluations per step)150 (50x3)150 (50x3)80 (40x2)150 (50x3)50 (50x1)

Per-video cost. Reconstruction and normal rendering take 2.8 min, about a quarter of our total. Generation alone takes 9.0 min, shorter than Phantom-MV and MV-S2V (10.8 min). All three use 150 DiT evaluations, although ours generates 105 frames in one pass. With the 3D stages added, our total of 11.8 min is comparable to these two baselines. MAGREF-MV is faster because it takes 80 DiT evaluations rather than 150. Our peak GPU memory is also the lowest (57.3 GiB vs 88.2 to 120.3 GiB).

Several videos of one subject. The normal turntable is not frame-aligned to the target, so one reconstruction serves every trajectory of a subject and its cost is paid once. Each further video needs only generation (9.0 min). 3DreamBooth also amortizes by reusing its fine-tuned weights but requires 19.2 min of fine-tuning first.

Thank you again for the constructive questions. For each of the three points you raised, we have added a measurement to the revision (Tables 4, 10 and 11). If these resolve your concerns, we would be grateful if this were reflected in your updated evaluation, and we would be happy to clarify anything further.

Best regards,

The Authors

AuthorsOpenReview ↗

Gentle Reminder by Authors to Reviewer n2Tg

Dear Reviewer n2Tg,

Thank you again for finding our paper "interesting and practically valuable". Table 10 now measures background consistency directly, and our background MEt3R matches the ground truth render (0.182 for both) at a similar camera rotation. Table 4 shows steady gains from 500 to 4,000 videos, and even 500 videos surpass every baseline on DINO and MEt3R. Table 11 reports 11.8 minutes per video with the lowest peak GPU memory, providing per-stage breakdown.

We understand that this must be a busy period for you. Before the discussion ends on November 18 (AoE), we would be happy if you could let us know whether our rebuttal addresses your concerns.

Best Regards,

The Authors

Reviewer_CUJdOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper aims to generate subject-consistent videos from a few reference images. It uses the reference images to preserve appearance and a reconstructed turntable of surface normals to guide geometry. The authors also build SHOWCsASE-4K using an automated pipeline that renders subjects from 3D assets and generates backgrounds. Authors evaluate the method on held-out synthetic data and real subjects from NAVI, achieving strong results in quality comparision.

Critical strengths
  1. The normal turntable is a simple but effective way to provide geometric guidance while keeping appearance and camera motion flexible. Since the normal sequence does not need to align frame by frame with the target video, the same geometry can be used to generate different camera motions.
  2. The data pipeline is well designed to preserve accurate subject geometry through 3D asset rendering while introducing diverse backgrounds.
  3. The experiments provide convincing evidence that geometry conditioning helps. On NAVI, removing the normal condition increases GT-CD from 0.0975 to 0.1987. The evaluation on real subjects and the user study also make the results more convincing beyond the synthetic benchmark.
Critical weaknesses

The method’s robustness to reconstruction errors is not separately evaluated. Training filters out unreliable reconstructions, while the reference-count ablation changes both image evidence and geometry quality. It remains unclear whether inaccurate normal conditions can mislead generation despite informative RGB references, or become detrimental relative to reference-only conditioning.

Questions for authors
  1. With the reference images fixed, how does realistic reconstruction error affect generation? Can inaccurate normal conditions degrade results relative to the reference-only variant?
  2. Why use surface normals rather than depth rendered from the same reconstruction? Does normal conditioning provide a measurable advantage over depth?
3 comments
AuthorsOpenReview ↗

First Rebuttal for Reviewer CUJd (1/2) - Q1

Thank you for the dedicated review and for describing the normal turntable as "a simple but effective way to provide geometric guidance." You were right that the reference-count ablation confounds two variables, and that the training filter removes exactly the cases where the question matters. The revision now includes an experiment that separates them (Appendix H.1, Table 9), and even the turntable built from a single view improves on the reference-only variant. We answer W1 and Q1 here, and Q2 in part (2/2).

Reviewer CUJd - W1/Q1: Robustness to reconstruction error

"Can inaccurate normal conditions degrade results relative to the reference-only variant?" (Q1)

As you pointed out, reducing the number of references in the reference-count ablation changes both the image evidence given to the model and the quality of the reconstruction behind the normal turntable, so the two effects cannot be separated from that table.

Design. We hold all four reference images fixed in every condition and vary only the normal turntable, replacing it with one reconstructed from one, two, three, or four views, which acts as a gradually corrupted normal condition at inference. The model input on the appearance side is therefore identical across conditions and only the geometry quality changes. We include the reference-only variant, which receives no normal turntable at all, as the floor you asked about. In the table below, the reference-only variant is None, and we also include Sparse from Table 3, which receives an accurate turntable with fewer views.

Results are in Appendix H.1 (Table 9), reproduced below. All rows use the four reference images, and the None, Sparse and four-view rows repeat Table 3. Held-out denotes ShowcaseBench held-out.

Table 9 (reproduced from the updated PDF). Fixed reference images, normal turntable reconstructed from fewer views.

SetGeometry conditionDINO↑CLIP↑MEt3R↓GT-CD↓RotCov↓ViCLIP↑Imaging↑Aesthetic↑
Held-outNone0.7520.9030.2650.20920.7270.24869.850.547
Sparse0.7520.9120.2630.17770.6980.25071.840.552
Normals from 1 view0.7650.9140.2390.14970.7080.25171.550.557
Normals from 2 views0.7760.9170.2270.13040.6960.25472.230.561
Normals from 3 views0.7880.9200.2090.11280.6810.25673.170.567
Normals from 4 views (ours)0.7990.9250.1900.08240.6650.25974.290.574
NAVINone0.7930.9040.3150.19871.1070.23867.670.532
Sparse0.7720.9090.2880.17691.0900.24067.720.531
Normals from 1 view0.7980.9090.2880.15651.0990.24571.400.538
Normals from 2 views0.8110.9120.2780.14361.0900.24972.360.542
Normals from 3 views0.8180.9150.2660.12651.0810.25373.180.546
Normals from 4 views (ours)0.8280.9180.2490.09751.0740.26074.580.551

Notably, even the least accurate turntable we tested, reconstructed from a single view, matches or improves on Sparse in DINO, CLIP, MEt3R and GT-CD on both benchmarks. Sparse keeps the accurate reconstruction but shows it from fewer angles, which suggests that the coverage of the turntable matters as much as the accuracy of each view.

About whether inaccurate normals can be worse than no normals. In short, they are not worse. The one-view turntable, the least accurate reconstruction we tested, improves on None in every metric on both benchmarks, lowering held-out GT-CD from 0.2092 to 0.1497 and NAVI GT-CD from 0.1987 to 0.1565. Within the range we tested, down to a single view, we did not find a reconstruction quality below which the conditioning becomes harmful, and performance degrades gradually rather than collapsing as fewer views are used. The training filter you mentioned (gate G3, Appendix I.1) is used only to curate training data and is not applied at evaluation or inference (Appendix K), so the benchmark subjects above are not filtered by it.

About robustness of dense normal turntable. The model never sees exact geometry in the condition, even in training. Each training turntable is rendered from the MV-SAM3D reconstruction of the four references, as at inference (Section 3.2), while the target is rendered from the source mesh. Such a reconstruction is consistent with itself across views but does not match the true shape exactly. Rendering it directly, as R&R does, gives GT-CD 0.1015 on held-out and 0.0988 on NAVI, against 0.0824 and 0.0975 for our output (Table 2). We believe the model thus learns to treat the turntable as an approximate geometry cue to harmonize with the references as exact appearance, so a worse turntable does not override them.

We continue our rebuttal in the next comment.

AuthorsOpenReview ↗

First Rebuttal for Reviewer CUJd (2/2) - Q2

Reviewer CUJd - Q2: Surface normals versus depth

"Why use surface normals rather than depth" (Q2)

Thank you for this thoughtful question. It is a natural one to ask of any geometry condition, and we are glad to answer it here.

The geometry is identical and only the signal differs. A depth turntable and a normal turntable are rendered from the same reconstruction. Our pipeline renders depth and converts each depth map to surface normals (Section 3.2), so the choice changes only the signal handed to the generator. Appendix H.5 now gives the full argument, which rests on three properties.

1. Independence from the camera. This property enables the strength you identified. You noted that the normal sequence does not need to align frame by frame with the target video, so the same geometry can drive different camera motions. Depth changes with the distance between the camera and the subject, which ties a depth turntable to the orbit radius it was rendered at, while the target trajectory given in the prompt has its own radius, elevation and framing. Surface orientation does not change with camera distance, so the same turntable serves target trajectories at different distances.

2. Scale. A four-view reconstruction fixes the subject only up to a similarity, and the recovered scale differs from subject to subject, so a depth map arrives with a range the generator would have to discount before using it. Surface normals are unit vectors and do not carry scale, so one condition serves subjects whose reconstructions were recovered at different scales.

3. Form of the information. Much of the variation in a depth map is a low-frequency ramp that only places the subject in front of the camera, while the orientation that separates one surface from another is its gradient. Rendering normals takes that derivative once, under known camera geometry, rather than leaving it to the generator. The same step removes the need for any per-subject normalization of the condition.

Unfortunately, we could not train a depth-conditioned variant. Training one variant takes about 5.5 days on four B200 GPUs (Appendix C), and during the rebuttal our GPUs were committed to the other requested experiments, including retraining for the data-scaling study. We hope this prioritization is understandable. The case for normals over depth therefore rests on the three properties above and on the Textured ablation. The paper does not claim that normals outperform depth. Its contribution is a geometry condition that does not need to align with the target camera.

The closest controlled evidence in the paper is consistent with this. The Textured ablation (Table 3) replaces the normals with a colored render of the identical reconstruction, keeping geometry and viewpoints fixed and changing only the representation. It is worse than normals on every metric on both benchmarks. This suggests that the gain comes from the representation of the geometry and not merely from geometry being present.

The controlled experiment in Appendix H.1 addresses the weakness you raised, and we would be grateful if this could be reflected in your evaluation. We appreciate questions that went straight to the design choices behind the method. We would love to answer if you have further questions.

Best regards,

The Authors

AuthorsOpenReview ↗

Gentle Reminder by Authors to Reviewer CUJd

Dear Reviewer CUJd,

Thank you again for describing the normal turntable as "a simple but effective way to provide geometric guidance". On whether OmniShowcase is robust to inaccurate normals, Table 9 keeps the references fixed and rebuilds the turntable from one to four views. Even the turntable from one view outperforms the variant without normals on every metric on both benchmarks, lowering GT-CD on ShowcaseBench from 0.2092 to 0.1497. Appendix H.5 explains our choice of normals over depth.

We understand that this must be a busy period for you. We would be grateful to hear whether our revisions address your concerns, before the discussion ends on November 18 (AoE).

Best Regards,

The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response and Summary of Revisions

Dear Reviewers, ACs and PCs,

Thank you so much for the careful reading and the concrete suggestions. We have taken each review carefully into account, and the comparisons and analyses you asked for have made OmniShowcase a more complete paper. We have uploaded a revised PDF.

Revisions are highlighted in green and new sections and additions in orange. Minor wording and spelling edits are not marked. Section, table and figure numbers in our replies refer to the revised PDF, and each reply will point directly to the updated numbers.

Main Changes

  • Reconstruct-and-Render (R&R) is now included as a non-generative alternative. (Sec. 4.1, Tab. 2 and Fig. 5)
  • Sec. 4.1 is now corrected regarding the four-view extension of the single-view baselines. (Sec. 4.1)
  • Sec. 4.3 now separates the contribution of the training data from other contributions, and adds a data-scaling analysis. (Sec. 4.3 and Tab. 4)
  • The new Appendix G discusses alternatives to OmniShowcase, with
    • R&R composition details consisting of implementation, cost, fitting and generalization, and discussion (Appendix G.1 and Fig. 12),
    • A qualitative comparison with the commercial reference-to-video models Seedance 2.5 and Kling 4.0 (Appendix G.2 and Fig. 13).
  • The new Appendix H adds
    • robustness to reconstruction error with the reference images fixed (Tab. 9),
    • background consistency of the training clips and the generated videos (Tab. 10),
    • training data scaling from 500 to 4,000 videos (Appendix H.3),
    • inference cost and memory of all methods with per-stage breakdown (Tab. 11),
    • and the full composition instruction used for R&R (Appendix H.6).
  • Sec. 3.2 is now titled Automated Data Synthesis with 3D-Consistent Subjects. (Sec. 3.2)
  • Sec. 3.2 and the new Appendix H.5 explain the choice of surface normals over depth. (Sec. 3.2 and Appendix H.5)
  • The discussion of Table 3 now uses the table terminology, and the captions of Tables 12 and 13 are corrected. (Sec. 4.3 and Appendix I.2)
  • Appendix C is now corrected regarding the training batch size and training time. (Appendix C)

The revised PDF already reflects all of these changes. Our reply to each review is posted in its own thread.

Best Regards,

The Authors

AuthorsOpenReview ↗

Final Remarks by Authors

Dear Area Chair and Reviewers,

Thank you for overseeing our submission and for the constructive feedback. We would like to briefly summarize the discussion.

Reviewer n2Tg found that the paper "addresses an important problem in subject-to-video generation", and Reviewer CUJd described the normal turntable as "a simple but effective way to provide geometric guidance" with experiments that "provide convincing evidence that geometry conditioning helps". After the rebuttal, Reviewer n2Tg wrote, "My concerns are largely addressed. I am happy to maintain my positive rating." We also thank Reviewer kFNm, whose close reading led to three corrections and whose suggestions led to the reconstruct-and-render and RefV2V comparisons. After the discussion, kFNm raised the score and wrote that they "don't have more questions regarding the paper".

Building on these strengths, the main results added during the discussion period are as follows.

  1. Reconstruct-and-render, R&R (kFNm). Even with the ground-truth camera path and matched colors, R&R trails OmniShowcase on DINO (0.714 vs. 0.799) and MEt3R (0.312 vs. 0.190) on ShowcaseBench and on NAVI (0.762 vs. 0.828 and 0.317 vs. 0.249), while OmniShowcase receives only a trajectory sentence. GT-CD is lower for OmniShowcase on ShowcaseBench and comparable on NAVI, and each R&R sample takes longer, partly in human effort (Table 2, Appendix G.1). In the RefV2V refinement Reviewer kFNm suggested (supplementary video), MiniMax H3 restores appearance well but occasionally swaps the subject's sides or leaves the commanded path, while Wan VACE, from our backbone's family, can drift from the references.
  2. Robustness to reconstruction error (CUJd). With the references fixed, even a turntable reconstructed from a single view improves on the variant without normals in every metric on both benchmarks, and every metric improves steadily from one to four views. The single-view turntable already lowers ShowcaseBench GT-CD from 0.2092 to 0.1497 (Table 9, Appendix H.1). Appendix H.5 adds the case for normals over depth.
  3. Background consistency (n2Tg, kFNm). Our background MEt3R equals that of the ground-truth render (0.182 for both) at a similar camera rotation (0.58 vs. 0.62), while the baselines rotate at most 0.24. Showcase-4K clips are also more consistent than Uni3C clips of the same subjects and camera paths (0.138 vs. 0.186), and Section 3.2 now limits the 3D consistency claim to the subject (Table 10, Appendix H.2).
  4. Data scale and cost (n2Tg). With only 500 training videos, the model already surpasses every baseline on DINO and MEt3R on both benchmarks, and GT-CD keeps improving up to 4,000 videos (Table 4). One video takes 11.8 minutes, comparable to Phantom-MV and MV-S2V, with the lowest peak GPU memory (Table 11).
  5. Commercial models. To show that the task is not solved by scale alone, Appendix G.2 qualitatively compares Seedance 2.5 and Kling 4.0 with the same references and prompt. Both swap a teapot's handle and spout at a single frame, which OmniShowcase avoids because its normal turntable states the orientation explicitly.

The discussion supports our core contribution. When the camera reveals unseen sides, the difficulty is geometric rather than purely photometric (Section 1). OmniShowcase answers it with two parts, training data with exact subject geometry and a normal turntable that does not need frame alignment. The backbone adaptation is deliberately simple, and the ablations attribute the gains to these two parts. Removing the normal turntable more than doubles GT-CD on both benchmarks (Table 3). Showcase-4K, ShowcaseBench and the data-synthesis code will be publicly released.

Every analysis above came from the reviewers' questions, apart from the additional commercial-model comparison that further motivates the problem. These analyses have made OmniShowcase substantially stronger. We would be grateful if these results could be considered in the final discussion, and we thank you for handling our submission.

Best Regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.