acceptodds
Under review as a conference paper at ICLR 2027

ViSPaC: Visual Spatial Pose Codec for Camera-Controlled Video Generation

Abstract

Camera encodings and scene geometry have enabled substantial progress in camera-controlled video generation. We view camera control as a representation problem of relating geometric motion to visual dynamics, and propose ViSPaC, a Visual Spatial Pose Codec for joint modeling of camera motion and video. ViSPaC represents camera trajectories as evolving projections of colored landmarks in a synthetic 3D reference structure. These projections expose the visual effects of camera motion, while landmark identities and image positions establish correspondences for geometric recovery, without requiring reconstruction of the underlying scene. We identify two coupled constraints on this representation: visual compression limits the number of reliably distinguishable landmark identities, while their spatial distribution determines coverage across camera views. ViSPaC separates color identity from spatial placement and uses a spatial codebook to adapt the reference structure's scale and layout within a limited identity budget. A shared video diffusion transformer models video and visual pose, supporting both camera-controlled generation and video-conditioned pose recovery. Controlled comparisons show that ViSPaC improves camera control and pose recovery over alternative representations, with further gains from its identity budget and spatial adaptation. On DL3DV and Tanks and Temples, ViSPaC outperforms both implicit camera-conditioning and explicit scene-geometry methods, reducing relative translation and rotation errors by an average of 24.7% and 31.7%, respectively, over the previous state of the art.

Primary area:
generative models
Keywords:
Camera-controlled video generation, Visual pose representation, Video diffusion.

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_2xX7OpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ViSPaC, a method for camera-controlled video diffusion. Instead of doing scene reconstruction or using camera pose embeddings, this work projects synthetic 3D landmarks into 2D colored points to represent camera motion as an image sequence. The video and landmark images are then trained together in a diffusion transformer. The authors also use a spatial codebook to change the landmark layout for different camera paths, and test on DL3DV and Tanks and Temples.

Critical strengths
  • Interesting representation idea: Using synthetic 3D landmarks to turn camera poses into visual point tracks is an interesting idea and avoids doing explicit scene reconstruction.
Critical weaknesses
  • No video results at all: This is a video generation submission, but there are zero video results. There is no supplementary website, no link, and no zip file with videos. The paper only shows a few static image frames in the pdf. But video generation papers need videos to check temporal consistency, motion smoothness, camera precision, and artifacts like flickering. You cannot evaluate a video paper from static frames. Having no videos is enough to reject the paper right away.
  • Only static scenes: The paper only evaluates on DL3DV and Tanks and Temples, which are just static scenes. Real videos have moving objects, but it is not clear at all how this landmark projection handles dynamic scenes or if it completely breaks down when objects move.
  • Decoding seems very fragile: Decoding camera poses needs detecting tiny colored dots, reading their RGB values, and running PnP. As shown in Table 5, VAE compression creates a lot of color mismatch errors. Putting layout settings into tiny corner dots also seems very easy to break if the video has blur or artifacts.
  • Method is overly complicated: The pipeline has too many hand-crafted parts like discrete color bins, corner marker palettes, layout codebooks, and disk size rules. It is not clear why we need such a complex pipeline when other camera conditioning methods are much simpler.
Questions for authors

The idea is interesting, but having no video results for a video generation submission is not acceptable. I would like authors to address following questions:

  • Why are there no videos or supplementary website included? How can reviewers judge the video quality and camera control without seeing any videos?
  • How does the method handle dynamic scenes with moving objects?
  • How robust is the pose decoding when the generated frames have blur or compression artifacts?
  • Why do we need such a complex landmark codec when simpler camera conditioning methods already work well?

I am negative and vote for clear rejection because of the missing videos, but I am happy to see what authors say in the rebuttal.

9 comments
AuthorsOpenReview ↗

Response to Reviewer 2xX7 – Part 1/2

We thank the reviewer for recognizing our representation of camera motion without scene reconstruction. Below, we provide video evidence and address dynamic scenes, decoding robustness, and the codec's design rationale.


W1 / Q1: Missing video results

We apologize for the missing video evidence. The accompanying supplement (ViSPaC_Supplementary_Videos.html) provides synchronized side-by-side comparisons on TNT, DL3DV, and dynamic scenes.


W2 / Q2: Dynamic scenes with moving objects

The fixed landmarks form a synthetic 3D reference for camera motion and are not attached to scene objects. Their projections depend only on the prescribed trajectory and intrinsics, so fixing this reference does not require the generated scene to remain static.

We test this with a separate exploratory evaluation using 30 images selected from the VBench-I2V Image Suite, each paired with a prompt describing scene motion and a prescribed camera trajectory. The table reports camera-following errors, Dynamic Degree (DD), Motion Smoothness (MS), Imaging Quality (IQ), and Overall Consistency (OC).

Table R7. Camera control and generation quality on dynamic scenes.

MethodATE ↓RPE-t ↓RPE-r (°) ↓DD ↑MS ↑IQ ↑OC ↑
Lyra-20.07170.01530.07476.6798.9171.6327.0366
Gen3C0.05970.01040.0777‾\underline{0.0777}13.3398.5468.8927.1848
Uni3C0.09570.02600.107340.0098.3570.0727.5590
RealCam-I2V0.10480.02690.093210.0098.5270.86‾\underline{70.86}26.9231
Geometry Forcing0.09640.02240.100450.0097.7360.3125.0291
SCoPE0.11370.03510.138063.33‾\underline{63.33}97.8070.2627.3695
ViSPaC0.0678‾\underline{0.0678}0.0148‾\underline{0.0148}0.099766.6798.55‾\underline{98.55}70.3127.3745‾\underline{27.3745}

Notes: Best results are bold; second-best are underlined. VBench scores are multiplied by 100. DD includes camera-induced motion; OC measures video–text consistency.

ViSPaC achieves the highest DD, the second-lowest ATE and RPE-t, and the second-highest MS and OC. In the supplementary comparisons, ViSPaC generates object motion alongside camera motion, while Lyra-2 and Gen3C show comparatively limited object motion. Compared with RealCam-I2V, Uni3C, and SCoPE, which also generate scene dynamics, ViSPaC has lower ATE and RPE-t. Together, these results support camera control with dynamic scene content.


W3 / Q3: Robustness of pose decoding

The codec is designed to tolerate visual reconstruction errors through discrete color encoding with separated codewords. Both landmark identities and configuration choices are decoded by matching observed colors to discrete codes. Color perturbations that remain within the accepted region of the same code leave the recovered identity or configuration unchanged, so exact RGB reconstruction is unnecessary.

This tolerance can be quantified analytically. With RGB normalized to [0,1], correct color decoding is preserved when the absolute error in each channel of the estimated interior color is below the following bounds:

Table R8. Analytic error bounds for color readout.

Encoded informationNumber of identities / choicesPer-channel error bound
Landmark identity2100.0504
Landmark identity4320.0378
Configuration: scale60.360
Configuration: grid20.235
Configuration: layout80.178

For landmarks, we limit the number of identities to retain sufficient color separation. Table 5 directly examines identity recovery after VAE compression: mismatch among matched landmark detections is 0.19%/2.98% for the 210/432 identities used by ViSPaC, compared with 18.53% at 1,404 identities. This motivates the chosen identity budgets. We also estimate landmark colors from disk interiors, reducing the influence of color mixing at blurred boundaries.

The corner configuration markers follow the same discrete-coding principle, with only a few choices per marker—for example, eight layouts—represented by widely separated palette colors. Each marker occupies a fixed location in a reserved region and is rendered as a radius-9-pixel disk. We read the median color from its central 9×9 region and select the nearest palette entry. The separated palette preserves configuration decoding for color errors within these margins, while interior median aggregation reduces the influence of blurred boundaries.


AuthorsOpenReview ↗

Response to Reviewer 2xX7 – Part 2/2

W4 / Q4: Necessity and complexity of the landmark codec

Methods based on scene geometry provide explicit visual guidance, but reconstruction errors can introduce inaccurate or inconsistent constraints that the generator must accommodate. Dynamic scenes further complicate maintaining consistent geometry and correspondences across frames. In the supplementary video demonstrations, Lyra-2 and Gen3C show comparatively limited object motion. Uni3C can produce natural object motion, yet its static-scene results still deviate from the prescribed camera path despite point-cloud guidance. Pose- and ray-based conditioning encodes geometric relationships, but the generator must still learn how these relationships translate into image changes. The higher camera errors of these baselines relative to Lyra-2 illustrate the remaining gap in control accuracy.

ViSPaC exposes camera-induced visual motion through projections of a known synthetic reference, without depending on scene reconstruction. Its identifiable point trajectories express motion through the spatial patterns and temporal changes that video models are trained to process, while preserving geometric correspondences for pose recovery. For this representation to remain useful, the points need to be distinguishable after visual reconstruction and distributed so that the camera continues to observe enough of them along its path.

As discussed in Q3, maintaining color separation for reliable readout limits the number of landmark identities. With a fixed identity budget, spreading points over a larger region reduces their density, and a fixed layout may provide poor coverage for some trajectories. The spatial codebook therefore adapts scale and layout to the trajectory while preserving color identities. Configuration markers specify the mapping used for decoding, since the same color identity can refer to different 3D positions across configurations. Rendering landmarks as disks provides an interior region for color estimation, with disk size balancing this support against overlap.

Under the shared joint framework in Table 3, ViSPaC reduces ATE from 0.1161 for raw pose and 0.0922 for Raxel to 0.0227, while also improving pose readout. Removing spatial adaptation or increasing the identity count to 1,404 raises ATE to 0.0285 and 0.0398, respectively; both variants also degrade pose readout. These results support both the visual-pose representation and the contribution of its capacity and coverage choices to camera control and geometric recovery.


Thanks for the rebuttal. The dynamic video results are not very clear. What is the target camera trajectory? The camera is barely moving? Ideally also you show multiple different camera trajectories per dynamic video.

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

The paper introduces ViSPaC, a camera pose representation for camera controllable video generation and also video-to-pose. Instead of using numerical camera parameters, plucker rays, or explicit point cloud conditions for video diffusion model, ViSPaC place a synthetic 3D grid of colored landmarks in the first camer’s space and project them into 2D as colored disk along the camera trajectory. By using this codec as a camera pose condition, ViSPaC generates consistent camera controlled videos compared to other conditons, and also can predicts camera poses from video.

Critical strengths

[S1] The core intuition and method is very interesting and compelling. Video diffusion models are known to be good at tracking moving 2D patches over time, and turning camera motion into a moving colored disks plays directly to this strength without needing additional depth estimation or noisy 3D point cloud at test time.

[S2] Strong results: With ViSPaC, the method scored state-of-the-art performance comparing to other style camera conditions.

[S3] The motivation is clear and the controlled ablations are convincing.

Critical weaknesses

[W1] Does V2P training help P2V?: In the training step, half of budget was used for V2P task. However, there is no ablation on wheter joing V2P+P2V training atually improves P2V camera controllable video generation compared to only training P2V alone. Also, as a camera pose predictor (V2P), even with the fallback, the camera pose accuracy falls behind feed-forward reconstruction models, despite ViSPaC running on 50 diffusion steps on a 5B model.

[W2] ViSPaC is finetuned from Wan2.2 5B, but several baselines in the Table 1 and 2 use weaker video backbones or smaller training sets. Finetuning the same video backbone with ViSPaC and other camera conditioning methods and compare them would strengthen the method.

[W3] ViPE and DA3 were used to generate the training data, and the same two models (ViPE and DA3) are used as the evaluators for ATE/RPE in Table 1 and 2. This may give ViSPaC an inherent alignment advantage.

Questions for authors

[Q1] If you train the model on P2V only, how does the performance compare to the V2P+P2V joint training?

[Q2] In V2P, how is the retry fallback percentage?

[Q3] During P2V, does the DiT actually use the 3D codebook lookup, or does it just track the 2D motion of the disks? Specifically, what happens to P2V ATE/RPE at test time, if you 1) mask out the 64x64 configuration markers, or 2) randomly permute the landmark colors (so each disk keeps a temporally consistent color for 2D tracking, but its color is no longer encoding its 3D grid indes)?

4 comments
AuthorsOpenReview ↗

Response to Reviewer NuBH – Part 1/2

We thank the reviewer for recognizing our motivation, strong results, and controlled ablations. Below, we further examine how joint training and the pose representation contribute to camera control.


W1 / Q1: P2V-only versus joint P2V/V2P training

We thank the reviewer for raising this concern. We compare P2V-only and joint P2V/V2P training on the same 10K-clip ablation subset, considering both the total update budget and the number of P2V updates. With equal task sampling, the 10K-step joint model receives 5K P2V updates.

Table R3. P2V-only versus joint P2V/V2P training on DL3DV.

Training configurationTotal updatesP2V updatesPSNR ↑ATE ↓RPE-t ↓RPE-r (°) ↓
P2V-only5K5K13.790.02930.00770.116
P2V-only10K10K14.380.02410.00740.090
Joint P2V/V2P10K5K14.100.02270.00560.070

At the same 10K total updates, joint training reduces RPE-t and RPE-r by 24.3% and 22.2%, respectively, and also lowers ATE. These gains are obtained with only half as many P2V updates, suggesting that V2P provides useful training signals for camera control. P2V-only achieves higher PSNR (14.38 versus 14.10). When the number of P2V updates is matched, joint training improves both PSNR and camera accuracy over P2V-only 5K.


W1 / Q2: V2P retry and fallback frequency

V2P first attempts generation without fixing a synthetic first-frame visual pose. When the fraction of undecodable frames exceeds a threshold (10%), we retry with a fixed synthetic first-frame visual pose, starting from the standard configuration.

Table R4. V2P retry frequency and generation overhead.

DatasetClips triggering a retryMeeting the criterion after one standard retryMean generation passes
DL3DV1/110 (0.91%)1/11.01
TNT23/63 (36.51%)23/231.37

Retries are rarely needed on in-domain DL3DV but are more frequent on out-of-domain TNT. In both datasets, all retried clips meet the decoding criterion after a single retry.


W2: Comparisons with the same video backbone

We adapt SCoPE and Lyra-2 conditioning to the same Wan2.2-TI2V-5B backbone, using identical pretrained weights, the same 10K-clip ablation subset, 10K total updates, and the same DL3DV evaluation protocol. Both baselines train P2V-only; ViSPaC uses the joint model from Table 3, with an expected 5K P2V updates.

Table R5. Camera conditioning with a shared backbone on DL3DV.

Conditioning methodPSNR ↑ATE ↓RPE-t ↓RPE-r (°) ↓
SCoPE12.450.03580.01050.1951
Lyra-213.350.03400.00900.1023
ViSPaC14.100.02270.00560.0700

ViSPaC improves all four metrics under the matched backbone, training data, and total update budget. Compared with Lyra-2, ViSPaC improves PSNR by 0.75 dB and reduces RPE-t and RPE-r by 37.8% and 31.6%, respectively.


W3: Potential alignment advantage from pose estimators

Thank you for raising this concern. ViPE provides the DL3DV training trajectories. For RealEstate10K and Stereo4D, DA3 supplies a sequence-level translation-scale correction and informs sequence filtering, while preserving source rotations and trajectory shape. The remaining datasets retain their provided camera annotations.

Our evaluation removes global scale differences through similarity alignment, so DA3 provides a complementary check on trajectory geometry beyond its role in scale correction. Appendix D.1 shows that ViSPaC achieves the lowest ATE and RPE under DA3 and ViPE separately on both datasets under the standard 81-frame protocol. In the shared-data comparison above, ViSPaC also retains lower ATE and RPE when all three methods are trained with the same camera annotations.


AuthorsOpenReview ↗

Response to Reviewer NuBH – Part 2/2

Q3: Configuration markers and landmark color identities

Thank you for suggesting these tests. We examine whether configuration information and the color–space correspondence contribute to P2V beyond temporally consistent 2D point tracks.

With the model and generation settings fixed, we mask the configuration markers or randomly permute the landmark palette at test time. Both interventions preserve landmark positions and motion; permutation keeps each landmark's assigned color fixed throughout the clip.

Table R6. P2V sensitivity to configuration markers and landmark colors.

DatasetConditioning inputATE ↓RPE-t ↓RPE-r (°) ↓PSNR (dB) ↑
DL3DVOriginal0.009260.003610.0429015.90
DL3DVConfiguration markers masked0.009830.003710.0449815.59
DL3DVTemporally consistent palette permutation0.011280.003880.0456915.60
TNTOriginal0.036530.014400.0817015.76
TNTConfiguration markers masked0.052170.020450.1077415.30
TNTTemporally consistent palette permutation0.043780.017970.0972615.28

Both interventions increase ATE/RPE and reduce PSNR on both datasets. Color permutation increases ATE by 21.8% on DL3DV and 19.8% on TNT, despite preserving landmark motion and temporal identity. This suggests that the learned color–space correspondence contributes to camera control beyond temporally consistent point tracks. Masking configuration markers also degrades performance, with a larger effect on TNT.


Thank you for your thorough rebuttal and the additional experiments. All my concerns were resolved, I will maintain my positive score.

Dear Reviewer NuBH,

Thank you for taking the time to review our responses and the additional experiments. We are glad they resolved your concerns. We appreciate your questions about the role of joint training and your suggestion to test how the model uses landmark colors and configuration markers. Thank you for your constructive feedback and continued positive assessment of our work.

Best regards,
The authors

Reviewer_zQBbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ViSPaC, which uses colored landmarks to represent camera pose and achieve camera control video generation and pose estimation in one model. To address issues such as identity confusion—caused by visual compression—and the limited field of coverage provided by a sparse set of landmarks, the authors decouple color-based identity from spatial location, utilizing a spatial codebook to adjust the scale and layout of the dot pattern. Experiments on DL3DV / TNT and ablations demonstrate its effectiveness.

Critical strengths
  1. ViSPaC utilizes the projections of colored 3D landmarks as a unified camera representation, supporting camera-controlled video generation and pose estimation.

  2. The experimental results are good, demonstrating competitive performance in camera-control precision. Furthermore, the approach performs well in cross-dataset testing, highlighting the potential of this representation.

Critical weaknesses
  1. The robustness of ViSPaC under large-scale camera motion remains unclear. When camera movement causes a significant number of feature points to exit the field of view, can the remaining, limited distribution of 3D points still provide effective control signals for the camera? Such as large-angle rotations (e.g., 60°–90°).

  2. The validation of generative capability and generalizability could be strengthened. The results in the paper seem to be static-scene camera control video generation. If the scene itself is dynamic, can current fixed landmarks still do good camera control? Besides, this paper focuses on video generation, but I do not find video results, which are vital for evaluating the results.

Questions for authors

Could the authors clarify how ViSPaC performs under large camera motions (e.g., 60°–90° rotations) and in dynamic scenes with dynamic objects?

Are there video demonstrations / comparison that could help evaluate camera control accuracy and temporal consistency?

5 comments
AuthorsOpenReview ↗

Response to Reviewer zQBb

We thank the reviewer for recognizing our unified representation, camera-control accuracy, and cross-dataset performance. We provide additional evaluations and video comparisons addressing large camera motions and dynamic scenes below.


W1 / Q1: Large camera motions and landmark coverage

We mitigate coverage loss by adapting the reference structure's scale and layout to the full camera trajectory. The structure extends beyond the initial view, allowing newly visible landmarks to provide pose information as others leave. These landmarks belong to the same 3D reference, so conditioning does not require the same points to remain visible throughout the clip.

Training covers both substantial translation and rotation: 10.3% of clips reach a displacement of ≥8 m from the initial camera position, and 15.5% reach a rotation of ≥60°. Taking the sparsest frame in each clip, the median projected landmark counts are 53 and 52, respectively.

To assess whether this conditioning remains effective under the reviewer's example of 60°–90° rotations, we evaluate all 10 corresponding benchmark clips (8 DL3DV, 2 TNT). Their per-clip minimum projected landmark counts range from 34 to 145. On these clips, ViSPaC achieves the lowest camera-control errors and highest PSNR among the evaluated methods:

Table R1. Camera control on the 60°–90° rotation subset.

MethodPSNR ↑ATE ↓RPE-t ↓RPE-r (°) ↓
Lyra-212.550.02350.00750.1387
Uni3C12.830.04140.00980.2188
Gen3C12.310.03930.01460.3972
SCoPE11.470.09950.01600.3814
Geometry Forcing10.180.04280.01190.2375
ViSPaC14.330.01530.00430.0597

Notes: Best results are bold.

Compared with the remaining 163 clips with <60° rotation, our rotation RPE increases slightly from 0.0569° to 0.0597° (4.9%), while PSNR decreases from 15.94 to 14.33 dB. ViSPaC nevertheless retains its control advantage over the compared methods on these large-rotation clips.


W2 / Q1: Dynamic scenes and moving objects

The fixed landmarks form a synthetic 3D reference for camera motion and are not attached to scene objects. Their projections depend only on the prescribed trajectory and intrinsics, so fixing this reference does not require the generated scene to remain static.

We test this with a separate exploratory evaluation using 30 images selected from the VBench-I2V Image Suite, each paired with a prompt describing scene motion and a prescribed camera trajectory. The table reports camera-following errors, Dynamic Degree (DD), Motion Smoothness (MS), Imaging Quality (IQ), and Overall Consistency (OC).

Table R2. Camera control and generation quality on dynamic scenes.

MethodATE ↓RPE-t ↓RPE-r (°) ↓DD ↑MS ↑IQ ↑OC ↑
Lyra-20.07170.01530.07476.6798.9171.6327.0366
Gen3C0.05970.01040.0777‾\underline{0.0777}13.3398.5468.8927.1848
Uni3C0.09570.02600.107340.0098.3570.0727.5590
RealCam-I2V0.10480.02690.093210.0098.5270.86‾\underline{70.86}26.9231
Geometry Forcing0.09640.02240.100450.0097.7360.3125.0291
SCoPE0.11370.03510.138063.33‾\underline{63.33}97.8070.2627.3695
ViSPaC0.0678‾\underline{0.0678}0.0148‾\underline{0.0148}0.099766.6798.55‾\underline{98.55}70.3127.3745‾\underline{27.3745}

Notes: Best results are bold; second-best are underlined. VBench scores are multiplied by 100. DD includes camera-induced motion; OC measures video–text consistency.

ViSPaC achieves the highest DD, the second-lowest ATE and RPE-t, and the second-highest MS and OC. In the supplementary comparisons, ViSPaC generates object motion alongside camera motion, while Lyra-2 and Gen3C show comparatively limited object motion. Compared with RealCam-I2V, Uni3C, and SCoPE, which also generate scene dynamics, ViSPaC has lower ATE and RPE-t. Together, these results support camera control with dynamic scene content.


W2 / Q2: Video demonstrations and comparisons

We apologize for the missing video evidence. The accompanying supplement (ViSPaC_Supplementary_Videos.html) provides synchronized side-by-side comparisons on TNT, DL3DV, and dynamic scenes.


Reviewer_zQBbOpenReview ↗

Thank you for rebuttal

Thank you for the two additional experiments and video comparisons. Could you show the same dynamic scenes from the supplement with larger camera rotations (60°–90°)? If comparing methods on those clips is difficult, an I2V comparison using their first frames and the same camera trajectory across methods would also be helpful. Testing 60°–90° rotations on out-of-domain inputs would make the comparison with the baselines more convincing.

AuthorsOpenReview ↗

Follow-up to Reviewer zQBb

Thank you for suggesting larger rotations. We kept the same six images and motion prompts and added two paths per scene. All 12 paths combine translation with rotation, with peak target rotations ranging from 62.1° to 86° relative to the first frame. For each comparison, all seven methods received the same first frame and target path. The videos are under Dynamics → Large motion 1 / 2.

In these comparisons, GEN3C and Lyra-2 often show little independent subject motion. SCoPE animates the subject but struggles to follow the prescribed camera path. Uni3C produces convincing subject motion and follows the camera in several examples, though it misses parts of the requested movement in some scenes.

We chose VGGT-Omega for its strong camera pose estimation results on dynamic-scene benchmarks. We used it to estimate camera trajectories in all 84 generated videos and compared them with the target paths. The table reports mean errors over the 12 trajectories for each method.

MethodATE ↓RPE-t ↓RPE-r (°) ↓
Lyra-20.34470.09510.5503
GEN3C0.19800.08940.6423
Uni3C0.71450.21100.9237
RealCam-I2V0.61680.24391.0157
Geometry Forcing1.07820.27632.0026
SCoPE1.69320.44651.5745
ViSPaC0.31450.15900.3759

GEN3C achieves the lowest mean ATE and translational RPE, but shows limited independent subject motion in these comparisons. ViSPaC achieves the lowest mean rotational RPE across all seven methods and lower mean errors on all three metrics than Uni3C, SCoPE, and RealCam-I2V, which also generate scene dynamics.

The Extreme rotations tab also compares six methods on one DL3DV scene with the camera center fixed during a uniform 90° turn, a fast 90° turn followed by a hold, and a uniform 180° turn.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.