acceptodds
Under review as a conference paper at ICLR 2027

SCoPE: Sightline-Coordinate Positional Encoding for Video Diffusion Transformers

Abstract

Video diffusion transformers address their tokens by position on the pixel-time grid: an address in the tensor, not in the world. The address we would want, the world point a token depicts, lies on a surface not yet generated, while its camera ray is fixed once the user specifies a trajectory. SCoPE therefore treats the ray as a second positional coordinate, and camera control becomes a property of the coordinate system, not an added module. The ray is added to the pretrained attention's queries and keys, and the score gains a term that reads the two rays alone. Its canonical raw-ray form recovers the reciprocal product, a classical coplanarity test for lines of sight. Normalize-Gate-Inject makes a single encoding trainable across metric and up-to-scale pose sources. The retrofit keeps RoPE bit-exact, starts from the unchanged pretrained DiT, and adds under 0.1% new parameters. On Wan2.2 at 5B and 14B under matched data and budget, SCoPE improves every camera-controllability and fidelity metric, leads all closed-loop revisit metrics, and shows widening margins with model size. At 14B, rotation error falls 29% and FVD 43% below the strongest baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video generation, positional encoding, diffusion transformer

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

SCoPE introduces camera control into video diffusion transformers by adding projected Plücker ray features to queries and keys after the pretrained RoPE. This produces content–geometry interactions and a geometry-only attention term. A Normalize-Gate-Inject mechanism handles variations in pose scale and feature magnitude. Experiments on Wan2.2 at 5B and 14B demonstrate improvements in camera controllability, visual quality, and closed-loop consistency.

Critical strengths
  1. The additive formulation is simple and practical (Section 4). It introduces explicit ray–ray and content–ray interactions while preserving pretrained RoPE and avoiding an additional attention branch. This provides a useful alternative to multiplicative camera encodings.

  2. The empirical evaluation covers two backbone scales and complementary aspects of generation, including camera accuracy, visual quality, dynamic scenes, and closed-loop revisits (Tables 1–3). The substantial reported improvements across these settings support the practical value of the method.

Critical weaknesses
  1. The conceptual advance is moderate, and the geometric interpretation has a limited scope. The reciprocal-product equivalence holds for a particular raw-ray configuration, while the practical encoding includes unconstrained projections, normalization, and gating. The flip is also an equivalent reparameterization, as acknowledged by the authors. Thus, the theoretical connection does not establish that the learned attention implements the corresponding geometric compatibility test.

  2. Some experimental details remain insufficiently specified. The cross-term ablation needs clarification because retaining zero initialization for both geometry branches could prevent the remaining ray–ray term from learning (Table 4 and Appendix G.4). Additionally, the trainable backbone components and optimization budgets should be documented for each baseline to substantiate the controlled comparison.

Questions for authors
  1. How was the “w/o (B)+(C)” ablation initialized and trained? If both geometry outputs start at zero, their remaining bilinear interaction gives zero gradients to both branches. Was initialization changed, a trained checkpoint used, or the intervention applied only at evaluation?

  2. Could you provide the trainable backbone components, total trainable parameter counts, and training budgets for each baseline? In particular, were backbone fine-tuning settings matched to SCoPE?

4 comments

We thank the reviewer for recognizing the practical value of SCoPE's additive attention design and its improvements at both scales and on closed-loop revisits. We address the initialization concern, baseline training fairness, and the scope of the geometric interpretation below.

Q1. Initialization of w/o (B)+(C)

This was a separately trained ablation, with (B) and (C) removed throughout training rather than only at evaluation. Both Q/K geometry projections were initialized with small nonzero Gaussian weights (mean 0, std 0.01), with the geometry scaling factor initialized to α=1\alpha=1. This differs from the full-model initialization because, after removing the content–geometry cross terms, (D) is the only remaining geometry-dependent term.

Specifically, the retained ray–ray term (D) is a dot product between the Q- and K-side geometry features. Its gradient with respect to either feature equals the feature from the opposite branch. If both geometry outputs are zero, gradients to both projections vanish. With small nonzero initialization and an active geometry scale, both projections can receive gradients from the first training step, avoiding this deadlock. Thus, the ablation evaluates a trainable ray–ray-only interaction rather than a degenerate geometry branch. We will clarify this initialization exception in Appendix G.4.

Q2. Baseline training budgets and finetuning fairness

The comparisons in Tables 1–3 span 5B and 14B. Within each scale, methods use the same pretrained backbone and matched training/evaluation protocols. For Wan2.2-I2V-A14B, all methods use the same four-dataset mixture, 30k optimization steps, and 81-frame, 480 × 832 first-frame I2V setting. Only the high-noise expert is finetuned for every 14B method; the low-noise expert remains frozen. The exact trainable modules and parameter counts are listed in Table R3 of our response to Reviewer k9tv. Importantly, the optimization budget is matched, but the methods' default trainable scopes differ.

To test whether SCoPE benefits merely from updating more backbone parameters, we conducted additional 14B controls in both directions: (i) expand UCPE and ReRoPE to train backbone self-attention, FFN, and norm3, and (ii) restrict SCoPE to backbone Q/K projections plus its geometry modules. All variants use the same 30k-step budget.

Table R6. Controlled finetuning comparisons (Wan2.2-I2V-A14B).

MethodRotErr ↓TransErr ↓CamMC ↓FVD ↓
UCPE0.0820.6930.760529.42
UCPE (+ backbone FT)0.0790.6880.755577.30
ReRoPE0.1140.8200.905493.30
ReRoPE (+ backbone FT)0.1120.8170.895527.58
SCoPE (Q/K FT)0.0620.5240.541261.54
SCoPE (full FT)0.0580.5170.530280.17

Expanding baseline finetuning yields only marginal camera-accuracy gains, while FVD worsens (UCPE: 529.42 → 577.30; ReRoPE: 493.30 → 527.58). Conversely, SCoPE with Q/K-only backbone finetuning retains nearly the same RotErr (0.062 vs. 0.058) and achieves better FVD (261.54 vs. 280.17) than full SCoPE. Moreover, PRoPE already trains more parameters than SCoPE (10.908B vs. 10.148B for the high-noise expert), yet achieves weaker RotErr (0.097 vs. 0.058) and FVD (472.11 vs. 280.17). These 14B controls indicate that the gains cannot be explained simply by differences in backbone finetuning scope. We will document the configurations and controls in the revision.

W1. Geometric interpretation

We agree with the reviewer that the Plücker reciprocal-product identity describes the canonical raw-ray pairing, not an exact compatibility test enforced by learned attention. After learned projections, normalization, and gating, (D) is a general geometry-conditioned bilinear interaction. The Q/K flip can be absorbed into an unconstrained key projection, so it does not add expressivity; its role is to motivate the canonical coordinate ordering.

The substantive distinction is in the attention-score structure. With only (A)+(D), the appearance and ray contributions are additive and separable: A(xi,xj)+D(ri,rj)A(x_i,x_j)+D(r_i,r_j). Terms (B) and (C) instead couple visual features and camera rays, allowing geometry-dependent attention to vary with token content. This preserves pretrained RoPE and requires no parallel attention branch. The 5B ablations in Table 4 support the usefulness of both interactions: removing (B)+(C) changes RotErr/FVD from 0.085/543.17 to 0.135/695.41, while removing (D) changes them to 0.159/733.08. We will distinguish this empirical benefit from the narrower canonical geometric identity in the revision.

Dear Reviewer,

Thank you again for your valuable feedback. We have clarified the initialization of the ablation experiments and provided additional details on baseline training configurations and comparisons. We would sincerely appreciate your thoughts on our responses and are happy to address any remaining questions.

Thank you for your time and consideration.

Reviewer_xwMEOpenReview ↗

Comments by Reviewer xwME

Thank you for the clarifications and additional experiments. The initialization explanation addresses my concern about the ray–ray-only ablation, and the controlled fine-tuning results substantially alleviate the concern about baseline fairness. I also appreciate the more precise scope of the geometric interpretation. I maintain my weak-accept recommendation.

Thank you very much for your positive feedback and for taking the time to review our additional experiments and clarifications. We sincerely appreciate your recognition of our rebuttal efforts and your continued support.

We will incorporate these clarifications and experimental details into the revised manuscript. Thank you again for your thoughtful review and consideration!

Reviewer_k9tvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

SCoPE is another method of controlling camera in generative video diffusion models. It adds camera-ray features directly to attention queries & keys, different from adding attention branches (UCPE) or replacing pos encodings (ReRoPE). They also have Normalize-Gate-Inject (NGI) approach, which normalizes ray features and controls their strength to handle different scalings of cam coordinates. SCoPE adds less than 0.1% new parameters to the 5B model (although parts of the backbone are also fine-tuned). On Wan2.2-14B, the paper reports 29% lower rotation error than UCPE and 43% lower FVD than ReRoPE.

Critical strengths
  1. The paper claims adding ray features directly to attention improves camera control compared to prior work. That benefits both visual quality (Tab 1-2) and runtime performance (Tab A3).
Critical weaknesses
  1. The abstract and some sections (Sec 4) are wordier than needed, which makes them hard to follow. The manuscript would benefit from clearer, simpler language.

  2. Training and advantage over baselines are unclear.

The “under 0.1%” in the abstract refers to added parameters, but SCoPE also fine-tunes substantial backbone layers (Appendix G.4). This raises the question of how the baselines were trained, and whether the benefit comes from the proposed components or from re-training existing parameters. For the “matched data and budget” re-implemented baselines, lines 308-309 refer to Appendix J, but no related details are present there.

In addition, some baselines don’t follow their original designs and some comparisons are not present. PRoPE is evaluated as a zero-initialized parallel branch instead of their original architecture (Appendix A.4). Similarly, ReCamMaster is a video-to-video model, and it’s not explained how it was adapted to the image-to-video. ReRoPE (the strongest FVD baseline at 14B) is also missing from Tab 2-3.

Together, this makes it very hard to compare methods fairly.

  1. The benefit of the Q/K flip is unclear.

Proposition 1 motivates swapping d and m on the key side by showing that the attention score can recover a known geometric formula. However, the authors don’t verify that this relationship holds in the trained model, and training doesn’t enforce it. Appendix J.5 also shows that the learned projection can absorb the flip anyway, so the flip shouldn’t change what the model can learn. It's therefore unclear what causes the FVD gain with the flip in Tab 4 (543.17 vs 579.83), and how it relates to the geometric motivation in Proposition 1.

  1. NGI’s claimed scale robustness needs a direct ablation, as it's not clear.

Preprocessing datasets already brings camera translations into roughly similar ranges across datasets and scenes (Appendix G.2–G.3). As far as I understand, NGI is meant to handle remaining scale differences and balance camera features against image features. Tab 4 shows that NGI helps on average, but it doesn’t directly test scale robustness. Using the same preprocessing, the authors should compare camera-control errors with and without NGI across clips with different translation magnitudes.

Questions for authors
  1. [W2] How were baselines adopted with the same training and data budget (especially PRoPE and ReCamMaster), and how does ReRoPE perform on Tab 2-3?

  2. [W3] What causes the flip’s FVD gain in Tab 4, given Appendix J.5?

  3. [W4] How does NGI affect camera errors across clips with different translation magnitudes?

5 comments
AuthorsOpenReview ↗

Response to Reviewer k9tv

We thank the reviewer for recognizing SCoPE's camera-control, video-quality, and efficiency gains. We clarify baseline implementations, matched finetuning, the Q/K flip, and NGI below.

W2. Baseline implementations and comparison fairness

W2a. Matched training settings

All Wan2.2-I2V-A14B methods use the same pretrained backbone, four-dataset mixture, 30k steps, and evaluation protocol (81 frames, 480 × 832). All are adapted to first-frame I2V.

PRoPE: Its projective encoding is placed in a parallel camera-attention branch because directly replacing pretrained RoPE was unstable in our initial Wan2.2 experiments. The camera branch, backbone self-attention, FFN, and norm3 are trained.

ReRoPE: We reproduced its published frequency-splitting design (training code was unavailable at the time): high-frequency temporal RoPE is retained; low-frequency channels use camera-aware rotary encoding. The standard setting trains self-attention.

ReCamMaster: In our Wan2.2 I2V adaptation, the first ground-truth frame is fused into the target latent without a source-video latent. Camera encoding is injected before self-attention; the camera encoder, projector, and self-attention are trained.

Table R3. Training configurations (Wan2.2-I2V-A14B).

MethodTrainable modulesTrainable params / expertTraining steps
SCoPEPlücker encoding, self-attn, FFN, norm310.148B30k
PRoPE (parallel)Camera attention branch, backbone self-attn, FFN, norm310.908B30k
UCPEParallel camera adapter; pretrained DiT frozen0.525B30k
ReRoPESelf-attn4.196B30k
ReCamMasterCamera encoder, projector, self-attn5.247B30k

W2b. Controlled finetuning comparison

To control for finetuning scope, we additionally train self-attention, FFN, and norm3 for UCPE and ReRoPE. Conversely, SCoPE (Q/K FT) trains only backbone Q/K projections and geometry-encoding modules. All retain the matched 30k-step budget.

Table R4. Camera control and video quality under controlled finetuning (Wan2.2-I2V-A14B).

MethodRotErr ↓TransErr ↓CamMC ↓FVD ↓
UCPE0.0820.6930.760529.42
UCPE (+ backbone FT)0.0790.6880.755577.30
PRoPE (parallel)0.0971.0601.117472.11
ReRoPE0.1140.8200.905493.30
ReRoPE (+ backbone FT)0.1120.8170.895527.58
ReCamMaster0.1090.9761.012675.23
SCoPE (Q/K FT)0.0620.5240.541261.54
SCoPE0.0580.5170.530280.17

Expanding UCPE and ReRoPE finetuning yields only marginal camera-accuracy gains, while FVD worsens (529.42 → 577.30 and 493.30 → 527.58). SCoPE retains nearly the same camera accuracy with Q/K-only backbone finetuning and achieves lower FVD (261.54 vs. 280.17). Under the submitted full-finetuning setting, it achieves 29% lower RotErr than UCPE and 43% lower FVD than ReRoPE. These controls support the effectiveness of our method beyond differences in backbone finetuning.

ReRoPE in Tables 2–3: We will add its results to both tables in the revised PDF and share the updated version subsequently.

W2c. Inference efficiency

SCoPE injects geometry through the existing Q/K projections and requires no parallel attention computation. The submitted Appendix B.3 reports the following 5B inference costs:

Table R5. Inference cost (Wan2.2-TI2V-5B).

MethodStep latency ↓Peak memory ↓
UCPE27.6 s21.8 GB
PRoPE34.6 s22.3 GB
SCoPE24.4 s11.6 GB

Relative to UCPE, SCoPE reduces step latency by 12% and peak memory by 47%; relative to PRoPE, the reductions are 30% and 48%.

W3. Q/K flip and geometric interpretation

The Q/K flip swaps the direction and moment components of the key ray, so that the ray–ray dot product recovers the Plücker reciprocal product, di⊤mj+mi⊤djd_i^\top m_j + m_i^\top d_j, in the canonical 6D setting (Eq⊤Ek=IE_q^\top E_k=I). With unconstrained learned projections, however, the flip can be absorbed into the key projection, making the two parameterizations equivalent in expressivity. Thus, the flip provides a geometrically motivated coordinate convention rather than enforcing the reciprocal product. The observed FVD difference (543.17 vs. 579.83) is empirical, not a consequence of greater expressivity. We will clarify this distinction in the revision.

W4. NGI scale robustness

Our 14B scale-stratified evaluation (Table R2: Camera control across translation magnitudes; also sent to Reviewer 4iaa) compares models with/without NGI on the same RE10K clips, with matched training steps and trajectory preprocessing. NGI improves accuracy for small, medium, and large translations. For large motions, TransErr falls from 1.424 to 0.959 (32.7%) and CamMC from 1.470 to 0.970 (34.0%), supporting NGI's scale robustness.

Dear Reviewer,

Thank you again for your detailed and thoughtful review. We have addressed your concerns regarding baseline fairness, geometric formulation, and NGI scale robustness.

We would also like to share an additional evaluation of NGI across different camera translation magnitudes on RE10K using Wan2.2-I2V-A14B:

TranslationMethodRotErr ↓TransErr ↓CamMC ↓ATE ↓
Small (0.03–0.60)SCoPE0.0240.0810.0920.093
w/o NGI0.0250.0950.1030.110
Medium (0.60–1.20)SCoPE0.0330.2080.2170.249
w/o NGI0.0360.2300.2330.338
Large (1.20-3.00)SCoPE0.0510.9590.9701.137
w/o NGI0.0601.4241.4701.522

NGI consistently improves camera control across all translation magnitudes, with particularly pronounced gains for large camera motions (32.7% lower TransErr and 34.0% lower CamMC).

We would greatly appreciate it if you could take another look at our responses and share any remaining concerns. Thank you for your time and consideration.

Dear Reviewer,

Thank you again for your detailed review. We would like to highlight several additional results addressing your main concerns.

1. Additional ReRoPE comparisons.

We have completed the previously missing ReRoPE evaluations for both dynamic-scene generation and closed-loop revisits, following the corresponding evaluation protocols.

Closed-loop revisit evaluation (Table 3).

MetricReRoPESCoPE
Return Trans. Ratio ↓0.5750.090
Return Rot. (°) ↓3.591.89
Norm. ATE ↓0.4300.275
Return LPIPS ↓0.4730.230
LPIPS Recovery ↑0.1050.587
DINOv2 Sim. ↑0.9170.952

Dynamic-scene evaluation (VBench; Table 2).

MetricReRoPESCoPE
Subject Consistency ↑0.9030.918
Background Consistency ↑0.9340.945
Motion Smoothness ↑0.9750.983
Dynamic Degree ↑0.7900.855
Aesthetic Quality ↑0.5220.561
Imaging Quality ↑0.6350.689

SCoPE outperforms ReRoPE across all reported metrics, demonstrating stronger camera consistency, appearance recovery, and dynamic-scene generation quality.

2. Baseline fairness and controlled finetuning.

Our additional 14B experiments (Table R4 in our rebuttal) directly examine the effect of backbone finetuning. Expanding backbone finetuning for UCPE and ReRoPE yields only marginal camera-control gains while degrading FVD. In contrast, SCoPE maintains strong camera accuracy with only the backbone Q/K projections and geometry modules trained, achieving an FVD of 261.54 compared with 280.17 under full finetuning. These results further support the effectiveness of the proposed geometry-aware attention design.

3. NGI scale robustness.

Our additional 14B evaluation shows that NGI consistently improves camera control across different translation magnitudes. For large camera motions, NGI reduces TransErr by 32.7% and CamMC by 34.0% compared with the variant without NGI.

We will also improve the clarity of the abstract and Section 4, clarify the distinction between additional and trainable parameters, and correct the appendix references.

As the discussion period approaches its end, we would greatly appreciate it if you could revisit our responses and share any remaining questions or concerns. Thank you very much for your time and consideration.

Reviewer_4iaaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To generate a video along a specified camera path, a video diffusion transformer needs to relate different views of the same scene. Standard positional encoding gives patches image and time coordinates. However, it does not describe the camera viewpoint. This paper proposes SCoPE (Sightline Coordinate Positional Encoding) to give each patch a camera ray that describes its viewing direction from the camera position. These rays are calculated from the camera settings and requested path before generation. SCoPE adds ray features to attention to help the model link patches that could show the same part of the scene using both their appearance and camera geometry. This contributes to preserving scene content as the viewpoint changes. Because camera datasets use different coordinate scales, ray features can become too large/small relative to image features. To address this, this paper also introduces Normalize-Gate-Inject (NGI) to normalize the ray features and learn their influence on attention and balance camera guidance with image appearance.

SCoPE was tested on the 5B/14B versions of Wan2.2 against baselines using the same backbone, training data and budget. Reported results show improved camera path accuracy, video quality, and recovery of the original view after the camera moves away and returns. Ablations on the 5B model show the largest overall performance drops when direct comparisons between rays or their interactions with appearance are removed.

Critical strengths
  1. Adding camera ray features to queries and keys creates a term based only on the two rays, alongside terms combining camera information with appearance. This gives attention a signal from the known camera path while the video content is still being generated. The design also keeps the pretrained positional encoding and requires no separate attention branch.

  2. The paper reports performance improvements in both camera control and video quality when SCoPE is applied to Wan2.2. On held-out RE10K videos, rotation error, trajectory error and FVD improve over methods using the same backbone, data and training budget. The paper also reports experiments where the camera moves away and returns to its starting viewpoint. SCoPE returns closer to the correct camera pose and recovers a view more similar to the original than the compared methods. This shows that it can keep the scene consistent during camera motion.

Critical weaknesses
  1. The paper does not show that the proposed ray encoding gives an advantage over a generic learned embedding of the same camera information added to the queries and keys in the same way. The reported ablations evaluate the importance of the ray related attention terms. However, they do not isolate the benefit of the proposed encoding.

  2. The reported experiments do not isolate NGI’s benefit for handling different camera coordinate scales. Appendix G describes adjustments that already make the datasets scales more comparable before NGI is applied. The experiment without NGI removes three components at oncer: the direction magnitude split used to represent camera scale, the gate that controls how much camera information is added to attention, and the normalization that rescales those camera features. The performance drop shows that these components help together but does not show that they help specifically by handling different camera scales. Some of the benefit could come from keeping camera and image features balanced even when the camera scales are already comparable.

Questions for authors
  1. Could you replace the ray encoding with a learned embedding of the same camera position and viewing direction, keeping the way it is added to attention and everything else unchanged? This would show whether the proposed ray encoding gives an advantage over a simpler representation of the same camera information.

  2. The datasets are already rescaled to make their camera coordinate scales more comparable. In addition, the experiment without NGI also removes the gate and normalization. Do you have a direct test showing that NGI improves camera path accuracy or video quality across different camera coordinate scales?

  3. Can you report results over multiple seeds at least for the 5B main comparison and the ablations?

4.In the experiment without the appearance ray interactions, was the model trained with these interactions removed or were they removed only during testing? If it was trained without them, did both camera branches start by outputting zero? If yes, how did the remaining ray-ray term receive a learning signal?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 4iaa

We thank the reviewer for recognizing SCoPE's improvements in camera control, video quality, and closed-loop consistency.

Q1. Generic learned camera embedding.

We would like to clarify the central contribution of SCoPE. SCoPE already uses learned camera embeddings through trainable projections EqE_q and EkE_k. Our key contribution is introducing camera rays as an additional geometric positional coordinate in pretrained self-attention. By injecting these learned ray features directly into Q/K, SCoPE enables geometry–geometry and content–geometry interactions, while preserving pretrained RoPE and avoiding additional attention branches.

We adopt Plücker coordinates [d,m][d,m], where m=o×dm=o\times d, as a standard geometric representation of viewing rays. The projections EqE_q and EkE_k learn how this geometric information interacts with visual features. A generic MLP operating on camera center and ray direction [d,o][d,o] can also approximate the Plücker mapping. The distinction lies primarily in the geometric inductive bias provided by the input representation.

To further examine this choice, we compare Plücker coordinates with a direct camera-center representation under the same backbone, training data, and Q/K injection mechanism.

Table R1. Camera representation (Wan2.2-TI2V-5B).

RepresentationRotErr ↓CamMC ↓FVD ↓
Plücker [d,m][d,m] (SCoPE)0.0850.802543.17
Center + direction [d,o][d,o]0.0900.816551.30

Both representations work effectively with our attention formulation, with Plücker coordinates achieving better camera accuracy and video quality.

More importantly, the submitted interaction ablations demonstrate the contribution of geometry-aware attention pairing. Removing the geometry-only ray–ray term increases RotErr from 0.085 to 0.159 and FVD from 543.17 to 733.08. These results highlight the importance of explicitly incorporating ray geometry into the pretrained attention mechanism.

Q2. NGI across different camera translation magnitudes.

NGI decomposes direction and magnitude, controls the strength of geometry injection, and matches the scale of injected features to the pretrained attention. We evaluate its effect across on RealEstate10K (RE10K) clips. Both variants use matched training steps, trajectory preprocessing, and evaluation settings. Table R2 uses additional Wan2.2-I2V-A14B checkpoints; the comparison with and without NGI uses the same clips and will be reported with matched training steps and trajectory preprocessing. SCoPE consistently outperforms the variant without NGI across all translation magnitudes, with particularly pronounced gains for large camera motions.

Table R2. Camera control across translation magnitudes (Wan2.2-I2V-A14B; additional evaluation).

TranslationMethodRotErr ↓TransErr ↓CamMC ↓ATE ↓
Small (0.03–0.60)SCoPE0.0240.0810.0920.093
w/o NGI0.0250.0950.1030.110
Medium (0.60–1.20)SCoPE0.0330.2080.2170.249
w/o NGI0.0360.2300.2330.338
Large (1.20-3.00)SCoPE0.0510.9590.9701.137
w/o NGI0.0601.4241.4701.522

Q3. Multiple seeds.

We additionally evaluate SCoPE, UCPE, and ReRoPE using five independent diffusion sampling seeds on the same test set, with fixed 5B checkpoints. The results show that SCoPE consistently outperforms the competing methods across sampling seeds, supporting the robustness of the reported improvements.

MethodRotErr ↓TransErr ↓CamMC ↓FVD ↓
UCPE0.113 ± 0.0040.860 ± 0.0300.906 ± 0.032709.55 ± 21
ReRoPE0.135 ± 0.0051.112 ± 0.0391.211 ± 0.043677.20 ± 21
SCoPE0.084 ± 0.0030.750 ± 0.0260.805 ± 0.028539.36 ± 16

Q4. Initialization of the w/o (B)+(C) ablation.

For the w/o (B)+(C) ablation, both Q/K geometry projections were initialized with small, nonzero random weights (mean 0, std 0.01), and the geometry scaling factor was initialized to α=1\alpha=1. This setting was chosen specifically for the ablation, where the content–geometry cross terms are removed and the ray–ray interaction is the only geometry-dependent term.

The retained term is qr⊤krq_r^\top k_r. Its gradients with respect to qrq_r and krk_r are krk_r and qrq_r, respectively. Since both geometry branches produce nonzero outputs at initialization, gradients propagate to both projections from the first training step. The ablation therefore does not suffer from the zero-gradient deadlock raised by the reviewer.

This experiment evaluates the contribution of the content–geometry interactions under a trainable ray–ray-only formulation. We will clarify the ablation-specific initialization in the revised appendix.

Dear Reviewer,

Thank you again for your constructive feedback. We have addressed your questions and provided additional experiments on camera representations, NGI scale robustness, and sampling-seed stability. We would greatly appreciate it if you could take another look at our response. We are happy to clarify any remaining concerns.

Thank you for your time and consideration.

We further assess the robustness of SCoPE using three independently trained 5B checkpoints with different random seeds, in addition to evaluating SCoPE, UCPE, and ReRoPE across five diffusion sampling seeds.

The sampling-seed evaluation shows consistent improvements in camera control and video quality over the competing methods. The independent training runs provide an additional assessment of SCoPE's optimization stability.

We report the mean and standard deviation across training runs below.

MethodRotErr ↓TransErr ↓CamMC ↓FVD ↓
SCoPE (3 training seeds)0.083 ± 0.0050.747 ± 0.0230.791 ± 0.031533.50 ± 10

The sampling-seed results confirm that SCoPE maintains a consistent advantage over UCPE and ReRoPE in camera control and video quality. The independent training-seed evaluation further assesses the stability of SCoPE's learned geometry-aware attention.

We would greatly appreciate it if you could revisit our rebuttal and share any remaining questions or concerns before the discussion period concludes. Thank you again for your time and thoughtful consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses

Dear Reviewers,

We thank reviewers 4iaa, k9tv, and xwME for their careful evaluation and constructive feedback. We appreciate your recognition of the practical additive attention design (4iaa, xwME), the improvements in camera control and video quality (all three reviewers), the inference efficiency (k9tv), and the closed-loop consistency (4iaa, xwME).

SCoPE integrates camera-ray features directly into the existing queries and keys, preserving pretrained RoPE while introducing ray–ray and content–ray interactions without an additional attention branch. This additive formulation allows attention to combine visual content with the prescribed viewing geometry. NGI controls the magnitude and strength of the injected features across heterogeneous pose sources.

We have incorporated the following experiments and clarifications into the revised manuscript. Revisions in the main text are highlighted in blue for ease of reference. The individual responses provide further discussion.

  • Baseline configurations and controlled finetuning. We document the baseline adaptations, trainable modules, parameter counts, and training budgets. We also add 14B controls that expand backbone finetuning for UCPE and ReRoPE and restrict SCoPE to Q/K projections, with camera modules remaining trainable. SCoPE retains lower camera errors and FVD across these settings. Section 5.4.2, Table 5. See our responses to k9tv, W2a–W2b, and xwME, Q2.

  • NGI across translation magnitudes. We compare SCoPE with and without NGI on the same clips and preprocessing, grouped by camera translation magnitude. NGI improves camera accuracy in all three groups; for large motions, translation error and CamMC decrease by 32.7% and 34.0%, respectively. The revision also specifies the grouping rule. Section 5.4.3, Table 6. See our responses to 4iaa, Q2, and k9tv, W4.

  • Geometric interpretation and attention interactions. We clarify the conditions of Proposition 1 and the role of the Q/K flip, and explain how the content–geometry terms make attention depend jointly on visual features and viewing rays. Section 4.1. See our responses to k9tv, W3, and xwME, W1.

  • Training and initialization of the interaction ablation. We specify that the ray–ray-only variant is trained with the content–geometry cross terms removed throughout optimization. Its geometry projections use small nonzero initialization with α = 1, allowing both projections to receive gradients from the first step. Appendix G.4. See our responses to 4iaa, Q4, and xwME, Q1.

  • Additional ReRoPE comparisons. We have added the requested ReRoPE results for dynamic scenes and closed-loop revisits. Tables 2–3. These complete the comparisons discussed in our response to k9tv, W2.

Thank you again for your time and feedback. We welcome further questions and discussion.

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.