acceptodds
Under review as a conference paper at ICLR 2027

WorldCast: World-Consistent Character Video Generation with Large-Angle Camera Control

45 views

Abstract

Precise camera control is essential for cinematic character video generation, yet existing methods struggle to follow large-angle trajectories that combine rotation, translation, and radius variation, with arbitrary initial azimuths. Such viewpoint changes also reveal character regions unseen in the initial frame, often leading to geometric distortions, appearance degradation, and identity drift. To systematically study these coupled challenges, we introduce CLACT, a large-scale Character-centric dataset and benchmark covering diverse Large-Angle Camera Trajectories, comprising 80,000 synthetic training samples and 42 synthetic and real-world evaluation scenes, with accurate camera geometry and multi-view character observations. We further propose WorldCast, a framework that couples scale-aligned geometric guidance with geometry-aligned multi-view character memory. Specifically, we back-project the initial view into a 3D point cloud and render it along the target trajectory to provide trajectory-aligned geometric guidance and coverage masks. To complement character information missing from the initial view, we introduce Geometry-Aligned Memory Attention (GAMA), which aligns multi-view anchor geometry with target-view geometry in a shared coordinate frame, enabling geometry-aware retrieval and injection of view-relevant character features. Experiments across diverse large-angle trajectories demonstrate improved camera-control accuracy and character consistency over the compared methods while maintaining competitive visual quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Controllable Video Generation, Camera-Controlled Video Generation, Large-Angle Camera Motion, Character Consistency, Multi-View Conditioning

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_2PoHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WorldCast combines projected geometric guidance with geometry-aligned multi-view character memory to generate identity-consistent videos under large-angle camera motion. It introduces CLACT and demonstrates improved camera adherence and cross-view consistency.

Critical strengths
  1. CLACT provides diverse large-angle camera trajectories and multi-view character observations, supporting joint evaluation of camera control and character consistency.
  2. Quantitative results and ablations support the effectiveness of GAMA, whose removal increases MEt3R from 0.228 to 0.333 (Table 2).
Critical weaknesses
  1. How are real-world anchors acquired, and how does alignment handle collinear trajectories or anchors mapped to the same target frame?
  2. Can you report fixed-scale camera errors on synthetic scenes and character consistency against held-out views?
  3. How does GAMA compare with standard multi-reference attention trained on the same data?
Questions for authors

see weakness

5 comments
AuthorsOpenReview ↗

Response to Reviewer 2PoH (1/2): Anchor Acquisition and Geometric Alignment

We thank the reviewer for recognizing the value of CLACT and the effectiveness of GAMA. We provide additional implementation details and evaluation results below.

W1.1. Real-world anchor acquisition.

For the 25 real-world cases, we capture multi-view observations of the same character. One frame is randomly selected as the initial input, and complementary front, back, and side observations are selected as character anchors.

W1.2. Alignment for collinear trajectories.

Collinear camera centers leave the global rotation underconstrained when alignment relies on position correspondences alone. Our implementation checks the number of distinct target correspondences and their spatial spread. For degenerate configurations with an available subject-center estimate, we determine the alignment using the shared initial-frame orientation and subject-center geometry:

  1. Rotation. The initial image is shared by the anchor observations and the target trajectory. After reconciling camera-coordinate conventions, its camera orientations in the two coordinate systems determine the global rotation.
  2. Scale. We compute the centroid of the character-anchor camera centers and their mean distance to this centroid, excluding the shared initial frame. The global scale is the ratio of the target initial camera's distance to the subject center to this mean distance.
  3. Translation. We translate the rotated and scaled anchor-camera centroid to the target subject center.

The subject center is obtained during benchmark preprocessing through least-squares intersection of camera viewing rays. For trajectories with parallel or collinear viewing rays, preprocessing uses an available non-degenerate orbit trajectory from the same static scene to estimate this center.

W1.3. Multiple anchors mapped to the same target frame.

Anchors sharing a target-frame correspondence are retained as independent observations. Their correspondences, together with other valid matches, are used to estimate one global Sim(3) transform. Applying this transform to each anchor's original camera pose preserves the relative geometric structure of the anchor configuration.

When repeated matches leave too few distinct target correspondences or insufficient spatial spread, alignment uses the shared initial-frame orientation and the available subject center, following the procedure in W1.2.

The resulting transform is applied to all anchor cameras, including those excluded from the fitting subset. Each anchor is independently encoded and augmented with a geometric embedding computed from its aligned camera pose. All anchor tokens are then concatenated into character memory, allowing video queries to retrieve complementary appearance and geometric information from the individual observations.

AuthorsOpenReview ↗

Response to Reviewer 2PoH (2/2): Camera Accuracy, Held-Out Views, and Controlled Comparisons

W2.1. Fixed-scale camera errors on synthetic scenes.

Following the reviewer's suggestion, we evaluate fixed-scale camera errors on the synthetic scenes in our evaluation benchmark. All methods use the same VGGT-Omega estimator and a common temporal grid.

For each case, we determine a reference scale srefs_{\mathrm{ref}} from the ratio of rendered ground-truth character depth to the corresponding depth estimated by VGGT-Omega on the initial frame. This scale is shared across all methods and held fixed during trajectory alignment, where only rotation and translation are fitted. We additionally report Sim(3)-aligned RMSE, which allows global scale adjustment, and the mean scale ratio sA‾/sref‾\overline{s_A}/\overline{s_{\mathrm{ref}}} (ideal: 1). The GT-render reference follows the same estimation and evaluation protocol.

MethodFixed-scale RMSE (cm) ↓Sim(3) RMSE (cm) ↓Scale Ratio (ideal: 1)
GT-render reference12.253.140.996
CamI2V120.95112.121.307
DualCamCtrl128.94108.821.463
GEN3C47.8921.470.892
Uni3C35.728.430.998
Lyra 2.057.2835.701.140
ViewCrafter134.22128.971.644
ViewCrafter*139.36133.151.995
MVGenMaster52.0513.731.035
MVGenMaster*49.3030.031.200
ReCamMaster116.44112.621.190
DepthDirector114.1297.212.308
CameraAnything81.8272.031.125
WorldStereo59.6642.901.175
WorldCast18.546.091.069

WorldCast achieves the lowest position RMSE among generation methods under both fixed-scale (18.54 cm) and Sim(3) alignment (6.09 cm). These results confirm its camera-control advantage even when scale adjustment is not permitted during trajectory alignment.

W2.2. Character consistency against held-out views.

Following the reviewer's suggestion, we further evaluate character consistency against held-out views on synthetic scenes. We select four held-out GT views per case, each offset by approximately 45° from the conditioning anchor views. These held-out views are never used as inputs for video generation.

Using these held-out GT views as references, we evaluate character consistency with DINO, CLIP, ArcFace, and MEt3R.

MethodDINO ↑CLIP ↑ArcFace ↑MEt3R ↓
CamI2V0.7120.8680.0970.718
DualCamCtrl0.6370.8540.0700.671
GEN3C0.8060.9270.3020.577
Uni3C0.7960.9290.2620.593
Lyra 2.00.7720.9210.1950.616
ViewCrafter0.6360.8430.1120.759
ViewCrafter*0.5790.8200.0910.779
MVGenMaster0.6590.8590.1520.634
MVGenMaster*0.6910.8810.1730.581
ReCamMaster0.7530.9160.2350.621
DepthDirector0.7670.9020.2970.643
CameraAnything0.7310.9180.2240.608
WorldStereo0.7290.9080.1760.645
WorldCast0.9100.9620.4190.461

WorldCast achieves the highest DINO, CLIP, and ArcFace similarities against held-out GT views and the lowest MEt3R error. These results demonstrate that WorldCast not only preserves consistency with the input character observations but also maintains character consistency at previously unseen viewpoints.

W3. Controlled comparison with standard multi-reference attention.

Following the reviewer's suggestion, we compared GAMA with standard multi-view memory attention using the same Wan2.1-I2V-14B backbone, training data, four character anchors, and training settings, including random anchor dropout.

MethodATE ↓RTE ↓RRE (°) ↓ArcFace ↑DINO ↑CLIP ↑MEt3R ↓Subj. ↑IQ ↑
Standard Multi-View Memory Attention0.1040.0250.8860.3960.6940.8210.2910.9090.716
WorldCast (GAMA)0.0650.0200.8960.5550.7190.8300.2280.9170.719

Under identical conditions, GAMA improves ArcFace by 0.159 and reduces MEt3R and ATE by 21.6% and 37.5%, respectively, while maintaining comparable RRE. These results further demonstrate the effectiveness of GAMA's geometry-aware memory retrieval in improving character consistency and camera-control accuracy under identical input and training conditions.

We sincerely hope that our detailed clarifications, supplementary evaluations, and manuscript revisions have fully addressed your concerns. Thank you again for your thoughtful and constructive comments. We welcome any further questions, comments, or discussion.

AuthorsOpenReview ↗

Follow-up to Reviewer 2PoH

Thank you again for your constructive feedback and valuable suggestions. As discussion deadline approaches, we would like to briefly follow up on our previous responses and manuscript revisions, including the additional experiments and clarifications addressing your concerns. We hope our responses and revisions have addressed your concerns and would welcome any further comments or feedback.

Thank you again for your consideration.

Reviewer_5jTSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WorldCast, a character-centric video generation framework for large-angle camera control. It combines geometric guidance with Geometry-Aligned Memory Attention to preserve character consistency across viewpoints. The authors also introduce a dataset containing 80,000 synthetic training samples and 42 evaluation scenes. Experiments demonstrate improvements in camera-control accuracy and character consistency.

Critical strengths
  1. The paper addresses the important challenge of maintaining character consistency under large camera movements.

  2. Combining geometric guidance with multi-view character memory is a sensible approach.

  3. The method achieves competitive performance, supported by component ablations.

Critical weaknesses
  1. The novelty is not clearly described. The contribution summary should differentiate this work from the literature, rather than describing the methods again.

  2. WorldCast uses multiple reference views, whereas some baselines use only a single image. This makes it difficult to attribute the improvements to the proposed architecture rather than additional input information.

  3. The framework introduces 20 additional memory-attention blocks on top of a 14B-parameter model, but inference time, memory overhead, and computational complexity are not reported.

Questions for authors
  1. What distinguishes GAMA from existing memory attention mechanisms?

  2. Can the authors provide controlled comparisons using identical reference views and training data?

  3. What are the additional computational costs and limitations of the proposed framework?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 5jTS (1/3): Methodological Distinctions and Multi-View Comparisons

We thank the reviewer for recognizing the importance of character consistency under large camera movements and the effectiveness of our approach. We provide additional analyses, controlled comparisons, and efficiency measurements.

W1. Clarity of the contribution summary.

We have revised the contribution summary to distinguish the contributions of CLACT, WorldCast, and GAMA. CLACT provides a character-centric dataset and benchmark combining large-angle camera trajectories with complementary multi-view character observations. WorldCast integrates trajectory-aligned geometric guidance with multi-view character memory to jointly address camera control and character consistency. GAMA establishes explicit geometric relationships between character anchors and target viewpoints, enabling viewpoint-aware retrieval of complementary character information under large viewpoint changes.

Q1 / Q2. Comparison with standard memory attention under controlled settings.

GAMA establishes an explicit geometric relationship between the target viewpoint and the character observations stored in memory. Standard multi-view memory attention retrieves information from reference-view features. GAMA additionally aligns the character anchors with the target camera trajectory and incorporates camera geometry into retrieval, enabling the model to retrieve character features relevant to the requested viewpoint. As large camera rotations reveal different character surfaces, complementary observations provide details that are insufficiently observed in the initial image, supporting identity and appearance consistency across viewpoints.

To evaluate this design, we conducted a controlled comparison with standard multi-view memory attention following the reviewer's suggestion. Both methods used the same Wan2.1-I2V-14B backbone, training data, four character anchors, and training settings, including random anchor dropout. The baseline employed conventional multi-view memory attention.

MethodATE ↓RTE ↓RRE (°) ↓ArcFace ↑DINO ↑CLIP ↑MEt3R ↓Subj. ↑IQ ↑
Standard Multi-View Memory Attention0.1040.0250.8860.3960.6940.8210.2910.9090.716
WorldCast (GAMA)0.0650.0200.8960.5550.7190.8300.2280.9170.719

Under identical input and training conditions, GAMA improves ArcFace by 0.159 and reduces MEt3R and ATE by 21.6% and 37.5%, respectively, while maintaining comparable RRE (0.896° vs. 0.886°). Subject Consistency and Image Quality also improve. These results support the contribution of GAMA's geometry-aware memory design to character consistency and translational camera-control accuracy when both methods receive the same multi-view observations.

W2. Fairness of comparisons with multi-view inputs.

Although some baselines support only a single input image, the original submission already evaluates anchor-conditioned ViewCrafter and MVGenMaster variants, which also receive complementary character observations. As discussed in the paper, these methods still face challenges in maintaining character consistency under large-angle camera motion, even with additional reference views. To further compare with methods supporting multi-view anchor inputs, we additionally evaluate GEN3C with additional character anchors and GEN3R with both its native input and additional character anchors.

MethodATE ↓RTE ↓RRE (°) ↓ArcFace ↑DINO ↑CLIP ↑MEt3R ↓Subj. ↑IQ ↑
GEN3C w/ Anchors0.1820.05841.660.3450.5990.7640.3800.8180.679
GEN3R0.2990.1324.060.2770.6410.8080.4030.8550.678
GEN3R w/ Anchors0.3110.1355.210.3740.7330.8440.3260.8510.677
WorldCast (Ours)0.06500.02000.8960.5550.7190.8300.2280.9170.719

Adding character anchors improves GEN3R's ArcFace from 0.277 to 0.374 and reduces MEt3R from 0.403 to 0.326, demonstrating the benefits of complementary character observations. However, its camera-control errors remain substantial. Although GEN3R w/ Anchors achieves higher DINO and CLIP scores, WorldCast performs better in camera-control accuracy, character identity preservation, geometric consistency, subject consistency, and image quality.

Overall, compared with existing state-of-the-art methods supporting multi-anchor inputs, WorldCast achieves stronger overall performance in camera-control accuracy, character identity preservation, geometric consistency, and generation quality.

AuthorsOpenReview ↗

Response to Reviewer 5jTS (2/3): Computational Costs

Q3. What are the additional computational costs and limitations?

Thank you for requesting these measurements. We report the inference time, parameter count, computational cost, and peak GPU memory on a single M403 GPU with BF16 precision. Runtime is measured end-to-end. Parameter counts cover the denoising backbone and control adapters.

Q3.1. Computational overhead relative to the backbone.

Both models generate 81 frames at 480×832 resolution, using 50 sampling steps and CFG 5.

ModelParams (B)FLOPs/video (P)Time (s)Peak Memory (GB)
Wan2.1-I2V-14B16.39168.6425.052.0
WorldCast17.39205.7520.056.1
Relative Overhead+6.1%+22.0%+22.4%+7.9%

The 20 memory-attention blocks introduce approximately 0.99B additional parameters. Compared with the Wan2.1-I2V-14B backbone, WorldCast increases the per-video computation by 22.0%, inference time by 22.4%, and peak memory by 7.9%.

Q3.2. Efficiency comparison with other methods.

We further compare WorldCast with existing methods in terms of parameter count, computational cost, inference time, and peak GPU memory.

MethodResolution (H×W)FramesStepsParams (B)FLOPs/video (P)Time (s) ↓Peak Memory (GB) ↓
CamI2V448×83216251.501.395.314.1
DualCamCtrl480×83261503.2518.1178.221.2
GEN3C480×832121357.2421.7156.958.0
GEN3R448×78481501.6223.0304.952.6
Uni3C480×832814017.39147.1442.471.3
Lyra 2.0480×832815017.06168.6666.888.1
ViewCrafter448×83225501.434.252.120.5
MVGenMaster480×83228500.873.767.836.6
ReCamMaster480×83281501.4928.3317.322.3
DepthDirector480×83281505.0034.4170.247.7
CameraAnything480×832815018.54167.81512.269.9
WorldStereo480×832814017.68101.71954.382.1
WorldCast480×832815017.39205.7520.056.1

Some methods cannot be evaluated under fully aligned settings, resulting in differences in output frame counts, resolutions, sampling steps, and inference configurations. Among methods generating 81 frames at 480×832 resolution, WorldCast achieves a shorter inference time than Lyra 2.0, CameraAnything, and WorldStereo. Its peak memory consumption is also lower than that of several large-scale counterparts, including Uni3C, Lyra 2.0, CameraAnything, and WorldStereo.

AuthorsOpenReview ↗

Response to Reviewer 5jTS (3/3): Acceleration and Limitations

Q3.3. Compatibility with four-step acceleration.

Notably, WorldCast can be directly adapted to pretrained few-step acceleration methods without additional training or fine-tuning. Specifically, we apply a pretrained LightX2V acceleration LoRA based on Improved Distribution Matching Distillation (DMD2) [1] to WorldCast. With 4 sampling steps and CFG 1, inference time decreases from 520.0 s to 43.0 s, achieving 12.1× acceleration, while peak memory increases only slightly, from 56.1 GB to 57.8 GB.

MethodResolutionFramesStepsParams (B)FLOPs/video (P)Time (s) ↓Peak Memory (GB) ↓
WorldCast480×832815017.39205.7520.056.1
WorldCast + LoRA480×83281417.769.943.057.8

We also evaluate generation quality before and after acceleration to examine whether the substantial reduction in sampling steps compromises camera control or character consistency.

SettingATE ↓RTE ↓RRE (°) ↓ArcFace ↑DINO ↑CLIP ↑MEt3R ↓Subj. ↑IQ ↑
WorldCast (50 steps)0.0650.0200.8960.5550.7190.8300.2280.9170.719
WorldCast + LoRA (4 steps)0.0450.0140.5540.5590.7420.8390.2250.9230.723

The accelerated model maintains or improves all reported metrics, including camera trajectory accuracy, character identity consistency, geometric consistency, subject consistency, and image quality. These results demonstrate that WorldCast is compatible with pretrained few-step acceleration LoRAs, achieving substantially faster inference without compromising generation quality on our benchmark.

Limitations and Efficiency.

The original 50-step configuration requires 520 s per video, indicating room for further acceleration. By directly incorporating a pretrained LightX2V acceleration LoRA, we reduce the inference time to 43 s, achieving a 12.1× speedup, with only a modest increase in peak memory from 56.1 GB to 57.8 GB.

Although the accelerated configuration still requires substantial GPU memory, its peak memory remains lower than that of several large-scale counterparts, including Uni3C (71.3 GB), Lyra 2.0 (88.1 GB), CameraAnything (69.9 GB), and WorldStereo (82.1 GB). Overall, these results demonstrate that WorldCast achieves a favorable balance between generation quality and inference efficiency, particularly when combined with pretrained few-step acceleration methods.

Reference

[1] Yin, T., Gharbi, M., Park, T., Zhang, R., Shechtman, E., Durand, F., & Freeman, W. T. (2024). Improved Distribution Matching Distillation for Fast Image Synthesis. Advances in Neural Information Processing Systems, 37, 47455–47487.

We hope that our comprehensive responses and the additional analyses and revisions have satisfactorily addressed your questions and concerns. We sincerely appreciate your valuable suggestions and welcome any further feedback or discussion.

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

This paper presents WorldCast, a framework for character-centric video generation under large-angle camera motion. The paper addresses two coupled challenges: accurately following complex camera trajectories involving large rotations and translations, and maintaining character identity, appearance, and geometric consistency when previously unseen character regions become visible.

The proposed framework combines scale-aligned geometric camera conditioning with Geometry-Aligned Memory Attention (GAMA). The former constructs trajectory-aligned guidance by back-projecting the initial frame into a 3D point cloud and rendering it along the target camera trajectory, supplemented by Plücker ray encoding. The latter aligns additional multi-view character observations with the target camera geometry in a shared coordinate frame and incorporates view-relevant character features through memory attention.

The authors also introduce CLACT, a character-centric dataset and benchmark comprising 80K synthetic training samples and 42 synthetic and real-world evaluation scenes. The framework is built upon Wan2.1-I2V-14B and trained in two stages. Experiments demonstrate improved camera-control accuracy and character consistency over the evaluated baselines, with ablation studies supporting the contributions of geometric conditioning and multi-view memory.

Critical strengths

1. Meaningful problem formulation and an effective integration of geometric and appearance conditioning.

The paper addresses an important challenge in camera-controlled video generation: large viewpoint changes simultaneously increase the difficulty of trajectory following and expose previously unseen character regions. The proposed combination of trajectory-aligned geometric guidance and multi-view character memory is well motivated. In particular, aligning the reference observations and target camera trajectory in a shared geometric coordinate frame provides a reasonable mechanism for retrieving view-relevant character information.

2. A useful benchmark and encouraging experimental results.

The proposed CLACT benchmark covers diverse large-angle camera trajectories and evaluates camera accuracy, cross-view geometric consistency, identity preservation, and visual quality. WorldCast achieves strong performance across these dimensions, including an ATE of 0.065, MEt3R of 0.228, and ArcFace similarity of 0.555. The component ablations and attention-routing analysis provide additional evidence that the proposed conditioning and memory mechanisms contribute to the overall performance. These results demonstrate the practical value of the framework on the evaluated settings.

Critical weaknesses

1. Insufficient isolation of the methodological contribution of geometry-aligned memory.

While GAMA is presented as a central methodological contribution, it remains unclear how much of the improvement in character consistency comes specifically from geometry-aware alignment rather than the additional multi-view observations and memory capacity.

For example, Table 2 shows that removing GAMA substantially decreases ArcFace similarity from 0.555 to 0.370 and increases MEt3R from 0.228 to 0.333. However, removing only anchor geometry alignment results in relatively small changes in ArcFace (0.555 to 0.547) and MEt3R (0.228 to 0.239), although the rotation error increases substantially. These results suggest that multi-view memory itself may account for a large portion of the identity and appearance consistency gains, while the specific contribution of geometry alignment to these metrics is less pronounced.

Furthermore, the proposed mechanism combines established components, including multi-view pose estimation, geometric alignment, camera-ray encoding, and attention-based memory retrieval. The paper discusses WorldStereo, which also explores geometry-constrained memory retrieval, but does not include it in the experimental comparisons. Although the two methods have different input assumptions and objectives, a more detailed methodological comparison would help establish the novelty of GAMA.

A controlled comparison against standard multi-view attention or memory retrieval, using the same backbone, training data, and reference inputs, would be important for attributing the observed improvements to the proposed geometric design.

2. Limited evidence for practical generalization and insufficient clarification of multi-view reference acquisition.

The proposed framework relies on additional front, back, and side character anchors. In the synthetic training pipeline, these anchors are directly rendered from existing 3D character assets using Blender. However, the acquisition procedure for multi-view anchors in the real-world evaluation cases is not sufficiently clarified. This is important because such complementary observations may not be readily available in practical image-to-video generation scenarios.

Moreover, the anchor images contain not only intrinsic character appearance but also scene-dependent illumination and shading. Although foreground masking removes background information, the proposed memory mechanism does not explicitly disentangle character identity and material appearance from lighting conditions. It is therefore unclear whether the same character anchors can be reused across different scenes and illumination conditions, or whether scene-specific references are necessary to achieve the reported performance.

The evaluation also raises questions about how character consistency is measured. According to Appendix C.3, MEt3R is computed using generated frames together with the four observation anchors, while ArcFace, CLIP, and DINO similarities are measured against the closest-view anchor. These metrics provide useful evidence of reference consistency, but do not fully establish long-range geometric consistency within the generated video itself or the ability to adapt character appearance to different scene illumination.

A more controlled evaluation of cross-scene anchor reuse and long-range consistency under 360-degree camera trajectories would strengthen the claims of world-consistent character generation.

Questions for authors

Q1. Contribution of geometry-aligned memory.

Could the authors compare GAMA with a standard multi-view cross-attention or memory-attention baseline using identical character anchors, backbone, and training data? Given the relatively small improvements in ArcFace and MEt3R from anchor geometry alignment in Table 2, this would help clarify the specific contribution of geometry-aware retrieval. Additionally, could the authors further distinguish GAMA from existing geometry-aware memory mechanisms such as WorldStereo and explain whether a direct comparison is feasible?

Q2. Real-world anchor acquisition and cross-scene generalization.

How are the front, back, and side anchors obtained for the 25 real-world evaluation scenes? Are these references captured or generated specifically for each scene, and do they share similar lighting conditions with the target video?

Could the authors evaluate whether a fixed set of character anchors can be reused across different environments and illumination conditions? This would help determine whether GAMA learns a reusable character representation or relies heavily on scene-specific appearance information.

Q3. Evaluation of long-range world consistency.

The current evaluation uses MEt3R together with anchor-based ArcFace, CLIP, and DINO similarities. Could the authors clarify the pairing and aggregation protocol used for MEt3R, particularly whether the reported metric evaluates consistency among generated frames independently of the reference anchors?

For large-angle camera trajectories, especially 360-degree orbits, could the authors provide a long-range consistency evaluation comparing generated frames at repeated viewpoints? An additional video-centric metric, such as the Multi-View Consistency metric in VBench-2.0, could also provide complementary evidence, although it is not necessarily a substitute for MEt3R.

8 comments
AuthorsOpenReview ↗

Response to Reviewer weSL (1/3): Comparison with Standard Multi-View Attention and WorldStereo

We thank the reviewer for the constructive feedback and recognition of our work. We have conducted additional experiments and provided further clarifications to address the reviewer's comments, as detailed below.

Q1.1. Controlled comparison with standard multi-view attention.

Following the reviewer's suggestion, we introduce a standard multi-view memory-attention baseline for a controlled comparison with GAMA. Both methods use the same Wan2.1-I2V-14B backbone, training data, four character anchors, and training settings, including random anchor dropout. The baseline employs conventional multi-view memory attention.

MethodATE ↓RTE ↓RRE (°) ↓ArcFace ↑DINO ↑CLIP ↑MEt3R ↓Subj. ↑IQ ↑
Standard Multi-View Memory Attention0.1040.0250.8860.3960.6940.8210.2910.9090.716
WorldCast (GAMA)0.0650.0200.8960.5550.7190.8300.2280.9170.719

Under identical training and input conditions, GAMA improves ArcFace by 0.159 and reduces MEt3R and ATE by 21.6% and 37.5%, respectively, while maintaining comparable RRE (0.896° vs. 0.886°). These results support the contribution of GAMA's geometry-aware memory design beyond access to the same multi-view anchors, with improvements in both translational camera accuracy and character consistency. We will include this controlled comparison in the revised manuscript.

Q1.2. Comparison with WorldStereo.

WorldStereo and GAMA both use geometric information to improve video consistency. Their memory mechanisms differ in the information stored and retrieved. WorldStereo maintains an incrementally updated scene-level 3D point cloud through Global-Geometric Memory (GGM). Its Spatial-Stereo Memory (SSM) retrieves historical views based on geometric overlap and establishes correspondences between target and reference views. GAMA uses complementary character observations as memory, aligns them with the target camera geometry in a shared coordinate frame, and retrieves character features relevant to the target viewpoint.

Both methods support camera-controlled video generation. Following the reviewer's suggestion, we evaluate WorldStereo using its original inference configuration, including resolution, frame settings, and inference procedure. Both methods are evaluated on the same benchmark using a unified evaluation protocol and metrics.

MethodATE ↓RTE ↓RRE (°) ↓ArcFace ↑DINO ↑CLIP ↑MEt3R ↓Subj. ↑IQ ↑
WorldStereo0.1330.0321.0110.3500.6650.7990.3320.8630.684
WorldCast (Ours)0.0650.0200.8960.5550.7190.8300.2280.9170.719

On this benchmark, WorldCast reduces ATE and MEt3R by 51.1% and 31.3%, respectively, and improves ArcFace by 0.205 compared with WorldStereo. These results demonstrate improved camera trajectory accuracy and character consistency under large-angle camera motion.

AuthorsOpenReview ↗

Response to Reviewer weSL (2/3): Cross-Scene Generalization

Q2.1. Real-world anchor acquisition.

For the 25 real-world evaluation cases, character anchors are obtained through real-world multi-view capture. Specifically, one frame is randomly selected from each multi-view capture as the initial input, while complementary observations from the same capture serve as the front, back, and side anchors. Therefore, in the original evaluation, the anchors for each case are captured in the corresponding scene and generally share similar environmental and illumination conditions with the target video.

Q2.2. Generalization across environments and illumination.

To investigate whether fixed character anchors can be reused across different environments and illumination conditions, we conduct an additional evaluation on 10 cases, each with three distinct scene/lighting variations. For each case, we keep its character anchors unchanged while modifying the initial frame's environment, lighting intensity, and color temperature, covering substantially different warm, cool, and darker illumination conditions.

We provide qualitative comparisons and an additional set of generated videos in the supplementary material for visual inspection. Despite pronounced scene and illumination differences between the modified initial frames and the fixed anchors, the generated videos generally follow the input scene and illumination style while preserving character identity, clothing textures, and appearance details from the anchors. We further report quantitative comparisons before and after these scene and lighting changes:

SettingArcFace↑DINO↑CLIP↑MEt3R↓
Original0.5180.8390.8880.315
Scene/Lighting Changes0.4980.7930.8630.353

After the scene and lighting changes, ArcFace decreases by only 0.020, while DINO and CLIP decrease by 0.046 and 0.025, respectively, and MEt3R increases by 0.038. Importantly, these metrics are all anchor-referenced consistency measures. When the initial frame has substantially different illumination from the anchors, the generated video needs to follow the input illumination, which may introduce appearance differences relative to the unchanged anchors and consequently lower the reference similarity scores.

Overall, the generated results maintain character identity and appearance while adapting to the input scene and illumination, demonstrating the reusability of fixed character anchors across different environments and lighting conditions and the cross-scene generalization capability of GAMA.

AuthorsOpenReview ↗

Response to Reviewer weSL (3/3): Long-Range World Consistency

Q3.1. Original MEt3R evaluation protocol.

The original MEt3R evaluates person-only consistency between generated frames and character anchors. Specifically, we apply SAM3 for person segmentation and place the foreground on a black background. For each sampled frame, we select the single nearest-view anchor according to the ground-truth camera azimuth. MASt3R is used to reconstruct dense 3D point maps, while DINO/FeatUp features are back-projected into 3D. We then calculate the cosine distance over overlapping foreground regions. Scores are averaged across sampled frames within each video and subsequently across evaluation cases. We will clarify this pairing and aggregation protocol in Appendix C.3.

Q3.2. Long-range consistency at repeated viewpoints.

Following the reviewer's suggestion, we introduce Repeated-View MEt3R to evaluate long-range consistency under large-angle camera trajectories. Specifically, we identify pairs of generated frames with overlapping camera viewpoints within each video, calculate pairwise MEt3R between these frames, and average the scores across all valid pairs.

MethodCamI2VDualCamCtrlGEN3CUni3CLyra 2.0ViewCrafterViewCrafter*
Repeated-View MEt3R0.0890.4530.0680.2190.2120.3260.468
MethodMVGenMasterMVGenMaster*ReCamMasterDepthDirectorCameraAnythingWorldStereoWorldCast (Ours)
Repeated-View MEt3R0.0700.1070.2560.3460.1490.3780.064

WorldCast achieves the lowest Repeated-View MEt3R (0.064), demonstrating stronger long-range consistency across repeated viewpoints.

Q3.3. Additional VBench-2.0 evaluation.

We further evaluate the VBench-2.0 Multi-View Consistency (MVC) metric suggested by the reviewer. MVC uses CoTracker2 to track points near the upper and lower image boundaries and identifies orbiting motion based on their opposite vertical movements, followed by RAFT optical flow and SIFT/RANSAC geometric matching. However, this orbit-detection criterion relies on scene-level parallax and is not well suited to character-centric camera motion, where image-boundary points may not exhibit the required motion pattern. Consequently, only 2.0% of videos in our large-angle camera-motion benchmark pass the validity filter, leaving insufficient valid samples for reliable comparisons across methods. Thus, MVC is unsuitable for quantitative method comparisons on our benchmark.

Other VBench-2.0 dimensions also have limitations for our task. Human Identity evaluates facial identity consistency relative to the first frame, while Human Clothes uses LLaVA-based yes/no question answering to assess clothing consistency. Neither directly measures full-body structural consistency across substantially changing viewpoints. Mechanics, Thermotics, and Material primarily assess physical phenomena and material properties, which are less relevant to character anatomy and structural consistency under large-angle camera motion.

We therefore additionally evaluate the more relevant VBench-2.0 Human Anatomy, which assesses anatomical abnormalities in the human body, face, and hands, providing a complementary measure of character structural plausibility. We conduct a full evaluation across all six camera trajectories and 42 cases.

MethodCamI2VDualCamCtrlGEN3CUni3CLyra 2.0ViewCrafterViewCrafter*
Human Anatomy0.8300.9620.9570.9710.9770.8640.855
MethodMVGenMasterMVGenMaster*ReCamMasterDepthDirectorCameraAnythingWorldStereoWorldCast (Ours)
Human Anatomy0.9430.9570.9570.8880.9600.9610.976

The additional Human Anatomy evaluation further supports the anatomical plausibility of characters generated by WorldCast, with a score of 0.976.

We sincerely hope that our detailed responses, additional experiments, and manuscript revisions have adequately addressed your concerns. Thank you again for your constructive feedback. We welcome any further questions, comments, or discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions in Response to Reviewer Comments

We thank all reviewers for their constructive feedback. The revised manuscript incorporates the following experiments and clarifications.

Reviewer weSL

  • Added comparisons with standard multi-view memory attention and WorldStereo (Secs. 4.2 and 4.3).
  • Evaluated cross-scene and illumination generalization using fixed character anchors (App. D.2).
  • Clarified the MEt3R evaluation protocol and added repeated-view consistency and VBench-2.0 Human Anatomy evaluations (Apps. C.3 and D.3).

Reviewer 5jTS

  • Refined the contribution summary and added multi-view input baseline comparisons (Sec. 1 and App. D.4).
  • Added computational cost analysis and inference efficiency comparisons with existing methods (App. D.5).
  • Evaluated four-step LoRA acceleration and its impact on generation quality (App. D.5).

Reviewer 2PoH

  • Clarified real-world anchor acquisition and geometric alignment under degenerate configurations (Apps. D.2 and E).
  • Added fixed-scale camera accuracy evaluation on synthetic scenes (App. D.7).
  • Added character consistency evaluation against held-out ground-truth views (App. D.6).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.