acceptodds
Under review as a conference paper at ICLR 2027

PAWA: Panoramic Action-conditioned World Modeling for Multiple Agents

Abstract

A multiplayer world model must predict how the independent actions of several players change a shared environment, including changes outside a player's current field of view. A point-of-view (POV) video captures such changes only after they enter its field of view, whereas panoramic video records the surroundings in every viewing direction and has served as an intermediate representation of the world state in recent world models. Learning this task therefore requires synchronized POV and panoramic videos of interacting players together with their actions, yet no existing dataset provides this combination in Minecraft, an established testbed in which players act independently in a shared, modifiable 3D world. We present PAWA, a dataset and framework for action-conditioned panoramic multiplayer world modeling. The dataset pairs each player's POV video with world-stabilized panoramic video and aligned actions. Its distributed pipeline decouples CPU gameplay capture from GPU rendering, scales beyond 64 simultaneous players, and re-renders each recorded session under new observation configurations without replaying the gameplay. On this dataset, we develop a dual-tower model that jointly predicts POV and panoramic video and couples the two streams through cross-attention that aligns tokens observing the same direction in a shared world frame. We further introduce a quadratic-phase player encoding that provides many more player codes than existing rotary codes at the same embedding dimension. Experiments show that PAWA outperforms state-of-the-art multiplayer world models in POV prediction quality while also producing high-fidelity panoramic predictions. We will release the dataset and collection pipeline to support further research on multiplayer world models.

Primary area:
generative models
Keywords:
world models, multiplayer world models, video generation, panoramic video, diffusion models, dataset

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

This paper proposes PAWA, a multiplayer world model that jointly predicts POV and panoramic videos in Minecraft. It introduces a scalable data collection pipeline, a dual-tower architecture with ray-aligned cross-attention, and a quadratic-phase player encoding for handling more players. Experiments show improvements over existing multiplayer world models.

Critical strengths
  • The paper introduces an approach to multiplayer world modeling by combining POV and panoramic video prediction. The motivation and technical details are clearly explained, and the paper is well written and easy to follow.
  • The paper includes extensive experiments and ablation studies. It also provides useful visualizations and qualitative results to illustrate the improvements.
Critical weaknesses
  • The evaluation mainly relies on visual quality metrics such as FVD and LPIPS, while largely overlooking consistency between the generated POV and panoramic videos. For this type of world modeling task, cross-view consistency is more important than the visual quality of each individual stream.
  • The paper does not compare ray-aligned cross-attention with simpler alternatives, such as vanilla cross-attention or explicit geometric warping. Since POV and panoramic views share the same camera center, their geometric correspondence can be directly computed. The necessity of the proposed ray-aligned design is therefore not sufficiently demonstrated.
  • The experiments use 12s rollouts and at most 16 players. I'm interesting about model's long-term consistency and scalability. In addition, the paper lacks an analysis of computational cost of the two DiT towers.
  • The comparison with existing methods is not very fair. PAWA uses additional panoramic inputs and a dual-tower architecture, while the baselines only use POV inputs.
Questions for authors

The experiments mainly focus on co-player tracking and action following, with limited evaluation of changes to the shared environment. I'm curious about another important aspect: whether changes made by one player, such as placing or breaking blocks, are correctly reflected in other players' views. This aspect is not evaluated.

3 comments

We thank all reviewers for their thoughtful reviews. We are glad that the reviewers acknowledge our contributions, including the clear motivation of modeling out-of-view interactions, the usefulness of the dataset and collection pipeline, the well-motivated geometric design, and the extensive experiments. Below we respond to your comments point by point.

W1: Cross-view consistency. We agree that a direct measure of agreement between the two generated streams is valuable, and we report one here. (The paper evaluates each stream against its own ground truth, in the POV and panorama rows of Table 1.) Since a player's POV camera and panorama share the same center, we reproject each generated panorama into the POV frustum using the camera orientation of that frame and compare it with the POV generated by the same model (LPIPS ↓, excluding the HUD rows). Because the two streams differ in resolution and the panorama contains no hand or HUD, the metric does not reach zero even for the recorded videos. We therefore report two reference levels measured on the recorded videos: the attainable level (recorded POV vs. recorded panorama) and the level of unrelated views (each POV frame paired with the panorama of the same video 6 s away). Results on the 30 two-player interaction episodes (frames 1–239):

SourceMatched (same moment)Time-shifted control (6 s apart)Matched − control [95% CI]Episodes with matched < control
Recorded videos0.4440.618−0.174 [−0.192, −0.157]30/30
PAWA (Causal)0.5660.608−0.042 [−0.053, −0.032]30/30
PAWA (Distilled)0.6190.638−0.019 [−0.023, −0.014]29/30

Both PAWA variants are significantly more consistent than their time-shifted controls, on 30 of 30 (Causal) and 29 of 30 (Distilled) episodes, and PAWA (Causal) remains so in all four time bands of the 12 s rollout (frames 189–239: −0.033 [−0.042, −0.026]). A generated POV thus agrees with its own panorama at the same moment significantly better than with an unrelated view of the same scene: the two towers generate the same world rather than two independently plausible ones. Prior models do not predict a panoramic stream, so the metric is undefined for them. We will add this metric to the paper.

W2: Ray-aligned vs. simpler cross-attention. We agree that Tables 9–11 primarily show the value of exchanging information, and that a vanilla cross-attention trained under the same conditions is the most direct test. Each such control requires retraining the model at full scale, which does not fit within the rebuttal period with the compute currently available to us. We will therefore add this comparison to the final paper: a vanilla cross-attention with image-coordinate 2D RoPE and an explicit-warping baseline, both trained under the same conditions as the full model. We add two points here:

  • What the ray alignment adds. The correspondence between the two image grids changes with yaw and pitch at every frame. A plain cross-attention would have to learn this mapping from a per-frame global orientation signal, whereas the ray phase supplies it in closed form and without parameters, so the attention starts from the correct geometric prior. The trained attention follows this geometry: at sites 19 and 23, POV queries place 43–57% of their attention inside their own frustum (10% by chance), and the most attended key lies 5°–9° from the query ray (Table 12).

  • Explicit warping. Warping imposes a fixed one-to-one pixel transfer, but the two streams differ in resolution, projection and content. The panorama covers 360°×180° at 512×256 pixels (about 1.4 px/°), whereas the POV covers 102°×70° at 480×320 (about 4.7 px/°), so a warp can supply at most about a third of the POV's linear resolution and nothing for the HUD and hand. Moreover, both streams are noisy at every denoising step, so copying pixels would carry the panorama's errors directly into the POV. The ray-aligned attention keeps the geometric prior while letting the model learn how much to read from the other stream through zero-initialized gates.

(Continued from 1/2.)

W3: Long rollouts, scalability and cost. The paper evaluates 12 s rollouts, 2.6× the causal training window (93 frames) and 5.3× the distillation window (45 frames). In Table 13's last band (frames 189–239), PAWA alone remains closer to the reference than a static first frame, both over the whole POV and within co-player regions.

We extended PAWA to continuous 24 s rollouts on 30 two-player episodes, with no mid-rollout reset. For PAWA (Distilled), POV PSNR decreases from 15.2512 to 13.5818 and LPIPS increases from 0.4717 to 0.5720 between the first and second 12 s windows of the same rollout. PAWA (Distilled) thus doubles the paper's rollout horizon without retraining, with moderate degradation in POV PSNR and LPIPS.

For scalability, PAWA is the only model that supports 16 players. There, PAWA (Causal) reaches a POV FVD of 92.2 (Table 2), and for 73–83% of the players their own actions best explain the camera motion of their generated video, against 6% by chance (Table 8). For cost, PAWA has 4.60B parameters: 2.18B in each tower and 0.24B (5%) in the seven cross-attention sites; Gamma-World has 2.18B. The table reports the average time per generated frame (sampling and decoding) and the peak memory on a single NVIDIA B200 GPU, over a 12 s rollout (6 s for 16 players); one frame comprises the POV and panorama of every player:

ModelPlayersStepsTime per framePeak memory
Gamma-World (Causal)2350.37 s41 GiB
Gamma-World (Distilled)240.07 s31 GiB
PAWA (Causal)2350.86 s86 GiB
PAWA (Distilled)240.23 s58 GiB
PAWA (Distilled), cross-attention removed240.22 s57.5 GiB
PAWA (Distilled)440.39 s91 GiB
PAWA (Distilled)840.73 s157 GiB
PAWA (Distilled)1641.28 s165 GiB

The cross-attention adds about 5% time and 1% memory; most of the additional cost over Gamma-World comes from the second tower, which predicts the panoramic stream. From 2 to 16 players, the time per frame grows roughly linearly with the number of players (0.23 s to 1.28 s).

W4: Fairness of the comparison. Predicting the panorama jointly with the POV is the central proposal of the paper rather than an auxiliary input: our claim is that, in multiplayer scenes, the panoramic stream gives each player's prediction access to what happens outside the current view. A comparison that removes all panoramic information would therefore evaluate a different model from the one we propose. We nevertheless ran such controls, to show what each difference contributes. Dual-tower architecture. Without cross-attention (Tables 9 and 11), the POV stream has no access to panoramic information, since pano→POV cross-attention is the only path from the panorama to the POV stream. The POV stream is then computed by a single tower initialized from Gamma-World and receives the same visual input as the baselines. Under the same training data and procedure, the arm of Table 11 that keeps the cross-attention reaches POV FVD 138.4 and LPIPS 0.455, against 191.7 and 0.511 for the arm without it. Additional panoramic input. We re-ran PAWA (Distilled) on the same 30 two-player episodes with the initial panorama replaced by a projection of the initial POV alone (regions not covered by the POV are filled with uniform gray), so that the model receives the same initial observation as the baselines. It still outperforms all three Gamma-World variants on every POV metric: PSNR 15.35, LPIPS 0.482, FID 30.2 and FVD 129.4, against 13.08, 0.567, 52.0 and 267.1 for the best Gamma-World variant on each metric. With the full initial panorama, the same re-run gives FVD 124.2 (138.1 in Table 1, which used a separate evaluation setup), so the advantage over the baselines does not depend on the additional initial panoramic observation.

Q: Block edits made by other players. We ran a preliminary analysis using the block placements logged during data collection. In the interaction tier, 7 placements occur within the forecast window and later become visible to the other player. Scoring the placed cell in that player's view against the recorded video, PAWA (Distilled) is closer to the recorded edit than a static copy of the conditioning frame (in which the block is absent) for 6 of the 7 events, in both the POV and the panorama; its POV LPIPS on the edited cell is 0.662, against 0.691 and 0.722 for Gamma-World (Bidirectional) and (Distilled). Our evaluation tiers were selected for co-visibility and camera motion, not for edits, so this sample is small. The final version will include a dedicated edit tier, covering both placing and breaking, in which one player edits the world outside the other's view; our pipeline supports this directly, since it records edits in the event log and renders any viewpoint offline.

Thank you for the rebuttal. The new results address most of my concerns about cost, long rollouts and fairness.

However, I am still not convinced by the cross-view consistency evaluation: the improvement over the time-shifted control is limited and does not directly show a consistent shared world across players. Also, the advantage of ray-aligned cross-attention over simpler alternatives remains unclear without direct comparisons.

Overall, I appreciate the authors' effort, but my concerns remain. I will raise my score to 3, but my actual assessment is more like 2.5.

Reviewer_AeRgOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PAWA, a framework for multiplayer world modeling using both POV and panoramic videos.

The authors first build a Minecraft dataset containing synchronized POV videos, panoramic videos, and actions from multiple players. They also introduce a data collection pipeline that separates gameplay recording from video rendering, allowing the collection of data with more than 64 players.

For the model, the authors propose a dual-tower architecture, where one tower generates POV videos and the other generates panoramic videos. The two towers exchange information through ray-aligned cross-attention. The paper also introduces a quadratic-phase player encoding to support more player identities.

Experiments show that PAWA performs better than existing multiplayer world models on several video generation metrics, including FVD, FID, and LPIPS.

Critical strengths

Paper's motivation is reasonable. A single player's POV cannot observe everything happening in a multiplayer world. Using panoramic videos to model the surrounding environment is an interesting idea, especially when other players can change the world outside the current view.

The dataset and collection pipeline are useful contributions. Separating gameplay recording from GPU rendering also seems practical, since the same recorded interactions can be rendered again with different camera settings.

The dual-tower architecture is also well motivated, and the ray-aligned cross-attention provides a reasonable way to connect POV and panoramic observations.

The experimental results are quite strong. PAWA outperforms the baselines by a large margin on most reported metrics. The paper also includes several useful ablation studies and qualitative comparisons.

Critical weaknesses
  1. I am not fully convinced that the comparison with existing models is fair.

In Section 5.1, PAWA receives both the initial POV and panoramic observations, while the baselines only receive the POV. This gives PAWA more information about the initial environment. I understand that this is part of the proposed framework. However, it makes it difficult to tell how much of the improvement comes from the proposed architecture and how much simply comes from having more visual information.

The authors mention that PAWA still outperforms the baselines when cross-attention is disabled (Table 9). However, this does not fully address the issue, since the model still uses the new dataset, its own training procedure, and panoramic conditioning in the panorama tower. That is, It is difficult to separate the effect of the proposed method from the new training data.

In fact, Table 11 shows that the model trained without cross-attention still reaches FVD 191.7 and LPIPS 0.511, which is already much better than the best baseline (267.1 / 0.567). This suggests that a large part of the gap comes from the training data rather than the method.

  1. The contribution of ray-aligned cross-attention is not fully isolated.

Most ablations compare the full model with a model where cross-attention is completely removed (Tables 9, 10, 11). I think this mainly shows that exchanging information between the two towers is helpful. It does not necessarily show that the proposed ray alignment is better than standard cross-attention.

The control in Table 12 re-initializes the cross-attention weights but keeps the ray encoding, so it also does not answer this question. Since the two cameras of one player share the same center, a plain cross-attention with image-coordinate positions may also learn the correspondence. It would be more convincing to compare with ordinary cross-attention, or cross-attention without the ray encoding, under the same training conditions.

Questions for authors

Please refer to weaknesses section

2 comments

We thank all reviewers for their thoughtful reviews. We are glad that the reviewers acknowledge our contributions, including the clear motivation of modeling out-of-view interactions, the usefulness of the dataset and collection pipeline, the well-motivated geometric design, and the extensive experiments. Below we respond to your comments point by point.

W1: How much of the improvement comes from the method vs. the data? As the reviewer notes, the panoramic stream is part of the proposed framework: predicting it jointly with the POV is our central proposal, so a comparison without any panoramic information evaluates a different model from the one we propose. We still report such controls to quantify each component, and we will add the following decomposition to the final version (POV metrics on the 30 two-player interaction episodes):

ConfigurationVisual input to the POV streamTraining dataPOV FVD ↓POV LPIPS ↓
Gamma-World (Bidirectional), best baselineinitial POVoriginal267.10.567
PAWA (Distilled), cross-attention removed at inference (Table 9)initial POV onlyours179.00.510
PAWA (Causal, intermediate), trained without cross-attention (Table 11)initial POV onlyours191.70.511
PAWA (Causal, intermediate), with cross-attention (Table 11, matched arm)POV + panoramaours138.40.455
PAWA (Distilled), fullPOV + panoramaours138.10.465

Without cross-attention, the POV tower receives no panoramic information, since pano→POV cross-attention is the only path from the panorama to the POV stream (Table 10). The POV stream in these rows therefore receives the same visual input as the baselines. The two Table 11 rows are intermediate causal checkpoints. We read the table in two steps and will state them in the final version:

  1. Data and training (267.1 → 179.0–191.7 FVD). A substantial part of the gap to prior models comes from our data and training, as the reviewer observes. We view this as part of our contribution: synchronized POV and panoramic multiplayer data at this scale exist only because the pipeline decouples CPU capture from GPU rendering.

  2. Panoramic information exchanged through the cross-attention (191.7 → 138.4 FVD). With the same data distribution, learning rate and batch size, the arm that keeps the cross-attention has 0.030 lower LPIPS (95% CI [0.016, 0.043]) after 2000 updates, during which both arms see identical batches, and 0.056 lower ([0.039, 0.073]) after 3000 updates, with lower LPIPS on 26 of 30 episodes. The LPIPS gap widens with training (0.012 → 0.030 → 0.056 over 0, 2000 and 3000 updates).

Removing the initial panorama. To test directly whether the advantage depends on the additional panoramic input, we re-ran PAWA (Distilled, 4 steps) on the same 30 two-player episodes with the initial panorama replaced by a projection of the initial POV alone (regions not covered by the POV are filled with uniform gray), so the model receives the same initial observation as the baselines. POV metrics, frames 1–239 (Gamma-World values are those of Table 1; the two PAWA rows were re-run together in a separate evaluation setup, under which the full initialization scores FVD 124.2 against 138.1 in Table 1, so they should be compared with each other):

Model and initializationStepsPSNR ↑LPIPS ↓FID ↓FVD ↓
Gamma-World (Bidirectional)3513.080.56756.6267.1
Gamma-World (Causal)3512.460.60652.0306.7
Gamma-World (Distilled)411.830.58458.4279.7
PAWA (Distilled), initial POV + panorama415.380.47028.4124.2
PAWA (Distilled), initial POV only415.350.48230.2129.4

With only the initial POV, PAWA (Distilled) remains better than all three Gamma-World variants on every POV metric, and close to its own result with the full initial panorama (FVD 129.4 vs. 124.2). The advantage over the baselines therefore does not depend on the additional initial panoramic observation.

W2: Isolating the ray alignment. We agree that Tables 9–11 show the value of the exchange, and that Table 12 shows that the trained attention follows the ray geometry rather than isolating the effect of the encoding. A vanilla cross-attention would have to learn a pixel mapping that changes with yaw and pitch at every frame, whereas the ray phase supplies this mapping in closed form and without parameters. We agree that a vanilla cross-attention trained under the same conditions is the direct comparison. Training it at full scale does not fit within the rebuttal period with the compute currently available to us, so we will add it to the final paper, together with a no-position-encoding variant and an explicit-warping baseline.

My concerns have either been resolved or will be resolved. I have raised the point after considering all the review comments

Reviewer_p52sOpenReview ↗
Rating
2(−1)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PAWA, a multiplayer world model that jointly generates each player’s POV video and world-aligned panoramic video to predict how players’ actions affect a shared environment. The authors develop a data collection pipeline that separates CPU-based gameplay execution and recording from GPU-based offline rendering, providing synchronized POV videos, panoramas, actions, and camera annotations.

The model consists of two DiT towers for POV and panoramic generation, connected through ray-aligned cross-attention that assigns identical rotary phases to tokens observing the same world-frame direction for the same player at the same time. It also introduces quadratic-phase player encoding to construct a larger identity codebook at a fixed phase dimension, a co-player-weighted flow-matching objective, and four-step distillation using Self Forcing and DMD.

Critical strengths
  1. The paper clearly identifies the important problem of out-of-view interactions. In a shared environment modified by multiple players, events occurring outside a player’s current POV should affect subsequent observations. PAWA addresses this by jointly predicting omnidirectional observations for each player. Using panoramas as both prediction targets and context for POV generation is well motivated by this problem.
  2. The geometric design connecting POV and panoramic observations is well motivated. Expressing both projections through world-frame viewing rays provides a common directional coordinate system despite their different image coordinates. Assigning zero relative phase to matching directions naturally accommodates changes in the POV camera’s orientation relative to the world-aligned panorama. The analysis of attention alignment also provides evidence that the model learns to exploit these directional correspondences.
Critical weaknesses
  1. The action-binding evaluation primarily measures whether each player’s camera rotation follows its own actions. It therefore does not directly establish whether changes in another player’s actions are correctly reflected in that player’s observations. LPIPS on reappearing co-player regions alone cannot distinguish reliance on the other player’s action inputs from reproduction of learned interaction patterns. Direct evaluation of whether POV and panoramic outputs, as well as observations across players, represent a consistent shared world state is also limited.
Questions for authors
  1. Could the authors hold the initial state and player A’s actions fixed, vary only player B’s actions, and evaluate whether the resulting changes in A’s observations match those in the actual game?
  2. The reported metrics assess video quality and similarity to ground truth, but do not directly measure whether multiple players’ videos simultaneously depict a consistent shared world. Could the authors provide a direct evaluation of cross-player consistency, as well as consistency between each player’s POV and panoramic outputs?
2 comments

We thank all reviewers for their thoughtful reviews. We are glad that the reviewers acknowledge our contributions, including the clear motivation of modeling out-of-view interactions, the usefulness of the dataset and collection pipeline, the well-motivated geometric design, and the extensive experiments. Below we respond to your comments point by point.

W1 / Q1: Do changes in another player's actions affect a player's observations? We ran a first counterfactual test during the rebuttal. Using PAWA (Distilled) on 30 two-player episodes, we keep the initial frames, the sampling noise and player A's actions fixed, and replace co-player B's actions with idle actions. In A's world-locked panorama, the resulting change concentrates in the direction of B: the ±20° azimuth window around B's position receives 1.47× the change of same-size control windows at other azimuths (95% CI [1.22, 1.76]). This tests the dependence of the panorama stream on B's actions; it is not yet a comparison against ground-truth counterfactual outcomes. Because our pipeline separates recording from rendering, we can replay modified actions and render the true outcome, and we will add this evaluation for both streams to the final paper. The re-entry analysis of the paper is also relevant (Table 7): a co-player stays outside a player's frustum for at least one second and then enters it; across 48 two-player, 111 four-player and 505 eight-player events, PAWA (Distilled) reconstructs the entering co-player with lower LPIPS than every baseline, with 95% intervals that exclude zero at every population, and the advantage holds when the co-player moves at least 2 m while hidden, where its new position is not in the initial frame.

Q2: Direct evaluation of cross-player and POV–panorama consistency. (a) POV–panorama consistency. A player's POV camera and panorama share the same center, so we reproject each generated panorama into the POV frustum using the camera orientation of that frame and compare it with the POV generated by the same model (LPIPS ↓, excluding the HUD rows). As references, we apply the same metric to the recorded videos, and to a time-shifted control that pairs each POV frame with the panorama of the same video 6 s away. Results on the 30 two-player interaction episodes (frames 1–239):

SourceMatched (same moment)Time-shifted control (6 s apart)Matched − control [95% CI]Episodes with matched < control
Recorded videos0.4440.618−0.174 [−0.192, −0.157]30/30
PAWA (Causal)0.5660.608−0.042 [−0.053, −0.032]30/30
PAWA (Distilled)0.6190.638−0.019 [−0.023, −0.014]29/30

Both PAWA variants are significantly more consistent than their time-shifted controls, on 30 of 30 (Causal) and 29 of 30 (Distilled) episodes, and PAWA (Causal) remains so in all four time bands of the 12 s rollout. The two towers thus generate the same world rather than two independently plausible ones. We will add this metric to the paper.

(b) Cross-player consistency. We do not yet have a direct metric for this. The closest evidence is the counterfactual test above (W1/Q1), in which changing co-player B's actions changes A's panorama in B's direction, together with the re-entry analysis (Table 7). In the final paper we will add a pose-based metric: the angular error between the direction in which B appears in A's generated panorama and the direction given by B's recorded pose, and the agreement of mutual visibility between the two players' views.

Presentation. In the final version we will move the definitions of the consistency and co-player metrics into the main text, add the table that separates the contribution of the data and training from that of the exchanged panoramic information, and use the additional page to bring the main conclusions of Tables 9–12 into the main text.

Thank you for the additional experiments and clarifications.

The main reason why author adapt the panorama view generation in multiple agent is sharing the state of world scene. I want to check that joint generation with panorama view and each ego centric view can really contains other agent’s actions and comunicates others actions by sharing the panorama space by detecting others.

Regarding Q1, my question was whether changing only player B’s actions causes player A’s world state or view exactly. For example, by measuring the location consistency in 3d space, the B’s movement should be exactly captured by A’s view. But proposed results that “1.47× the change of same-size control windows” is not represent the sharing the exact same state of world, even proposed Table. 7 don’t represent the result that panorama generation can contain the other’s action exactly.

Regarding Q2, similarly I asked that whether proposed multiple players’ videos simultaneously depict a consistent shared world in each panorama. but the rebuttal provides neither a result from the proposed metric nor a synchronized qualitative comparison that directly addresses this question.

Tracking the consequences of other players’ actions outside the current POV is a central motivation for using panoramas. But most of consistency is not thoughtfully provided and concerned in metric and experiments. I therefore consider the evidence for a consistent shared world across players insufficient, and I am lowering my score from 3 to 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.