acceptodds
Under review as a conference paper at ICLR 2027

WorldCeption: Unifying Perception and Generation for World Modeling

Abstract

Visual world modeling requires both inferring structure from observations and generating visual predictions under structural conditions. Although generative priors support visual perception, accurate perceptual readout alone does not establish that its representations can guide synthesis. A central challenge is to make dense perceptual fields and sparse states serve both purposes within a unified model. We introduce WorldCeption, a compact framework that unifies perception and conditional generation through a shared dense–sparse interface. Given RGB video, WorldCeption predicts dense perceptual representations and continuous or discrete sparse states; conversely, it synthesizes RGB video conditioned on dense representations or sparse states. Built upon a pretrained video transformer, the model combines shared attention with role-dependent feed-forward computation and task-dependent information access, accommodating perceptual readout and conditional synthesis within the same network. These complementary mappings can be composed, enabling estimated states to guide generation and perceptual representations to be translated through an RGB intermediate. Experiments show that estimated camera states retain reconstruction quality close to annotated conditions, while changing supplied states redirects recovered camera motion. Composed perceptual and generative paths further enable cross-representation completion without direct pairwise supervision. WorldCeption offers a compact architectural extension toward perception-centric world modeling through reusable visual representations. Code and models will be available upon acceptance.

Primary area:
generative models
Keywords:
Video Generation, Video Understanding, Video Perception, Controlled Generation, World Model

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_QTnROpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

WorldCeption raises an interesting and widely concerned question: Can a pre-trained Wan video model be reused for perception? The core approach of WorldCeption is a shared dense-sparse interface to support both perception and conditional generation simultaneously. It also introduces task-specific ffn and attention masks. The experiment evaluates perceptual accuracy, conditioned generation, reuse of camera state estimation and cross-representation completion through the generated RGB.

Critical strengths

The shared interface supports multiple readouts, including depth, normal, instances, etc. , and relies mainly on pre-trained video generation priors with limited additional parameters. This is a useful architectural approach for connection awareness and generation. Camera parameter reuse experiments also provide meaningful evidence that the estimated state can be used as a generation condition: the reconstruction quality is close to the GT condition generation.

Critical weaknesses

This article links perception and generation, which is a major difference from GenCeption, but I still haven't figured out what benefits this brings. The VGGT experiment, while retaining the jointly trained WorldCeption generator, replaced the camera estimation, which was insufficient as it did not compare the joint training with the individual training. I hope to see separate training versions of the two modules, using comparable data and training budgets, and compare their combined results. Regarding this, since this is a major difference from GenCeption, an unclear description of the benefits will greatly undermine the contribution of this article.

The generated results also lack strong comparisons. The reconstruction quality close to the gt condition generation indicates that the perception works with the generator, but it cannot tell us how good the generator itself is. Under the same first frame and camera state, the exclusive model will help evaluate video quality and control accuracy. Examples of dense conditional generation also lack comparisons with existing conditional generation methods. Given that generation is also a major new feature outside of GenCeption, I think the assessment of it is insufficient to support this contribution.

Questions for authors
  1. Have you compared the perception and generative models trained jointly with those trained separately? I want to know whether joint training can improve perception, generation or their combined performance.

  2. How does World Ception compare with the version of the same backbone trained only on the generation task? For video quality and camera-control accuracy, under the same conditions, is it compared with the existing methods?

  3. Have you tested multiple rounds of trajectory editing, where each generated video is passed back through perception before the next edit? Is the cumulative error in this process controllable? I think this will be a strong evidence of the model's reliability.

4 comments
AuthorsOpenReview ↗

Response to Reviewer QTnR (1/2) — The value of unification and cross-model state reuse

We sincerely thank the reviewer for the thoughtful comments, particularly the question of what unifying perception and generation provides beyond connecting separate models. We have added direct independent/joint comparisons and cross-model state-reuse experiments.

1. Independent versus joint training. P-only/G-only train for 2k updates each; shared-native S and role-aware W train jointly for 4k each. All start from the same pretrained source, use batch 64, and have matched per-task exposure. This matches task exposure, not FLOPs or optimizer trajectories.

EvaluationSpecialistShared SRole-aware W
VKITTI OOD depth AbsRel ↓0.3092 (P)0.47610.6266
VKITTI OOD semantic mIoU ↑0.3265 (P)0.31540.3474
Weather OOD accuracy ↑79.17% (P)87.50%91.67%
RE10K static LPIPS ↓0.3959 (G)0.43880.4470
RE10K dynamic LPIPS ↓0.3923 (G)0.43190.4323
VKITTI paired-weather LPIPS ↓0.3843 (G)0.40270.3703

The results show task-dependent transfer, not universal joint-training gains. Depth and most generation metrics favor specialists; selected semantic, categorical and paired-weather results favor W.

2. Cross-model state reuse. On 16 RE10K sources:

State providerG-only generatorJoint W generator
GT0.39590.4470
P prediction0.40540.4576
W prediction0.39150.4197

Values are LPIPS (↓). Independent P→G has a lower observed LPIPS than W→W (0.4054 vs. 0.4197); joint training is not necessary for state exchange. However, W states are usable by G, and with W fixed, W states improve over P states. This separates state compatibility from generator quality.

The demonstrated benefit is architectural integration and direct interoperability: one shared backbone provides perceptual state readout and conditional synthesis through the same explicit state interface, without requiring two separately trained backbones at deployment. We do not establish that this is more accurate or computationally cheaper than a separate P→G pipeline. Rather, we quantify the task-dependent trade-offs while demonstrating that both capabilities coexist in one model.

In the final-model reuse evaluation on 32 sources with two paired sampling-noise seeds, GT/self-estimated conditions yield LPIPS 0.2294/0.2307. Own prediction − GT gives PSNR −0.116 dB (paired 95% CI [−0.345, 0.084]) and LPIPS +0.0013 (paired 95% CI [−0.0043, 0.0073]). Both intervals include zero; this is not proof of statistical equivalence. Perception uses the observed video, not future prediction.

The P/G/S/W training comparisons use one training seed. We agree multiple seeds would strengthen the evidence, but additional full runs were infeasible within the rebuttal compute budget and timeline. We limit conclusions to the observed runs and intend to extend the study with multiple seeds when resources permit.

AuthorsOpenReview ↗

Response to Reviewer QTnR (2/2) — External conditional generation and recursive editing

We appreciate the reviewer’s suggestions to evaluate the generator against external conditional-generation systems and to test whether repeated state reuse accumulates error. We have added both experiments.

1. External camera-conditioned generation. We compare WorldCeption with Wan2.1-Fun-V1.1-1.3B-Control-Camera on 32 RE10K sources, using the same real first frame and factual/reversed-translation trajectories. Both generate 81 frames at 336×192 with native settings (WorldCeption: 40 steps/CFG 3; Fun: 50 steps/CFG 6); this is not an equal-compute comparison.

MetricWorldCeptionFun Camera
Factual LPIPS ↓0.23320.2533
Factual rotation error ↓0.4850°0.2572°
Factual direction error ↓18.7404°10.2848°
Edited rotation error ↓0.4969°0.2669°
Edited direction error ↓23.3645°12.3683°

WorldCeption has a lower observed factual LPIPS, while Fun follows camera motion more accurately. Six video-quality dimensions are mixed; dynamic degree measures motion amount, not monotonically better quality.

2. External depth-conditioned generation. We compare with Wan2.1-VACE-1.3B on four procedural compositions/12 windows. Both receive the same raw depth and task text, without an RGB reference, using native codecs and sampling settings.

MetricWorldCeptionVACE
Independent depth AbsRel ↓0.10030.1507
δ1 ↑0.92230.8600
RMSE ↓6.22048.1714
Auxiliary LPIPS ↓0.53890.6098

Depth evaluation uses per-frame median scale alignment and measures relative geometry, not absolute-scale preservation. These external comparisons use different native sampling configurations and are not equal-compute comparisons.

3. Multi-round trajectory editing. We evaluate 16 RE10K sources over three rounds (+δ, +δ, −2δ), retaining the original real first frame at every round. An independent motion evaluator scores outputs against fixed nominal targets.

Direction error ↓Round 1Round 2Round 3
Recursive edit16.39°20.07°22.53°
Direct edit16.39°17.78°17.54°
Recursive identity17.54°18.27°22.17°
Direct identity17.54°17.54°17.54°

Recursive feedback accumulates more motion deviation than direct execution, including without edits. Command drift grows from 0° to 10.49° and 14.07°. This is a three-round diagnostic, not evidence of long-horizon stability. Appearance/identity stability has not been independently quantified.

These external and recursive evaluations use one fixed sampling seed. The two T2V noise seeds are paired sampling seeds, not independent training seeds. We agree multi-seed training evaluation would strengthen the evidence, but additional full runs were infeasible within the rebuttal timeline and compute budget. We explicitly limit our conclusions to the observed runs and intend to extend the study with multiple training seeds when resources permit.

Thanks for the rebuttal and the additional experiments. The response addressed most of my questions. I appreciate the effort the authors put into it. I still find the contribution somewhat limited, so I’ll keep my score at 2. I’m open to either outcome, if the AC decides to accept the paper, that’s fine with me.

Thank you very much for your thoughtful reassessment and for taking the time to consider our additional experiments. We are glad that the rebuttal has addressed most of your concerns, and we sincerely appreciate your constructive feedback throughout the review process.

We understand your remaining reservations regarding the scope of the contribution, and we respect your assessment. Your comments have nevertheless helped us substantially strengthen the experimental evidence and clarify the contributions of our work. We also appreciate your openness to the AC's final decision.

Reviewer_2M1wOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WorldCeption adapts a pretrained video diffusion model to support both perception and conditional generation. Dense representations, camera trajectories, and weather categories serve as prediction targets and generation conditions through shared attention, role-dependent FFNs, and task-dependent attention masks. Experiments evaluate individual tasks, reuse of estimated camera states for synthesis, and cross-representation completion through generated RGB.

Critical strengths

The shared interface supports a useful range of perception and generation tasks with only 1.21% additional backbone parameters. Using the same final weights for dense prediction, camera estimation, and conditional synthesis is a meaningful integration result.

The camera-state reuse experiment provides concrete evidence that predicted states remain useful for generation. Across 32 sources and two seeds, replacing annotated states with estimated states changes PSNR by only −0.116 dB and LPIPS by +0.0013. The wrong-source intervention also shows that generated motion responds to the supplied trajectory.

Critical weaknesses

The architectural benefit is not cleanly established. Table 3 uses different expert initializations across routing variants and only one training seed, so it cannot isolate the effect of role-dependent routing. Its control metrics measure sensitivity rather than motion accuracy. Table 2 supports state reuse, but does not isolate the benefit of joint training over a separately trained perception–generation pipeline. This matters because the shared architecture is a central contribution.

The cross-representation completion evidence is too narrow to establish a reliable capability. The 96 conversions all come from one Scene01 time window, and the direct reference feeds annotations into a path trained for RGB input rather than providing a trained competing method (Appendices J and N). The paper also reports missing objects and semantic-region drift, without an independent quantitative evaluation of constraint preservation. The examples show plausible completion, but leave its reliability unclear.

Questions for authors

Can you provide a routing comparison with identical expert initialization, evaluated using recovered-motion accuracy and reconstruction quality? This would help determine whether the proposed routing improves control beyond simply increasing sensitivity.

Do you have cross-representation results on independent held-out scenes, using the final model and an independent measure of input-constraint preservation? A comparison with a separately trained perception–generation pipeline would also help establish the value of unification.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2M1w (1/2) — Routing and independent versus joint training

We sincerely thank the reviewer for the careful technical assessment. We have added an initialization-matched routing experiment and task-exposure-matched independent/joint comparisons.

1. Initialization-matched control routing. Additive and Role-dependent use identical initialization and trainability for all 971 tensors, including nonzero experts, and the same 64,000-sample execution ledger over 1,000 updates. Only the known-control FFN changes (W+A vs. A); dense and query paths remain W and W+A.

HistoryRoutingLPIPS ↓PSNR ↑Rotation ↓Direction ↓
StaticAdditive0.469413.1803.548°67.79°
StaticRole-dependent0.419512.9502.122°63.97°
DynamicAdditive0.467913.2882.821°59.22°
DynamicRole-dependent0.414612.8601.633°53.75°

Correct-versus-null LPIPS utility improves from approximately −0.0006 to 0.0485/0.0557 for static/dynamic history. Motion is independently recovered using VGGT. Thus, the result is not merely output sensitivity. PSNR decreases and translation differences remain inconclusive.

2. Independent versus joint training. P-only/G-only train for 2k updates each; shared S and role-aware W train for 4k each. All start from the same pretrained source, use batch 64, and have matched per-task exposure. This does not match FLOPs or optimizer trajectories.

EvaluationP/G specialistShared SRole-aware W
VKITTI OOD depth AbsRel ↓ (P/S/W)0.30920.47610.6266
VKITTI OOD semantic mIoU ↑ (P/S/W)0.32650.31540.3474
Weather OOD accuracy ↑ (P/S/W)79.17%87.50%91.67%
Blender normal mAE ↓ (P/S/W)12.5121°12.1012°11.9148°
RE10K static LPIPS ↓ (G/S/W)0.39590.43880.4470
RE10K dynamic LPIPS ↓ (G/S/W)0.39230.43190.4323
VKITTI paired-weather LPIPS ↓ (G/S/W)0.38430.40270.3703

Joint training yields task-dependent transfer, not universal gains: selected semantic, categorical, normal and paired-weather results favor W, while depth and most generation results favor specialists. Blender normal is a training-source diagnostic; Weather OOD has 48 samples.

These historical comparisons use one training seed. We agree multi-seed evaluation would strengthen the evidence, but additional full runs were infeasible within the rebuttal compute budget and timeline. We therefore limit conclusions to these observed runs; source-level variation is not training-seed robustness. We intend to extend the comparisons with multiple training seeds when resources permit.

AuthorsOpenReview ↗

Response to Reviewer 2M1w (2/2) — Cross-model state reuse and cross-representation completion

We appreciate the reviewer’s concern that state reuse alone does not establish reliable cross-representation composition. We have added a multi-composition evaluation and explicitly delimit what it demonstrates.

1. Cross-model state reuse. On 16 RE10K sources, we evaluate GT, P-predicted and W-predicted camera states with both the independent G and joint W generators:

State providerG-only generatorJoint W generator
GT0.39590.4470
P prediction0.40540.4576
W prediction0.39150.4197

Values are LPIPS (↓). Independent P→G has a lower observed LPIPS than W→W (0.4054 vs. 0.4197), so joint training is not necessary for state exchange and is not universally better. However, W-predicted states are usable by G; with W fixed, W states improve over P states (0.4197 vs. 0.4576). This separates state compatibility from generator quality. Lower LPIPS does not imply greater physical accuracy than GT.

2. Completion on additional procedural compositions. We evaluate four preselected evaluation-only procedural compositions, each with three non-overlapping 81-frame windows and one fixed sampling seed. The 48 generated RGB intermediates cover four dense→RGB→dense chains:

CompositionCompletion metricReal-RGB readout reference
Depth→Normal, mAE ↓8.4545°6.9443°
Normal→Depth, AbsRel ↓0.31590.2007
Semantic→Depth, AbsRel ↓0.34300.2007
Instance→Semantic, mIoU ↑0.13700.2649

These are new procedural compositions, not necessarily unseen assets; 12 windows are not 12 independent scenes. Real RGB is an information-richer reference, not a same-input translator baseline.

For independent depth-constraint checking, Depth Anything V2-L recovers generated depth with AbsRel 0.1003 and δ1 0.9223, using per-frame median scale alignment. Independent normal, semantic and instance constraint scores remain unavailable. The results therefore support representation-dependent composition, not arbitrary or natural-scene generalization.

An early-budget pipeline comparison also favors G2k→P2k over W4k→W4k: Depth→Normal mAE is 11.3973° vs. 12.0653°, and Semantic→Depth AbsRel is 0.1880 vs. 0.2242. These are separate from the final-model results.

The completion experiment uses one fixed sampling seed. We acknowledge that multiple training seeds would strengthen robustness analysis, but could not complete additional training runs within the rebuttal timeline and compute budget. We make no training-seed robustness claim.

With the same initialization, role-dependent routing reduced camera rotation error in both settings. That answers my main question about camera control, so I’m raising my score from 2 to 3.

Thank you very much for reconsidering our work and for acknowledging the additional controlled experiment. We are glad that the initialization-matched routing comparison addresses your main concern regarding camera control. We also appreciate your constructive feedback, which helped us strengthen the experimental evidence and clarify the contribution of our design.

Reviewer_Wb4hOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose WorldCeption, adapting a video transformer to a bidirectinal interface that transfers between dense representations, sparse states and RGB observations. Video VAE is applied for dense modalities like semantics, depth, etc. while role-dependent feedforward paths are used for sparse controls and prediction queries.

Critical strengths
  1. The model design is well motivated: It's a sensible choice to separate unknown-state queries from supplied controls. The attention-mask argument clearly limits its guarantee to latent-prefix noninterference.
  2. The presentation is sound and clear. Notions are easy to follow.
Critical weaknesses
  1. Experiemental results: The results in Table 1 do not demonstrate better or even comparable performance compared to those specialist models.
  2. The architectural ablations do not establish the proposed mechanisms’ contribution. The results in Table 3 do not show clear advantages of a specific design choice.
Questions for authors
  1. Can generation retention be compared directly with original Wan, including temporal quality and runtime?
1 comment
AuthorsOpenReview ↗

Response to Reviewer Wb4h

We sincerely thank the reviewer for the thoughtful evaluation and constructive feedback. We appreciate the recognition of our architectural motivation and the suggestions regarding specialist comparisons, routing ablations, and generation retention. In response, we have conducted extra controlled experiments and revised the manuscript to clarify both the capabilities and limitations of WorldCeption.

1. Performance relative to task-specialized models and the value of unification.

We agree that WorldCeption does not match every task specialist. Under our common local evaluation protocols, Depth Anything V2-L remains stronger on VKITTI2 depth (AbsRel 0.0570 vs. 0.1146), and VGGT is stronger on RE10K camera direction (11.86° vs. 23.10°). Normal estimation is dataset-dependent:

EvaluationReference modelReferenceWorldCeption
Sintel Clean mean angular error ↓GenCeption-1.3B32.46°30.54°
Sintel Final mean angular error ↓GenCeption-1.3B30.26°33.67°

Our contribution is not universal specialist-level accuracy, but a shared interface through which heterogeneous world states can be estimated and directly reused as generation conditions. The same backbone supports dense/sparse state readout, categorical readout, conditional generation, and T2V, with 17.16M additional interface parameters (1.21% of the backbone size).

We evaluate this reuse directly on 32 RE10K sources with two paired sampling-noise seeds:

Own prediction − GT conditionDifferencePaired 95% CI
PSNR−0.116 dB[−0.345, 0.084]
LPIPS+0.0013[−0.0043, 0.0073]

Both intervals include zero. This indicates a small observed reconstruction cost, not statistical equivalence. Perception uses the observed video, so this is state reuse rather than future prediction.

2. Initialization-matched routing ablation.

We agree that the original routing ablation did not fully isolate the effect of expert initialization. We therefore conducted an additional controlled comparison. Both variants use identical initial values and trainability for all 971 parameter tensors, including nonzero experts, and follow the same 64,000-sample execution ledger over 1,000 updates. The only architectural difference is the FFN path for supplied control tokens: Additive uses W+A, while Role-dependent uses A. Dense and query pathways remain unchanged.

HistoryRoutingLPIPS ↓PSNR ↑Recovered rotation error ↓
StaticAdditive0.469413.1803.548°
StaticRole-dependent0.419512.9502.122°
DynamicAdditive0.467913.2882.821°
DynamicRole-dependent0.414612.8601.633°

The correct-versus-null LPIPS utility also increases from approximately −0.0006 to 0.0485 (static) and 0.0557 (dynamic). These results indicate improved use of supplied camera conditions rather than merely greater output sensitivity. However, PSNR decreases, and translation-direction errors remain substantial. We therefore limit the conclusion to the tested camera subsystem.

3. Original-Wan generation retention and runtime.

We directly compare original Wan2.1-T2V-1.3B with WorldCeption using 32 fixed prompts and two paired sampling-noise seeds at 256×448 resolution and 81 frames.

DimensionOriginal WanWorldCeptionΔ
Aesthetic quality ↑0.470530.49447+0.02394
Imaging quality ↑0.609100.70438+0.09528
Overall consistency (ViCLIP) ↑0.236270.18846−0.04781
Subject consistency ↑0.937620.95642+0.01880
Background consistency ↑0.964870.96148−0.00339
Motion smoothness ↑0.983800.99261+0.00881
Dynamic degree0.515620.42188−0.09375

Generation is partially retained, not losslessly preserved. Overall consistency decreases, while the lower dynamic-degree point estimate does not establish a reliable reduction in motion amount because its paired interval includes zero.

Runtime measurements give median sampling-through-decode times of 18.45 s for original Wan and 26.93 s for WorldCeption. Both peak at 5.22 GiB on the tested native T2V backbone path; this does not represent all interfaces resident simultaneously. We report the measured slowdown rather than claim an inference-speed advantage.

4. Training-seed limitation.

The routing ablation uses one training seed. The two T2V noise seeds are sampling seeds, not independent training runs. Additional full training-seed repetitions were not feasible within the rebuttal compute budget and timeline. We agree that multi-seed evaluation would strengthen the evidence and intend to extend these controlled comparisons when resources permit. We do not interpret source-level intervals as training-seed robustness.

The revised manuscript includes the corresponding controlled ablation, state-reuse results, original-Wan comparison, and runtime measurements.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Discussion Outcomes

We sincerely thank the Area Chair and all reviewers for their careful evaluation and constructive feedback. The discussion has helped us strengthen the experimental evidence and clarify the contribution and limitations of WorldCeption.

1. Discussion Outcomes

Following the rebuttal:

  • Reviewer 2M1w raised the score from 2 to 3, explicitly acknowledging that the initialization-matched routing experiment reduced camera rotation error in both settings and addressed the main concern regarding camera control.
  • Reviewer QTnR maintained the score of 2, but acknowledged that the rebuttal addressed most of the questions and expressed openness to either outcome based on the AC's decision.
  • Reviewer Wb4h maintained the score of 3 after considering our additional experiments and the other reviewers' comments.

We appreciate that the additional evidence helped address several technical concerns, while some differences regarding the overall contribution remain.

2. Central Contribution

We would like to clarify that WorldCeption is not simply a multi-task model combining perception and generation. It establishes a shared, explicit world-state interface connecting visual understanding and conditional synthesis.

In conventional perception pipelines, estimated states are primarily treated as prediction outputs. In WorldCeption, heterogeneous dense representations and sparse states serve two complementary roles: they are targets inferred from RGB observations and explicit conditions guiding RGB video synthesis. This allows the same model to estimate world states, reuse its predictions for generation, and compose perceptual and generative mappings through RGB intermediates.

The central contribution is therefore the operational reuse of world representations across understanding and generation within a shared pretrained video model, rather than simply increasing the number of supported tasks.

3. Additional Evidence

We conducted targeted experiments addressing the reviewers' main concerns:

  • Routing: An initialization-matched comparison across 971 parameter tensors demonstrates improved recovered camera rotation accuracy in both static and dynamic settings.
  • Joint vs. Separate: P/G/S/W comparisons and cross-model state reuse quantify task-dependent transfer and the trade-offs of joint training.
  • Completion: Four additional procedural scene compositions and independent depth evaluation provide broader evidence of cross-representation composition.
  • Generation: Comparisons with Fun Camera and VACE, together with a seven-dimensional original-Wan evaluation, characterize conditional-generation quality and capability retention.
  • Recursive Editing: Three-round trajectory experiments reveal measurable motion-error accumulation under recursive feedback.

Importantly, we retain negative results. Specialized models remain stronger on several individual tasks, independent P→G can outperform the jointly trained pipeline, and recursive feedback introduces motion drift. These findings characterize the actual benefits and limitations of the proposed approach.

4. Remaining Limitations and Significance

We acknowledge that our experiments do not establish universal superiority over specialists, arbitrary cross-representation generalization, or long-horizon recursive stability. The additional training comparisons also use a single training seed. Due to limited computational resources and the rebuttal timeline, we could not complete multiple full training runs. We agree that multi-seed evaluation would strengthen the evidence and intend to pursue it in future work.

Overall, we believe WorldCeption makes a meaningful contribution to perception-centric world modeling: heterogeneous world-state estimation, conditional visual synthesis, and explicit state reuse can coexist within a shared pretrained video model. Our additional experiments provide controlled evidence for this integration, quantify its performance trade-offs, and identify its current limitations.

We sincerely thank the reviewers and the Area Chair for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.