- 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
-
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.
-
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?
-
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
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.
| Metric | WorldCeption | Fun Camera |
|---|---|---|
| Factual LPIPS ↓ | 0.2332 | 0.2533 |
| 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.
| Metric | WorldCeption | VACE |
|---|---|---|
| Independent depth AbsRel ↓ | 0.1003 | 0.1507 |
| δ1 ↑ | 0.9223 | 0.8600 |
| RMSE ↓ | 6.2204 | 8.1714 |
| Auxiliary LPIPS ↓ | 0.5389 | 0.6098 |
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 1 | Round 2 | Round 3 |
|---|---|---|---|
| Recursive edit | 16.39° | 20.07° | 22.53° |
| Direct edit | 16.39° | 17.78° | 17.54° |
| Recursive identity | 17.54° | 18.27° | 22.17° |
| Direct identity | 17.54° | 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.
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.
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:
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.