acceptodds
Withdrawn from ICLR 2027

Unveil as You Go: Rethinking Spatial Exploration for Single-Image World Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The integration of generative models and reconstruction networks has emerged as a promising paradigm for 3D Gaussian world generation. However, existing methods mostly treat camera exploration as a plug-and-play input rather than scene-adaptive understanding, and the spatial priors encoded in pretrained vision–language models remain underexploited. To address this, we propose UnveilWorld, in which the explorer performs recursive planning and understanding and the generator unveils unobserved views, jointly contributing to the final world reconstruction. To further investigate the synergy between spatial understanding and generation, UnveilWorld is post-trained with hard-state step-wise AWM, pairing stage-wise group-relative credit assignment with failure-biased state propagation. The joint optimization of the full pipeline closes the loop among exploration, generation, and reconstruction end-to-end. Experiments demonstrate that UnveilWorld generates high-quality 3D worlds with strong cross-view consistency in 18.12 seconds on a single NVIDIA A800 GPU.

Primary area:
generative models
Keywords:
World Generation, Spatial Navigation, Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4ihDOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UnveilWorld, a pipeline for single-image 3D Gaussian world generation. Instead of using fixed camera paths, the method uses a depth model and vision-language model to plan scene-adaptive camera trajectories and descriptions. These trajectories then condition a 4-step diffusion model initialized from FlashWorld to decode 3D Gaussians. The authors train the system with an RL post-training setup called Hard-State Step-wise AWM, and evaluate on WorldScore-Static benchmarks.

Critical strengths
  • Neat planning idea: Using depth and vision-language priors to plan camera exploration paths instead of using fixed trajectories is a neat idea.
Critical weaknesses
  • No video results or project website: This is a 3D world generation submission, but there are zero video results. There is no project website, no anonymous demo link, and no zip file with videos. The paper only shows static images in the pdf. For 3D world and scene generation, you need videos to check 3D consistency, novel view stability, geometry collapse, and rendering artifacts. It is impossible to evaluate a world generation paper from still images. Having no videos is an immediate reason to reject.
  • Low image metrics: In Table 2 and Table 3, the reconstruction numbers on DL3DV are very low. The PSNR is only 13.5 - 14.3 dB, and SSIM is around 0.35 - 0.42. These values mean the rendered novel views have very low image quality and look blurry or washed out.
  • Unclear RL post-training value: The paper introduces a complicated RL scheme with hard-state curriculum and step-wise VLM rewards. But looking at the ablation in Table 3, the improvement is very small (PSNR only goes from 12.26 dB to 12.67 dB). It is not clear if all this extra RL complexity is actually needed.
  • Complex pipeline: The entire pipeline stacks many separate models together (MoGe-3 depth, Qwen2.5-VL planner, DiT trajectory generator, T5 text encoder, VAE, DiT world generator, and 3DGS decoder). If the depth estimator or VLM planner fails early on, those errors propagate and break the whole world generation.
Questions for authors

The idea of scene-adaptive exploration is cool, but having no video results makes it impossible to accept this submission.

I would like authors to address following questions:

  • Why are there no videos or demo website provided? How can reviewers check 3D consistency and novel view synthesis quality without videos?
  • Why are the PSNR numbers in Table 2 and Table 3 so low (around 13-14 dB)?
  • Why is the hard-state AWM RL setup needed when the metric gains in Table 3 are so marginal?
  • How robust is the pipeline when depth estimation fails or the VLM outputs a bad path?

I am clear reject because there are no video results, but happy to consider a rebuttal and other reviews.

Reviewer_UcWsOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes UnveilWorld, a framework for single-image 3D Gaussian world generation that integrates scene-adaptive exploration, generation, and reconstruction. Specifically, the authors introduce an exploration module that leverages pretrained vision-language models and depth-based spatial priors to generate scene-aware camera trajectories and corresponding textual descriptions. These conditions are then used by a four-step 3DGS generator initialized from FlashWorld. The authors also propose hard-state step-wise Advantage Weighted Matching (AWM), which incorporates stage-wise reward feedback and propagates low-scoring candidates during training. Experiments demonstrate improved 3D world generation quality and cross-view consistency compared to existing methods.

Critical strengths
  • I think integrating scene-adaptive exploration with 3D world generation is an interesting and practical direction. Instead of relying on predefined camera trajectories, the proposed framework uses pretrained spatial priors to determine where to explore and what content to generate.

  • The use of a shared depth-based spatial anchor to align exploration trajectories with scene geometry is a reasonable approach. It provides a practical way to address the scale ambiguity of camera trajectories across different scenes.

  • The authors provide extensive experimental comparisons with recent world generation methods. The results demonstrate promising improvements in visual quality and cross-view consistency while maintaining efficient inference.

Critical weaknesses

The novelty and effectiveness of the proposed exploration strategy could be better justified. The authors argue that existing world generation methods primarily rely on predefined or scene-agnostic camera trajectories. However, previous Text-to-3DGS works such as Director3D [1], SplatFlow [2], and VideoRFSplat [3] have already explored camera pose generation alongside multi-view or 3D scene generation. In particular, SplatFlow and VideoRFSplat jointly generate multi-view images and camera poses, while HY-World 2.0 also incorporates scene-aware trajectory planning. Although these methods differ in their conditioning settings and exploration objectives, the authors should better clarify the methodological novelty of the proposed reconstruction-oriented exploration strategy. Moreover, Table 2 compares the proposed method with a fixed trajectory that observes in place and a Common VLA trajectory whose implementation is not sufficiently described. Stronger comparisons with existing learned trajectory generation or scene-adaptive planning methods would help establish the contribution of the proposed exploration module.

[1] Li et al., Director3D: Real-world Camera Trajectory and 3D Scene Generation from Text, NeurIPS 2024.
[2] Go et al., SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis, CVPR 2025.
[3] Go et al., VideoRFSplat: Direct Scene-Level Text-to-3D Gaussian Splatting Generation with Flexible Pose and Multi-View Joint Modeling, ICCV 2025.

Questions for authors
  • Could the authors better distinguish the proposed exploration strategy from prior works that jointly generate camera trajectories and scene content, such as SplatFlow and VideoRFSplat? In particular, what additional benefits does reconstruction-oriented exploration provide over learned camera trajectory generation?
Reviewer_GhPeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper generates an explorable 3D Gaussian world from a single image. A vision-language planner initialized from a navigation model, together with a trajectory DiT and a depth-based scale normalization, plans a camera trajectory and step-wise text descriptions. A generator initialized from FlashWorld then builds the world along this trajectory in four denoising stages. The generator is post-trained with Hard-State Step-wise AWM: at each intermediate stage a VLM scores a group of candidates and the lowest-scoring valid one is passed on to the next stage, and the final worlds are scored against ground truth. Finally, the planner, trajectory DiT and generator are trained jointly, which the paper presents as closing the loop between exploration, generation and reconstruction.

Critical strengths
  1. The main idea is interesting and changes what is optimized. The camera trajectory is not treated as a fixed input but becomes part of what is learned, so that planning and generation can in principle improve each other. Table 2 supports that this is worth doing: with the generator fixed, the construction trajectory makes a large difference (e.g. +2.5 dB PSNR over a fixed trajectory). This is the main reason I think the paper could be accepted once the evidence below is in place.

  2. The final system is strong and fast. In Table 1 it is first on four of six metrics and second on the other two against recent single-image world generators, and it takes 18.12 s per world on one A800 (Appendix A.5 states clearly what the timing includes). So the system clearly works; my open questions are about which components the gains come from, not whether the method works.

Critical weaknesses
  1. It is unclear whether the loop is actually closed. The planner's training data are described (Sec. 3.5: about 60k trajectory–description tuples), but its training procedure is not. According to Appendix A.2, AWM updates only the generator, and the only stage that updates the planner is joint training. I could not find a supervised stage or a loss that uses the data-engine tuples, and the paper does not say how the objective in Eq. (5) reaches the planner. Appendix A.4 makes this harder to follow: candidates in a group "share the reference image, spatial anchor, camera conditions, and text", and candidate generation and scoring "do not retain a gradient graph", so the group advantages compare noise samples under the same trajectory. The only test of joint training is the last row of Table 3, which gets 300 extra updates (600 vs. 300) and has mixed gains (LPIPS 0.5833 → 0.5867). This is consequential because if the planner is in effect not trained, the learnable exploration module reduces to the pretrained InternVLA-N1 components, and two of the three contributions, the closed loop and the learnable planner, are not demonstrated.

  2. The planner's contribution is not separated from the generator. Because the two are trained together, a gain could come from them adapting to each other rather than from better exploration. Table 4 includes FlashWorld driven by the planner's trajectory (67.65 vs. 72.45 for the full method), but there is no FlashWorld result with its own default trajectory, and FlashWorld is missing from Table 1. Table 2 does not say which generator checkpoint it uses, the fixed baseline observes in place, and the Common VLA baseline is not specified. There are also no statistics on path length or coverage, so I can't rule out that the planned trajectories score better simply because they cover more of the scene. Without this, the value of the planner, which the paper is built around, can't be assessed. I think both weaknesses can be addressed in the rebuttal; Questions 1–3 list what would raise my score.

Questions for authors
  1. How exactly is the planner trained? Is the data-engine data used in a supervised stage, and with which loss? In joint training, which planner outputs receive gradients from Eq. (5), given that candidates share the trajectory and no gradient graph is kept for generation? How does joint training compare with simply continuing AWM for 300 more updates with the planner frozen?

  2. The experiment I would most like to see: the original FlashWorld (no post-training) driven by (a) its default trajectory, (b) a simple trajectory with similar path length and coverage, and (c) the planner's trajectory, all under the Table 1 protocol. Could the authors also say which generator Table 2 uses, describe the Common VLA baseline, and report statistics of the planned trajectories (path length, coverage, distance to the real camera path)?

  3. The group-relative weighting comes from AWM and GRPO, so the new element of the generator training is passing the worst valid candidate forward. Table 3's single-sample baseline changes both the number of candidates and the objective. How does propagating the worst valid candidate compare with propagating a random or the best one, with the same candidate count and update budget?

  4. Appendix B.1 states a scene-level exclusion only for the DL3DV experiments (Tables 2–3). Can the authors confirm that the WorldScore-Static scenes in Tables 1 and 4 do not overlap with the training data, including the private videos? Also, Table 2's GT views come from real DL3DV camera paths, and the planner's data are built from the same kind of real capture paths, so this evaluation may favor a planner that imitates typical real trajectories. How do the authors see this?

  5. The depth-based anchor is described as necessary ("only geometry-aligned trajectories can reliably steer the generator"), but I could not find an ablation against distance-normalized trajectories. Could the authors add one?

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

Related papers

Discussion (0)

Sign in to comment.