World building with 3D scene representations is increasingly important for content creation, simulation, and interactive experiences, yet real workflows are inherently iterative: creators repeatedly extend existing scenes under user control. Motivated by this gap, we study text-guided 3D scene expansion via free-form view insertion. Starting from a real scene captured by multi-view images, a user specifies a text expansion intent, which a generative model materializes as an inserted view extending the scene coverage. Unlike simple object editing or style transfer within a fixed scene, the inserted view may be 3D-misaligned with the original reconstruction, introducing geometric shifts, hallucinated content, or view-dependent artifacts that disrupt global multi-view consistency. To address this challenge, we propose SceneExpander, which applies test-time adaptation to a parametric feed-forward 3D reconstruction model with two complementary distillation signals: anchor distillation stabilizes the captured scene using geometric cues from the captured views, while inserted-view self-distillation retains insertion-supported predictions to accommodate the misaligned view. Experiments on ETH scenes and online data demonstrate improved expansion behavior and reconstruction quality under misalignment.
Figures & tables
Figure 2: Challenges in text-guided 3D scene expansion via free-form view insertion. (1) A 3D-misaligned inserted view can corrupt reconstruction quality in the captured region. (2) Newly introduced content must remain consistent under novel viewpoints. Red boxes highlight ghosting artifacts caused by conflicting 3D constraints from the inserted view.
Figure 3: Overview of SceneExpander. (1) A user provides a text expansion instruction, which an image generative model materializes as a free-form inserted view Ig ; a feed-forward reconstructor fθ predicts an initial 3D scene from captured views. (2) To accommodate the text-guided inserted view that may be 3D-misaligned, we adapt fθ at test time using two distillation losses: anchor distillation on captured views and inserted-view self-distillation on Ig . (3) The adapted model produces an expanded 3D scene consistent with the captured region while incorporating the inserted content.
Figure 4: Geometry-perturbation augmentation. We show four augmentation modes: identity (original), global affine warp, blockwise piecewise-affine warp (grid 3×3 ), and their combination (global+block), designed to mimic pose drift and local geometric inconsistencies.
Figure 5: Qualitative comparison of scene expansion under misaligned insertion. We show an indoor scene from the Online collection (left) and an outdoor scene from the ETH dataset (right). The first row presents the generated inserted view (purple box) together with a subset of captured multi-view images from the original scene (blue box). Each subsequent row shows renderings at novel camera poses for different methods, illustrating both fidelity to the captured region and insertion satisfaction. Please refer to the media supplement for video visual results.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
GPT-IS ↑
GPT-RQ ↑
H-IS ↑
H-RQ ↑
OminVGGT ( Peng et al. 2025 )
14.969
0.3647
0.370
3.16
2.66
2.26
2.40
TTT3R ( Chen et al. 2025b )
13.483
0.2747
0.457
2.94
2.38
1.72
1.71
WorldMirror ( Liu et al. 2025 )
19.295
0.614
0.357
3.50
2.61
4.04
4.07
3DGS ( Kerbl et al. 2023 )
14.046
0.3045
0.418
3.16
2.61
2.40
2.32
SceneExpander (Ours)
21.206
0.776
0.276
3.72
3.05
4.56
4.53
Table 1: Quantitative comparison. We report preservation on the captured region using image-space fidelity metrics (PSNR/SSIM/LPIPS) and expansion quality using LLM-based and human evaluation. GPT-IS/H-IS measure insertion satisfaction, and GPT-RQ/H-RQ measure overall rendering quality. ↑ higher is better, ↓ lower is better.
Variant
PSNR
SSIM
LPIPS
GPT-IS
GPT-RQ
Baseline
19.295
0.614
0.357
3.50
2.61
+ Anchor distill.
20.763
0.714
0.329
3.52
2.96
+ Subset samp. & aug.
20.892
0.720
0.322
3.56
3.24
+ Self-distill. ( p=0.5 )
21.124
0.757
0.294
3.70
3.08
+ Self-distill. ( p=1.0 )
20.856
0.704
0.318
3.84
2.88
+ Stoch. rest. (Full)
21.206
0.776
0.276
3.72
3.05
Table 2: Ablation study. Each row adds one component to SceneExpander. “distill.” denotes distillation, “samp.” denotes sampling, “aug.” denotes augmentation, and “stoch. rest.” denotes stochastic restoration.
Figure 6: Multi-round expansion. Starting from an initial captured scene, users can repeatedly insert new views to expand the reconstruction. Across consecutive rounds, SceneExpander progressively enlarges the 3D scene while preserving previously reconstructed regions.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: More cases of SceneExpander. Left of the dashed line: input images, where the purple dashed box indicates the generated inserted view and the remaining images are captured views. The second and third cases show a representative subset of captured views for brevity. Right of the dashed line: novel views and corresponding depth maps rendered by camera navigation in the expanded scene.
Figure 8: User study interface. Participants rank videos A–E by insertion satisfaction (H-IS) based on consistency with the edit goal, with the edited region highlighted in red. Options A–E include our method and four baselines, shown in a randomized order. The H-RQ interface follows the same format.
Figure 9: A failure case of SceneExpander. The purple box shows the generated inserted view, and the red boxes highlight artifacts. SceneExpander alleviates misalignment on the two side pillars and the left stone pedestal, but still produces errors on the top crossbeam.
Figure 10: Prompt for VLM-based evaluation on insertion satisfaction.
Figure 11: Prompt for VLM-based evaluation on rendering quality.
We present GuidedSceneGen, a text-to-3D generation framework that produces metrically accurate, globally consistent, and semantically interpretable indoor scenes. Unlike prior text-driven methods that often suffer from geometric drift or scale ambiguity, our approach maintains an absolute world coordinate frame throughout the entire generation process. Starting from a textual scene description, we predict a global 3D layout encoding both semantic and geometric structure, which serves as a guiding proxy for downstream stages. A semantics- and depth-conditioned panoramic diffusion model then synthesizes 360° imagery aligned with the global layout, substantially improving spatial coherence. To explore unobserved regions, we employ a video diffusion model guided by optimized camera trajectories that balances coverage and collision avoidance, achieving up to 10x faster sampling compared to exhaustive path exploration. The generated views are fused using 3D Gaussian Splatting, yielding a consistent and fully navigable 3D scene in absolute scale. GuidedSceneGen enables accurate transfer of object poses and semantic labels from layout to reconstruction, and supports progressive scene expansion without re-alignment. Quantitative results and a user study demonstrate greater 3D consistency and layout plausibility compared to recent panoramic text-to-3D baselines.
Stefan Ainetter, Thomas Deixelberger, Edoardo A. Dominici +3
Huawei Technologies, Austria · Graz University of Technology, Austria · Huawei Technologies, Switzerland
The synthesis of immersive 3D scenes from text is rapidly maturing, driven by novel video generative models and feed-forward 3D reconstruction, with vast potential in AR/VR and world modeling. While panoramic images have proven effective for scene initialization, existing approaches suffer from a trade-off between visual fidelity and explorability: autoregressive expansion suffers from context drift, while panoramic video generation is limited to low resolution. We present Stepper, a unified framework for text-driven immersive 3D scene synthesis that circumvents these limitations via stepwise panoramic scene expansion. Stepper leverages a novel multi-view 360° diffusion model that enables consistent, high-resolution expansion, coupled with a geometry reconstruction pipeline that enforces geometric coherence. Trained on a new large-scale, multi-view panorama dataset, Stepper achieves state-of-the-art fidelity and structural consistency, outperforming prior approaches, thereby setting a new standard for immersive scene generation.
Felix Wimbauer, Fabian Manhardt, Michael Oechsle +4
Generating explorable 3D scenes from a single image is a highly challenging problem in 3D vision. Existing methods struggle to support free exploration, often producing severe geometric distortions and noisy artifacts when the viewpoint moves far from the original perspective. We introduce One2Scene, an effective framework that decomposes this ill-posed problem into three tractable sub-tasks to enable immersive explorable scene generation. We first use a panorama generator to produce anchor views from a single input image as initialization. Then, we lift these 2D anchors into an explicit 3D geometric scaffold via a generalizable, feed-forward Gaussian Splatting network. Instead of treating the panorama as a single image for reconstruction, we project it into multiple sparse anchor views and reformulate the reconstruction task as multi-view stereo matching, which allows us to leverage robust geometric priors learned from large-scale multi-view datasets. A bidirectional feature fusion module is used to enforce cross-view consistency, yielding an efficient and geometrically reliable scaffold. Finally, the scaffold serves as a strong prior for a novel view generator to produce photorealistic and geometrically accurate views at arbitrary cameras. By explicitly conditioning on a 3D-consistent scaffold to perform reconstruction, One2Scene works stably under large camera motions, supporting immersive scene exploration. Extensive experiments show that One2Scene substantially outperforms state-of-the-art methods in panorama depth estimation, feed-forward 360° reconstruction, and explorable 3D scene generation. Project page: https://one2scene5406.github.io/