T3lescope: Arbitrary-Resolution High-Fidelity Generative Surface Reconstruction from Images

Atsuhiro Noguchi; Tianhan Xu; Yiming Liang; Yuta Kikuchi; Masahiro Ishiyama; Shintaro Takagi; Hitoshi Murai; Eiichi Matsumoto

cs.CVOct 2, 2026

T3lescope: Arbitrary-Resolution High-Fidelity Generative Surface Reconstruction from Images

Authors: Atsuhiro Noguchi, Tianhan Xu, Yiming Liang, Yuta Kikuchi, Masahiro Ishiyama, Shintaro Takagi, Hitoshi Murai, Eiichi Matsumoto

Organizations: Preferred Networks, Inc. · The University of Tokyo

Abstract

We reconstruct high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization, across scales ranging from single objects to large outdoor scenes. Per-scene optimization methods lack the learned 3D prior needed when observations are sparse or surfaces are glossy or transparent. Existing generative methods leverage such priors to complete geometry in sparsely observed regions, but typically operate at a fixed resolution over a limited spatial extent, trading spatial coverage against detail. Reconstructing a large scene therefore often requires partitioning it into independently processed overlapping local regions, making it difficult to maintain global geometric consistency. To address these issues, we propose T3lescope, which applies a single fixed-resolution generator across scene scales in an inference-time coarse-to-fine cascade. A coarse level establishes the scene layout, and finer levels perturb and denoise geometry inherited from the coarser level within progressively finer spatial cells to recover surface detail. The model is trained on individual cells at multiple scales and shares its weights across all levels, so no hierarchy is fixed during training, and the number of levels, cell scales, and cell locations are determined at inference time. On indoor, outdoor, and city-scale scenes, T3lescope outperforms feed-forward and generative baselines, matches or surpasses per-scene optimization, and recovers fine structures as well as glossy and transparent surfaces. These results show that our method generalizes across diverse scenes, view counts, and image resolutions. Project page: https://pfnet-research.github.io/t3lescope/

Figures & tables

Panel of Figure 1 — Figure 1: High-fidelity mesh reconstruction without per-scene optimization. Our method takes multi-view images as input and directly outputs a high-resolution mesh with mesh flow models. We adopt a coarse-to-fine cascade: the entire scene is generated at a coarse resolution first, and regions of interest are then upscaled. The cascade layout and the mesh resolution are chosen at inference time. Colors mark the cascade level each surface was generated at.

Panel of Figure 2 — Figure 2: Method overview. L $n$ denotes level $n$ . (a) DA3 points place the L0 cells and select their views. Each finer level halves the cell size and refines the mesh of the level above. (b) Each image is encoded once as a tile pyramid, and from each view a cell takes the tiles at the pyramid level that matches its voxel spacing. (c) All cells at all levels share the SS and mesh generators, which attend to these tiles. At finer levels, the SS stage starts from the noised parent (SDEdit).

Panel of Figure 3 — Figure 3: Qualitative results. (a) ScanNet++ 7b6477cb95 and (b) Tanks and Temples truck . (c) Real-world captures (ours): mesh, one input view, and a close-up of the boxed region.

w/ per-scene optimization
	32 Views							256 Views
	2D				3D			2D				3D
Method	MAE $\downarrow$	AbsRel $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$	MAE $\downarrow$	AbsRel $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$
2DGS	0.1426	0.1141	38.376	0.9058	0.0723	0.4837	0.7181	0.0929	0.0846	32.208	0.9747	0.0513	0.5345	0.8006
MILo	0.2523	0.1780	42.156	0.9848	0.1127	0.4241	0.6910	0.1406	0.1226	33.612	0.9998	0.0817	0.5508	0.7985
GaussianWrapping	0.3467	0.2762	35.974	0.9648	0.2830	0.4413	0.7505	0.3197	0.3071	30.118	0.9761	0.2136	0.6075	0.8113

Table 1: Quantitative comparison on ScanNet++. Mean over the 25 largest scenes. 2D metrics compare depth and normals rendered at held-out views, and 3D metrics compare the mesh surfaces visible from those views. L0 to L2 are our cascade levels. The top three results are highlighted.

w/ per-scene optimization
	Small, 16 Views		Small, 256 Views		Large, 16 Views		Large, 256 Views
Method	F@ $\tau$ $\uparrow$	CD $/\tau$ $\downarrow$	F@ $\tau$ $\uparrow$	CD $/\tau$ $\downarrow$	F@ $\tau$ $\uparrow$	CD $/\tau$ $\downarrow$	F@ $\tau$ $\uparrow$	CD $/\tau$ $\downarrow$
2DGS	0.355	9.42	0.448	9.64	0.101	44.95	0.233	28.08
MILo	0.346	14.79	0.527	11.63	0.053	54.04	0.268	51.29
GaussianWrapping	0.345	8.80	0.650	5.54	0.094	37.61	0.370	42.51
w/o per-scene optimization

Table 2: Quantitative comparison on Tanks and Temples and the city-block dataset. (a) Mean over the three small (Caterpillar, Ignatius, Truck) and three large (Barn, Courthouse, Meetingroom) scenes with 2K input images. F@

\tau

and CD

/\tau

use the per-scene threshold

\tau

. (b) Mean over three blocks, evaluated against the BIM ground truth. L

x

denotes level

x

, with its cell size in parentheses. The top three results are highlighted.

Panel of Figure 4 — Figure 4: Qualitative comparison on the city-block dataset. (a) Three of the 256 input views. (b) Each method’s reconstruction, with zoomed crops of ours and the ground truth to compare detail.

Panel of Figure 5 — Figure 5: Effect of parent initialization. Every finer cascade level starts from a latent noised to $t_{s}$ , where $t_{s}=1$ generates from pure noise. (a) Mean over the 25 largest ScanNet++ scenes, with the best results in bold. (b) One block of the city-block dataset.

Panel of Figure 6 — Figure 6: Effect of image conditioning. The baseline uses the GenRecon-style conditioner, our voxel-queried attention replaces it, and the ConvNeXt branch is then added, all at the finest cascade level. (a) Mean over the 25 largest ScanNet++ scenes, with the best results in bold. (b) A multi-view capture of people.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

32 Views
	Small (25)						Large (25)						All (50)
	Depth				Normal	Coverage	Depth				Normal	Coverage	Depth				Normal	Coverage
Method	MAE $\downarrow$	RMSE $\downarrow$	AbsRel $\downarrow$	SqRel $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	AbsRel $\downarrow$	SqRel $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	AbsRel $\downarrow$	SqRel $\downarrow$	Ang $\downarrow$	Comp $\uparrow$
2DGS	0.1261	0.2493	0.1329	0.0591	40.149	0.9542	0.1426	0.3412	0.1141	0.0799	38.376	0.9058	0.1344	0.2953	0.1235	0.0695	39.262	0.9300
MILo	0.1748	0.3699	0.1855	0.1530	44.731	0.9913	0.2523	0.6171	0.1780	0.2623	42.156	0.9848	0.2136	0.4935	0.1817	0.2077	43.444	0.9881
GaussianWrapping	0.4688	1.4378	0.6228	6.9407	40.150	0.9365	0.3467	1.3934	0.2762	2.8323	35.974	0.9648	0.4077	1.4156	0.4495	4.8865	38.062	0.9507

Appendix

Table 3: Quantitative comparison on ScanNet++ (2D metrics). Mean over the 25 smallest, the 25 largest, and all 50 scenes. Depth and normals are rendered from the meshes at held-out views. The top three results are highlighted.

32 Views
	Small (25)							Large (25)							All (50)
	Whole mesh			Virtual scan				Whole mesh			Virtual scan				Whole mesh			Virtual scan
Method	CD $\downarrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	F@10 $\uparrow$	NC $\uparrow$
2DGS	0.0682	0.7878	0.6587	0.0741	0.4453	0.7698	0.7065	0.1017	0.7748	0.6399	0.0723	0.4837	0.8070	0.7181	0.0850	0.7813	0.6493	0.0732	0.4645	0.7884	0.7123
MILo	0.0605	0.8043	0.6696	0.1013	0.4264	0.7381	0.6779	0.0910	0.7708	0.6532	0.1127	0.4241	0.7382	0.6910	0.0757	0.7876	0.6614	0.1070	0.4252	0.7382	0.6845
GaussianWrapping	0.0633	0.8085	0.6728	0.4135	0.4433	0.7501	0.6723	0.0766	0.7909	0.6963	0.2830	0.4413	0.8073	0.7505	0.0699	0.7997	0.6846	0.3483	0.4423	0.7787	0.7114

Appendix

Table 4: Quantitative comparison on ScanNet++ (3D metrics). Mean over the 25 smallest, the 25 largest, and all 50 scenes. We report the whole-mesh protocol of GenRecon ( Schmid et al., 2026 ) and the virtual-scan protocol of GaussianWrapping ( Gomez et al., 2026 ) . The top three results are highlighted.

Panel of Figure 7 — Figure 7: Qualitative comparison on ScanNet++ with 32 input views. All meshes are rendered from a held-out view of each scene.

Panel of Figure 8 — Figure 8: Qualitative comparison on ScanNet++ with 32 input views. All meshes are rendered from a held-out view of each scene.

Panel of Figure 9 — Figure 9: Qualitative comparison on ScanNet++ with 256 input views. All meshes are rendered from a held-out view of each scene.

Panel of Figure 10 — Figure 10: Qualitative comparison on ScanNet++ with 256 input views. All meshes are rendered from a held-out view of each scene.

Panel of Figure 11 — Figure 11: Qualitative comparison with GenRecon on ScanNet++ in the 256-view setting. Both meshes are rendered from the same held-out viewpoint, with close-ups of the boxed regions below.

16 Views
	Barn		Caterpillar		Courthouse		Ignatius		Meetingroom		Truck
Method	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$
2DGS	0.1680	0.4530	0.2467	0.0667	0.0773	1.1360	0.3975	0.0260	0.0566	0.4412	0.4213	0.0314
MILo	0.0622	0.5128	0.1522	0.1275	0.0235	2.3359	0.5351	0.0210	0.0740	0.1741	0.3510	0.0594
GaussianWrapping	0.1520	0.1269	0.2098	0.0777	0.0457	2.0951	0.5061	0.0081	0.0840	0.1635	0.3192	0.0409
DA3	0.1123	0.2928	0.1088	0.0876	0.0418	1.3350	0.1417	0.0349	0.1878	0.2334	0.1926	0.0414

Appendix

Table 5: Per-scene quantitative comparison on Tanks and Temples. F@

\tau

uses the official per-scene threshold

\tau

, and CD is in meters. Rows marked 4K use the 4K input images instead of the 2K ones. The top three results are highlighted.

Panel of Figure 12 — Figure 12: Qualitative comparison on Tanks and Temples with 16 input views. All meshes of a scene are rendered from the same viewpoint.

Panel of Figure 13 — Figure 13: Qualitative comparison on Tanks and Temples with 16 input views. All meshes of a scene are rendered from the same viewpoint.

Panel of Figure 14 — Figure 14: Qualitative comparison on Tanks and Temples with 256 input views. All meshes of a scene are rendered from the same viewpoint.

Panel of Figure 15 — Figure 15: Qualitative comparison on Tanks and Temples with 256 input views. All meshes of a scene are rendered from the same viewpoint.

Panel of Figure 16 — Figure 16: Qualitative comparison with GaussianWrapping on Tanks and Temples with 256 input views. Both meshes are rendered from the same viewpoint, with close-ups of the boxed regions below.

w/ per-scene optimization
	Scene 025			Scene 033			Scene 034			Mean
Method	CD $\downarrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@10 $\uparrow$	NC $\uparrow$	CD $\downarrow$	F@10 $\uparrow$	NC $\uparrow$
2DGS	4.701	0.101	0.231	5.011	0.088	0.157	4.863	0.039	0.134	4.858	0.076	0.174
CityGaussianV2	6.278	0.119	0.216	4.262	0.123	0.217	4.926	0.094	0.252	5.155	0.112	0.228
w/o per-scene optimization
DA3	4.514	0.014	0.039	4.754	0.010	0.027	6.285	0.022	0.079	5.184	0.015	0.048

Appendix

Table 6: Per-block quantitative comparison on the city-block dataset. All methods use 256 input views. The cell size of each level of our method is shown in parentheses. CD is in meters. The top three results are highlighted.

Panel of Figure 17 — Figure 17: Qualitative comparison with GaussianWrapping on real-world captures. Both meshes are rendered from the same viewpoint, with close-ups of the boxed regions below.

Panel of Figure 18 — Figure 18: Failure cases with distant objects. Input images containing distant objects, and our meshes from 256 input views at 4K, rendered from the same cameras and from the top. The distant objects are not reconstructed well.

Panel of Figure 19 — Figure 19: Examples of our procedural scene generation. The first two images show a synthetic multi-room house, and the last two show procedural grass scenes.

32 Views
	Small (25)						Large (25)
	2D			3D virtual scan			2D			3D virtual scan
$t_{s}$	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$
0.8	0.0324	15.31	0.9990	0.0190	0.6718	0.9296	0.0565	16.33	0.9962	0.0212	0.7155	0.9247
0.9	0.0715	16.97	0.9909	0.0394	0.6730	0.8945	0.1450	19.16	0.9773	0.0621	0.6779	0.8580
1.0	0.0998	18.31	0.9823	0.0548	0.7255	0.8646	0.2170	22.15	0.9535	0.1051	0.6306	0.7881

Appendix

Table 7: Effect of parent initialization on ScanNet++. Every finer cascade level starts from a latent noised to

t_{s}

, where

t_{s}=1

generates from pure noise. Mean over the 25 smallest and the 25 largest scenes, with the best results in bold.

	Barn		Caterpillar		Courthouse		Ignatius		Meetingroom		Truck		Mean
$t_{s}$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $/\tau$ $\downarrow$
0.8	0.2080	0.0895	0.1949	0.0637	0.1159	0.5905	0.4704	0.0055	0.3084	0.0476	0.3563	0.0311	0.2756	9.6855
0.9	0.2336	0.0831	0.1960	0.0575	0.1169	0.6717	0.4436	0.0058	0.3071	0.0884	0.3850	0.0304	0.2804	10.5876
1.0	0.2304	0.1030	0.2057	0.0555	0.1084	0.6974	0.4584	0.0084	0.2654	0.1313	0.4300	0.0343	0.2830	12.0138

Appendix

Table 8: Effect of parent initialization on Tanks and Temples with 16 input views. Every finer cascade level starts from a latent noised to

t_{s}

, where

t_{s}=1

generates from pure noise. Inputs are the 4K images. F@

\tau

and CD

/\tau

use the official per-scene threshold

\tau

, and CD is in meters. The best results are in bold and the second best underlined.

Panel of Figure 20 — Figure 20: Effect of parent initialization on Tanks and Temples. Meshes at the finest cascade level from 16 input views at 4K, rendered from the same viewpoint per scene.

L0
	Barn		Caterpillar		Courthouse		Ignatius		Meetingroom		Truck		Mean
Conditioning	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $/\tau$ $\downarrow$
Baseline	0.1781	0.0691	0.1124	0.0839	0.0836	0.4428	0.2092	0.0113	0.3383	0.0360	0.2472	0.0344	0.1948	9.2747
Voxel attn.	0.1704	0.0645	0.1223	0.0814	0.0875	0.3684	0.2381	0.0100	0.3531	0.0359	0.2482	0.0347	0.2033	8.5584
+ ConvNeXt	0.1463	0.0787	0.1328	0.0805	0.0965	0.3619	0.3223	0.0073	0.3561	0.0349	0.2955	0.0317	0.2249	8.4522
L1

Appendix

Table 9: Effect of image conditioning on Tanks and Temples with 32 input views. The baseline uses the GenRecon-style conditioner, our voxel-queried attention replaces it, and the ConvNeXt branch is then added. Inputs are the 4K images. F@

\tau

and CD

/\tau

use the official per-scene threshold

\tau

, and CD is in meters. The best results are in bold and the second best underlined.

Panel of Figure 21 — Figure 21: Effect of image conditioning on Tanks and Temples. Meshes at the finest cascade level from the 4K images, all rendered from the same viewpoint per scene. The baseline uses the GenRecon-style conditioner, our voxel-queried attention replaces it, and the ConvNeXt branch is then added.

L0
	Small (25)						Large (25)						All (50)
	2D			3D virtual scan			2D			3D virtual scan			2D			3D virtual scan
Tile level	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$
Coarse	0.0456	18.85	0.9984	0.0241	0.5633	0.9073	0.0718	19.14	0.9986	0.0231	0.6349	0.9097	0.0587	19.00	0.9985	0.0236	0.5991	0.9085
Adaptive (ours)	0.0438	18.74	0.9986	0.0230	0.5816	0.9110	0.0714	19.21	0.9984	0.0235	0.6336	0.9090	0.0576	18.98	0.9985	0.0233	0.6076	0.9100
Fine	0.0578	20.01	0.9967	0.0313	0.4919	0.8894	0.0892	19.89	0.9977	0.0275	0.5849	0.8966	0.0735	19.95	0.9972	0.0294	0.5384	0.8930

Appendix

Table 10: Effect of tile pyramid level selection on ScanNet++ with 32 input views. Coarse fixes every cell to the coarsest pyramid level whose image height is at least 512 pixels, and Fine to the full-resolution level; ours matches the level to the cell’s voxel spacing. Mean over the 25 smallest, the 25 largest, and all 50 scenes, with the best results in bold.

L0
	Barn		Caterpillar		Courthouse		Ignatius		Meetingroom		Truck		Mean
Tile level	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $\downarrow$	F@ $\tau$ $\uparrow$	CD $/\tau$ $\downarrow$
Coarse	0.1554	0.0740	0.1273	0.0730	0.0908	1.0964	0.2996	0.0086	0.3052	0.0365	0.3289	0.0296	0.2178	13.0521
Adaptive (ours)	0.1463	0.0787	0.1328	0.0805	0.0965	0.3619	0.3223	0.0073	0.3561	0.0349	0.2955	0.0317	0.2249	8.4522
Fine	0.1675	0.0619	0.1176	0.0648	0.0921	0.7501	0.2672	0.0091	0.3274	0.0390	0.2671	0.0348	0.2065	10.5084
L1

Appendix

Table 11: Effect of tile pyramid level selection on Tanks and Temples with 32 input views. Coarse fixes every cell to the coarsest pyramid level whose image height is at least 512 pixels, and Fine to the full-resolution level; ours matches the level to the cell’s voxel spacing. Inputs are the 4K images. F@

\tau

and CD

/\tau

use the official per-scene threshold

\tau

, and CD is in meters. The best results are in bold.

32 Views, L0
	Small (25)						Large (25)						All (50)
	2D			3D virtual scan			2D			3D virtual scan			2D			3D virtual scan
Window	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$	MAE $\downarrow$	Ang $\downarrow$	Comp $\uparrow$	CD $\downarrow$	F@2 $\uparrow$	NC $\uparrow$
$w=6$	0.0437	18.65	0.9985	0.0230	0.5845	0.9111	0.0723	19.27	0.9984	0.0238	0.6316	0.9090	0.0580	18.96	0.9984	0.0234	0.6080	0.9100
$w=2$ (ours)	0.0438	18.74	0.9986	0.0230	0.5816	0.9110	0.0714	19.21	0.9984	0.0235	0.6336	0.9090	0.0576	18.98	0.9985	0.0233	0.6076	0.9100
32 Views, L1

Appendix

Table 12: Effect of the attention window size

w

on ScanNet++. Each voxel attends to the

w\times w

DINO tokens around its projection in every view. Mean over the 25 smallest, the 25 largest, and all 50 scenes, with the best results in bold.

Panel of Figure 22 — Figure 22: Effect of the synthetic training data. Meshes at the finest cascade level with 32 input views, from the model trained with our synthetic multi-room houses and grass scenes (Full) and from one trained on SAGE-10K and 3D-FRONT alone (w/o house & grass). Without the synthetic data, walls of the multi-room ScanNet++ scene disappear, and the grass around the loader becomes a flat surface.

	Count		Time (s)						Memory (GiB)
			View sel.	Parent
Level	#Cells	#Tiles	+ tiles	transfer	SS	Shape	Decode	Total	Peak
L0 (4 m)	31	329	23	–	38	48	2.4	112	36
L1 (2 m)	470	473	26	21	466	394	7.6	914	57
L2 (1 m)	2,490	591	27	60	2,428	1,740	29	4,285	79
Total			76	82	2,932	2,181	39	5,310 (89 min)	79

Appendix

Table 13: Inference time per cascade level on ScanNet++. Wall-clock seconds on one NVIDIA H200 GPU with 32 input views, averaged over the 5 large scenes. View sel. + tiles includes DA3 depth, per-cell view selection, tile cropping and tile feature encoding (DINOv3 and ConvNeXt); parent transfer decodes the parent level to initialize the child SS. Model loading is excluded. Peak is the peak allocated GPU memory.

Large-Scale 3D Reconstruction

See all 17 papers

May 22, 2026

GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction

We introduce a new approach to high-fidelity 3D scene reconstruction from multi-view RGB images that tightly couples reconstruction with a strong generative 3D prior. We cast scene reconstruction as conditional 3D generation over a set of spatially-localized, overlapping chunks that together tile the scene, scaling generation to large scene extents. Crucially, we inherit the fidelity and completeness of state-of-the-art generative shape models -- we use Trellis.2 as an example -- which we generalize to the scene level. To this end, we propose a projection-based conditioning mechanism that lifts posed multi-view image features into a coherent 3D representation aligned with the generative model, independent of view ordering and spatially anchored to the scene, yielding high-fidelity, multi-view consistent generated geometry. This enables lifting the strong object-level prior of Trellis.2 to multi-view, scene-scale generation, producing faithful, editable PBR mesh reconstructions of indoor environments. As a result, we obtain high-fidelity results that outperform cutting-edge reconstruction methods by 16%.

Katharina Schmid, Nicolas von Lützow, Jozef Hladký +2

Technical University of Munich · Computing Systems Lab, Huawei Technologies, Switzerland

3D Reconstruction Multi-View 3D Reconstruction 3D Mesh Reconstruction Indoor Scene Reconstruction Conditional Generative Modeling

Ask PDF

Oct 13, 2026

WorldRecon: Generative Multi-View Mesh Reconstruction for Large-Scale 3D World

Reconstructing complete, high-quality meshes of large-scale scenes from unposed multi-view images requires faithful geometric grounding while plausibly completing surfaces that are unobserved or only sparsely captured. We present WorldRecon, a generative multi-view reconstruction framework for recovering complete, large-scale 3D scene meshes from unposed RGB images. WorldRecon uses a coarse, noisy global point cloud and camera poses estimated from the input images by a reconstruction model as a global scaffold for whole-scene generation and partitions it into overlapping spatial blocks. To robustly leverage these imperfect geometric priors, we equip the native 3D generative model TRELLIS.2 with a dedicated reconstruction-based conditioning architecture that generates each block from its local point-cloud crop and block-relevant image observations, producing sparse structure, detailed geometry, and PBR materials. To maintain scene-level coherence, we introduce neighbour-aware latent inpainting and a causal world-generation procedure that propagates structural, geometric, and material context across overlapping blocks. Together, these designs scale multi-view generative reconstruction to large-scale scenes, yielding coherent, unified meshes. Extensive experiments demonstrate that WorldRecon achieves state-of-the-art reconstruction performance, reducing Chamfer distance by 14.9% and 43.1% while improving F-score by 17.3% and 12.0% over the strongest baselines on synthetic and real-world indoor scenes, respectively, thereby demonstrating faithful and complete generative reconstruction as well as strong sim-to-real generalization. Project page: https://kszpxxzmc.github.io/WorldRecon-website/.

Mengchen Zhang, Yang Li, Chunshi Wang +2

Zhejiang University · Tencent Hunyuan · Fudan University

3D Reconstruction 3D Mesh Reconstruction Multi-View 3D Reconstruction Large-Scale 3D Reconstruction Indoor Scene Reconstruction 3D Diffusion Models

Ask PDF

Aug 25, 2026

SceneReGen: Generative Reconstruction of 3D Scenes from a Single Image

Single-image 3D scene reconstruction must complete partially observed objects and place them coherently in a shared observation-aligned scene frame. Object-level generative priors offer strong completion ability, but their centered, scale-normalized outputs are typically expressed in an object frame, creating a fundamental representation gap between object generation and scene reconstruction. We introduce SceneReGen, a generative reconstruction framework that reinterprets scene reconstruction as the generation and assembly of complete object assets in a shared observation-aligned scene frame. SceneReGen addresses the generation-reconstruction gap through selective pose factorization: each object's observed orientation is encoded directly in the generated mesh, while translation and scale are estimated from instance-level and global scene evidence. Given a scene image and instance masks, a geometry encoder extracts dense cues; learnable shape queries condition a pretrained DiT-based 3D generator to produce complete meshes in their observed orientations, while position queries fuse object and scene features to assemble them in the shared frame. On the 3D-FUTURE evaluation subset, SceneReGen achieves the best scene-level CD, scene-level F-Score, and 3D bounding-box IoU among the evaluated methods, ties the best object-level CD, and ranks second in object-level F-Score. Qualitative outputs in autonomous-driving and embodied-AI scenes further illustrate the potential of asset-centric reconstruction beyond indoor furniture.

Zefan Tian, Yuteng Ye, Yiheng Zhang +5

Huawei · Northwestern Polytechnical University

3D Reconstruction Single-View 3D Reconstruction Generative Modeling 3D Asset Generation

Ask PDF

Abstract

Large-Scale 3D Reconstruction

Explore similar work

GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction

WorldRecon: Generative Multi-View Mesh Reconstruction for Large-Scale 3D World

SceneReGen: Generative Reconstruction of 3D Scenes from a Single Image