Authors: Yanbo Wang, Siqi Luo, Siyu Zhang, Chenming Zhu, Wenzhao Zheng, Qian Yu, Juncheng Yan, Bin Fu, Tai Wang, Yu Qiao, Yihao Liu, Jiwen Lu
Organizations: Tsinghua University · Shanghai Artificial Intelligence Laboratory · Shanghai Jiao Tong University · Fudan University · The University of Hong Kong · University of California, Berkeley
Spatial intelligence has recently been assessed through generative tasks such as spatially grounded image editing, yet language-guided generation of spatially consistent views from partial multi-view observations remains underexplored. We introduce MindView, a benchmark for generative multi-view spatial intelligence with 1,500 samples from 513 scenes across three indoor datasets. Its seven tasks form two families: geometry-dominant tasks specify changes in camera position and orientation, while semantics-dominant tasks require inferring target viewpoints by grounding referred objects and interpreting spatial relations and reference-frame cues. Evaluation of eleven generators reveals substantial room for improvement on MindView. A paired context analysis further shows that all eleven generators obtain lower overall scores under the expanded context setting. These findings motivate us to investigate whether task-specific supervision can improve target-view generation and whether models can use additional multi-view evidence effectively. We therefore fine-tune BAGEL on the MindView training dataset, improving its overall score by 24.3 points and yielding gains across all seven MindView tasks. Beyond these generation gains, the fine-tuned model also yields higher scores on three spatial-understanding benchmarks without training on their annotations, suggesting that learning to generate spatially grounded views may benefit spatial understanding. To improve evidence utilization under expanded context, we propose MindCue, a training-free spatial evidence selector guided by a target-view hypothesis. MindCue reverses the drop in overall score under expanded context and outperforms the standard-context baseline. Overall, MindView provides a concrete testbed for generative multi-view spatial intelligence.
Figures & tables
Figure 1: Overview of MindView, showing seven task examples, benchmark statistics, six metrics, and representative results. Red A marks anchor views, and targets are real images.
Figure 2: Five-stage offline construction pipeline for MindView.
Figure 3: Distribution of the 1,500 MindView samples across source datasets and tasks.
Model
DINO-I
VLM Judge
Geometry
Overall
Semantic
Spatial
Quality
RScore
TScore
Open-Source Models
BAGEL
48.3
44.5
33.6
47.1
34.7
31.9
40.0
FLUX.2
52.2
49.5
39.4
80.4
35.5
38.6
49.3
Qwen-Image-Edit-2511 †
52.4
49.2
38.6
69.5
37.2
40.3
47.9
Emu3.5 †
44.9
41.0
31.2
67.2
36.0
49.9
45.0
Table 1: Quantitative comparison on MindView. Task-macro scores over 1,500 samples. and mark best and second-best across displayed models, respectively; † denotes tiled context views.
Figure 4: Qualitative comparison on Composition and Multi-Cue Viewpoint Inference.
Model
Geometry-Dominant Tasks
Semantics-Dominant Tasks
Trans.
Rot.
Comp.
ROG
TFR
OCR
MCVI
Open-Source Models
BAGEL
41.7
38.8
36.1
43.2
38.7
44.1
37.6
FLUX.2
53.0
47.0
44.1
50.5
49.9
53.9
46.6
Qwen-Image-Edit-2511 †
53.4
45.9
41.6
49.9
46.6
57.0
40.7
Emu3.5 †
55.7
43.6
40.7
42.5
45.0
52.3
35.5
Table 2: Task-wise comparison on MindView. Per-task Overall scores across all displayed models. Highlighting and † follow Table 1 ; abbreviations are defined in Section 3.1 .
Figure 5: Overall score trajectories across nested context settings.
Figure 6: Overview of the MindCue framework.
Model
Fixed C6 (6.0)
Fixed C10 (10.0)
CounterSel (8.6)
RandDrop (7.4)
MindCue (7.4)
ΔC6
ΔC10
BAGEL
40.0
39.0
37.1
39.6
41.1
+1.1
+2.1
BAGEL-SFT
64.4
63.2
62.6
64.1
65.1
+0.8
+1.9
Table 4: Task-macro Overall under context selection. Parentheses give mean view counts; and mark the best and second-best policies per model, respectively. ΔCk is MindCue’s score-point gain over Fixed Ck , computed before rounding.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Translation example from ARKitScenes.
Figure 8: Rotation example from ScanNet.
Figure 9: Composition example from ARKitScenes.
Figure 10: Relational Object Grounding example from ScanNet++.
Figure 11: Target-Facing Reorientation example from ScanNet.
Figure 12: Object-Centric Repositioning example from ScanNet.
Figure 13: Multi-Cue Viewpoint Inference example from ScanNet++.
Model
Semantic
Spatial
Quality
Human
VLM
Human
VLM
Human
VLM
GPT-Image-2
68.4
63.4
63.3
50.2
76.3
84.8
Nano Banana Pro
62.6
59.5
59.5
47.4
72.5
79.8
SenseNova-U1.5
57.6
54.1
51.5
43.5
66.8
70.3
BAGEL
45.5
45.4
43.0
33.4
44.1
48.6
Appendix
Table 5: Human and VLM-judge component scores on the shared 210-sample subset. Human scores are three-expert averages; all values are on a 0–100 scale, and higher is better.
Task
N
Real target (GT)
Contexts
ΔR ( ∘ )
ΔC (cm)
ΔD (%)
ΔR ( ∘ )
ΔL (cm)
ΔD (%)
Translation
30
0.85
4.96
1.09
1.09
6.02
1.39
Rotation
30
0.52
1.92
0.58
0.65
2.14
0.79
Composition
30
0.96
4.60
0.95
0.88
4.71
1.18
ROG
30
0.82
3.25
0.99
0.93
2.62
0.97
TFR
30
0.63
1.60
0.55
0.70
1.88
0.80
Appendix
Table 6: VGGT reference stability across generators. Only the generation changes; ten contexts and the real target stay fixed. Entries are sample medians after averaging six generator pairs per sample. ΔR : relative-rotation change; ΔC : aligned GT-center displacement; ΔL : context-pair distance change; ΔD : symmetric relative depth change. Distances are depth-calibrated; lower values indicate greater stability.
Figure 14: VGGT evaluation of GPT-Image-2 and BAGEL outputs. For the ScanNet TFR query “Turn to face the telephone,” the top row shows six generation contexts ( A : anchor); evaluation uses ten fixed references, the real target, and one generation. Reconstructions are independently depth-calibrated and rigidly aligned using only the ten reference camera centers, without scale fitting. All cameras are VGGT estimates; GT denotes the real target image. Dark red marks points contributed by BAGEL’s generation; all other scene points are gray.
Task
Target-view hypothesis
Geometry-Dominant Tasks
Trans.
Move from the reference position along its horizontal camera axes by the requested displacements, keeping its orientation and intrinsics.
Rot.
Keep the reference position and intrinsics. Rotate about world-up by the requested yaw, with left positive and right negative. A turn-around is 180∘ .
Comp.
Apply translations and rotations in query order from the reference pose. Each displacement uses the current horizontal camera axes, accounting for preceding rotations. Keep the reference intrinsics.
Semantics-Dominant Tasks
ROG
For each core-supported target centroid, place cameras at the requested distance along rays toward input-camera centers. Face the centroid and inherit source intrinsics; select the highest-scoring object–camera pair.
Appendix
Table 7: Task-specific target-view hypotheses in MindCue.
Configuration
Mean views
Mean (s)
BAGEL, C6
6
38.528
BAGEL, C10
10
50.625
BAGEL, selected views
7.30
42.476
MindCue + BAGEL, same views
7.30
43.913
MindCue component
Mean (s)
VGGT reconstruction
0.572
Appendix
Table 8: Runtime on a single A100 80GB. View counts refer to generator inputs, with identical images and order in the selected-view rows. Means are weighted and rounded independently.
Model
Trans.
Rot.
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
BAGEL
47.7
44.2
34.5
48.6
47.8
27.5
45.8
41.4
30.4
46.4
33.7
35.1
FLUX.2
53.2
49.8
40.9
85.3
59.0
29.7
46.7
42.4
33.9
78.5
31.9
48.7
Qwen-Image-Edit-2511 †
57.3
51.8
41.0
67.3
67.8
35.4
43.1
36.9
29.3
69.8
32.2
63.8
Emu3.5 †
52.0
46.0
38.0
69.6
80.1
48.4
38.7
32.0
24.5
63.3
30.3
72.5
DreamOmni2
56.3
51.3
41.8
61.8
74.2
41.1
39.8
31.8
25.5
60.8
23.7
63.7
Appendix
Table 9: Component scores on MindView for Translation and Rotation.
Model
Comp.
ROG
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
BAGEL
42.7
34.9
28.0
45.7
29.8
35.5
50.9
48.8
36.2
45.2
42.5
35.4
FLUX.2
43.6
40.0
33.5
80.8
32.9
33.6
54.1
52.0
40.5
81.3
33.2
42.2
Qwen-Image-Edit-2511 †
43.7
38.3
30.7
66.5
31.6
39.0
55.2
54.6
42.0
69.3
38.4
39.8
Emu3.5 †
38.0
31.3
26.9
63.8
26.8
57.3
49.0
45.5
32.8
58.5
29.9
39.1
DreamOmni2
36.0
28.5
24.4
64.0
19.5
42.6
35.4
29.4
23.6
70.5
11.2
32.3
Appendix
Table 10: Component scores on MindView for Composition and Relational Object Grounding.
Model
TFR
OCR
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
BAGEL
43.4
42.0
30.9
45.0
32.4
38.3
61.7
59.2
44.1
50.8
27.0
22.0
FLUX.2
50.5
48.5
38.3
76.1
32.7
53.4
70.0
68.4
52.9
77.6
28.4
25.8
Qwen-Image-Edit-2511 †
48.8
45.6
35.3
75.6
31.0
43.5
74.2
79.4
60.6
61.9
31.1
34.8
Emu3.5 †
40.6
40.8
29.8
73.3
25.7
59.4
62.7
61.2
43.8
68.5
37.7
40.0
DreamOmni2
31.0
25.6
19.2
67.7
13.2
53.4
62.7
60.2
45.6
63.4
23.8
31.5
Appendix
Table 11: Component scores on MindView for Target-Facing Reorientation and Object-Centric Repositioning.
Model
MCVI
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
BAGEL
45.8
41.2
31.4
47.9
30.0
29.3
FLUX.2
47.5
45.8
36.0
82.9
30.4
36.7
Qwen-Image-Edit-2511 †
44.6
37.9
31.4
76.2
28.8
25.5
Emu3.5 †
32.9
30.6
22.6
73.0
21.4
32.3
DreamOmni2
31.0
24.9
21.2
70.1
11.3
34.5
Appendix
Table 12: Component scores on MindView for Multi-Cue Viewpoint Inference.
Model
C3
C10
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
BAGEL
46.7
43.2
32.4
47.3
33.0
34.0
45.9
40.8
31.0
48.5
32.4
35.5
FLUX.2
49.6
46.4
37.7
85.9
33.7
41.5
51.5
48.7
38.1
69.3
32.5
38.0
Qwen-Image-Edit-2511 †
50.2
47.1
37.2
69.8
33.8
38.1
51.6
48.1
37.9
71.1
36.2
38.3
Emu3.5 †
44.5
38.9
28.6
59.5
34.8
46.8
44.9
39.9
30.6
68.3
33.3
49.1
DreamOmni2
42.7
37.6
30.8
76.9
25.1
36.1
32.2
27.7
22.1
59.0
25.9
51.8
Appendix
Table 13: Task-macro component scores under reduced and expanded contexts.
Model
Policy
DINO-I
Sem.
Spa.
Qual.
RScore
TScore
BAGEL
Fixed C6
48.3
44.5
33.6
47.1
34.7
31.9
Fixed C10
45.9
40.8
31.0
48.5
32.4
35.5
CounterSel
43.7
38.4
29.3
48.4
29.5
33.0
RandDrop
47.3
42.8
32.0
47.9
34.0
33.3
MindCue
49.2
44.9
33.9
47.6
36.6
34.7
BAGEL-SFT
Fixed C6
68.0
71.0
57.1
61.8
61.7
66.5
Appendix
Table 14: Task-macro component scores for the five context policies in Table 4 .
Figure 15: Qualitative comparison across all seven MindView tasks. Each row shows a query summary, six input contexts, the held-out real target (GT), and outputs from the selected generators named above the columns. A red border and A identify the anchor when one is required.
Current Large Multimodal Models (LMMs) struggle with spatial reasoning tasks requiring viewpoint-dependent understanding, largely because they are confined to a single, static observation. We propose Thinking with Novel Views (TwNV), a paradigm that integrates generative novel-view synthesis into the reasoning loop: a Reasoner LMM identifies spatial ambiguity, instructs a Painter to synthesize an alternative viewpoint, and re-examines the scene with the additional evidence. Through systematic experiments we address three research questions. (1) Instruction format: numerical camera-pose specifications yield more reliable view control than free-form language. (2) Generation fidelity: synthesized view quality is tightly coupled with downstream spatial accuracy. (3) Inference-time visual scaling: iterative multi-turn view refinement further improves performance, echoing recent scaling trends in language reasoning. Across four spatial subtask categories and four LMM architectures (both closed- and open-source), TwNV consistently improves accuracy by +1.3 to +3.9 pp, with the largest gains on viewpoint-sensitive subtasks. These results establish novel-view generation as a practical lever for advancing spatial intelligence of LMMs.
Spatial intelligence is essential for multimodal large language models, yet current benchmarks largely assess it only from an understanding perspective. We ask whether modern generative or unified multimodal models also possess generative spatial intelligence (GSI), the ability to respect and manipulate 3D spatial constraints during image generation, and whether such capability can be measured or improved. We introduce GSI-Bench, the first benchmark designed to quantify GSI through spatially grounded image editing. It consists of two complementary components: GSI-Real, a high-quality real-world dataset built via a 3D-prior-guided generation and filtering pipeline, and GSI-Syn, a large-scale synthetic benchmark with controllable spatial operations and fully automated labeling. Together with a unified evaluation protocol, GSI-Bench enables scalable, model-agnostic assessment of spatial compliance and editing fidelity. Experiments show that fine-tuning unified multimodal models on GSI-Syn yields substantial gains on both synthetic and real tasks and, strikingly, also improves downstream spatial understanding. This provides the first clear evidence that generative training can tangibly strengthen spatial reasoning, establishing a new pathway for advancing spatial intelligence in multimodal models.
Muzhi Zhu, Shunyao Jiang, Huanyi Zheng +9
Zhejiang University, State Key Laboratory of CAD & CG · Ant Group · Zhejiang University of Technology +1
Current Large Language Models have achieved Olympiad-level logic, yet Vision-Language Models paradoxically falter on elementary spatial tasks like block counting. This capability mismatch reveals a critical spatial intelligence gap,'' where models fail to construct coherent 3D mental representations from 2D observations. We uncover this gap via diagnostic analyses showing the bottleneck is a missing view-consistent spatial interface rather than insufficient visual features or weak reasoning. To bridge this, we introduce **3ViewSense**, a framework that grounds spatial reasoning in Orthographic Views. Drawing on engineering cognition, we propose a Simulate-and-Reason'' mechanism that decomposes complex scenes into canonical orthographic projections to resolve geometric ambiguities. By aligning egocentric perceptions with these allocentric references, our method facilitates explicit mental rotation and reconstruction. Empirical results on spatial reasoning benchmarks demonstrate that our method significantly outperforms existing baselines, with consistent gains on occlusion-heavy counting and view-consistent spatial reasoning. The framework also improves the stability and consistency of spatial descriptions, offering a scalable path toward stronger spatial intelligence in multimodal systems.~\footnote{https://github.com/Jasaxion/3ViewSense}
Shaoxiong Zhan, Yanlin Lai, Zheng Liu +6
Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Pengcheng Laboratory, Shenzhen, China · School of Software Engineering, Chongqing University, Chongqing, China