acceptodds
Under review as a conference paper at ICLR 2027

MindView: Benchmarking Generative Multi-View Spatial Intelligence

Abstract

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. Our findings further highlight task learning and query-guided evidence selection as complementary strategies for improving performance in this setting.

Primary area:
datasets and benchmarks
Keywords:
generative spatial intelligence, multi-view generation, viewpoint reasoning, multimodal generative models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Lsi6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MindView evaluates whether image generators can produce a requested viewpoint of an indoor scene from multiple partial observations and a language instruction. It contains 1,500 samples from 513 scenes across three datasets and covers seven tasks, ranging from camera translation and rotation to viewpoint inference based on objects and spatial relations. The authors evaluate 11 existing generators and release 20,000 training examples. Fine-tuning BAGEL raises its Overall score from 40.0 to 64.4. They also propose MindCue, a training-free view selector designed to address the performance decline observed when more input views are provided.

Critical strengths
  1. The task design is valuable. The seven tasks cover both explicit camera motions and viewpoint inference requiring object grounding and spatial relations. Composition or multi-cue inference is the hardest task for each evaluated generator, revealing specific limitations beyond aggregate scores (Table 2).

  2. Benchmark construction and review are careful. The authors use captured target images and filter candidates using camera poses, visibility, and scene annotations. Three reviewers check ambiguity and context adequacy, and the training and evaluation sets have disjoint scenes (§3.2, Appendices A and C.1).

  3. The evaluation is broad, and the main adaptation result is clear. The paper reports separate measures of content, viewpoint, visual quality, and geometric consistency, with partial validation against human ratings (§3.3, Appendix B.4). BAGEL-SFT improves over BAGEL on all seven tasks and all 42 task–metric combinations (Table 2, Appendix Tables 9–12).

Critical weaknesses
  1. A single target image may penalize valid responses to semantic instructions. ROG’s distance, centering, and visibility constraints may not uniquely determine a camera position in continuous space; MCVI’s uniqueness checks mainly cover recorded frames (Appendix A.2 and Limitations). Yet the VLM judge compares the generation with one target image without seeing the instruction, while the geometric metrics measure pose agreement with that target (§3.3, Appendices B.2–B.3). A different view that satisfies the instruction could therefore receive a lower score. The ambiguity checks mitigate this risk, but the paper does not quantify it.

  2. The contribution relative to existing generative spatial benchmarks could be explained more concretely. GSI-Bench[1] also evaluates spatially constrained image generation or editing and studies transfer from generative training to spatial understanding. MindView differs meaningfully through multi-view observations, camera-viewpoint inference, and captured target views. However, §1–2 mainly describe differences in task setup without showing which failure modes or conclusions specifically require these distinctions. A task-level comparison would clarify the benchmark’s added value.

[1] Zhu, M., Jiang, S., Zheng, H., et al. Exploring Spatial Intelligence from a Generative Perspective. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026.

  1. The cause of the performance decline with additional views remains unclear. Expanding the context from six to ten views increases target-content coverage from 88.4% to 92.0%, yet lowers Overall for all 11 existing generators; their Semantic and Spatial scores also decline (§4.3, Figure 5, Appendix Table 13). MindCue’s advantage over count-matched random selection suggests that view relevance matters, but does not distinguish irrelevant-view interference from difficulty integrating more images or effects of input order and format. The decline persists for BAGEL-SFT after task-specific fine-tuning, making this distinction especially relevant. The evidence establishes that more views do not necessarily help, but does not isolate why.

  2. The spatial-understanding transfer results lack an absolute performance reference. Table 3 compares only BAGEL and BAGEL-SFT. This paired comparison establishes the effect of fine-tuning, but representative understanding or unified models evaluated under the same input and scoring protocols would help place BAGEL-SFT’s performance in context. Moreover, the gains on MMSI-Bench and MindCube-Tiny are only 1.2 and 1.8 percentage points, respectively. Without a control matched for training volume, the source of these gains remains uncertain (§4.4, Table 3).

Questions for authors
  1. For ROG and MCVI, could the authors use the original instructions to manually assess generated views that differ from the target image but remain valid, and quantify how often single-target evaluation penalizes them?
  2. Compared with GSI-Bench and related benchmarks, which failure modes does MindView’s multi-view setting reveal that single-image spatial editing tasks cannot? Could the authors provide task-level examples?
  3. When expanding from six to ten views, how much of the decline is attributable to low-relevance additions versus difficulty integrating more images? Could the authors compare relevance-based selection with changes in view order and simpler count-matched selection policies?
  4. Could the authors evaluate representative spatial-understanding or unified models using Table 3’s protocols, and add a training-volume-matched control for the transfer results?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Lsi6 (Part 1/3)

W1 / Q1. Single-GT Evaluation: Rationale and Robustness

We agree that valid alternative views may receive lower scores under a single GT, as noted in our Limitations. We examine the rationale for fixed-GT LLM scoring through comparison with the Instruction-aware LLM Judge and human ratings, the robustness of model comparisons through reference GT replacement, and the remaining penalties on valid alternatives through an oracle-view test.

  1. Comparison with the Instruction-aware LLM Judge. We evaluate 300 outputs from five models on 30 ROG and 30 MCVI queries, with independent ratings from the three experts in the original Table 5. Both automatic judges use the same Qwen3-VL-30B-A3B-Instruct model. Our judge sees only the GT and candidate. The Instruction-aware LLM Judge and experts share the same criteria and see the instruction, six input images, a GT example, and candidate, explicitly accepting valid alternatives. Scores are Spatial / Semantic, and correlations use five model means.
Model / StatisticHumanInstruction-aware LLM JudgeOriginal LLM Judge (Ours)
BAGEL39.1 / 40.281.7 / 85.028.3 / 39.7
BAGEL-SFT77.5 / 74.992.0 / 93.749.3 / 61.3
GPT-Image-272.6 / 73.599.7 / 99.742.3 / 54.3
Nano Banana Pro65.0 / 62.695.7 / 95.742.0 / 50.7
SenseNova-U1.551.7 / 51.394.3 / 95.339.3 / 49.0
Pearson r vs. Human—0.7254 / 0.75570.9394 / 0.9386
Spearman ρ vs. Human—0.40 / 0.401.00 / 1.00

The Instruction-aware LLM Judge gives consistently high scores (GPT-Image-2: 99.7 on both metrics), possibly rewarding plausible layouts without fully understanding the input scene or verifying all spatial constraints, both of which can be challenging for LLMs. Our judge agrees more closely with human scores (Pearson r: 0.9394 / 0.9386 vs. 0.7254 / 0.7557) and rankings (Spearman ρ: 1.00 / 1.00 vs. 0.40 / 0.40). This supports a fixed GT as a practical scoring reference for model-level comparisons.

  1. Reference GT replacement. We next test whether model comparisons depend on the choice of a valid reference. Construction records provide 65 ROG groups and 143 MCVI groups with saved captured alternatives satisfying the instruction and image-clarity requirements. For each sample, we select the candidate with the largest combined pose difference from the original GT. We keep all 12 models’ outputs and original Quality scores fixed and recompute the five reference-dependent metrics. Overall averages six metrics, and parentheses show Alternate minus Original.
ModelROGMCVIModelROGMCVI
BAGEL43.5 → 43.6 (+0.1)36.2 → 35.9 (−0.3)SenseNova-U1.558.3 → 58.9 (+0.6)51.3 → 50.9 (−0.4)
FLUX.251.9 → 51.8 (−0.1)46.3 → 46.3 (0.0)Nano Banana54.6 → 54.8 (+0.2)47.5 → 47.4 (−0.1)
Qwen-Image-Edit-251150.0 → 50.5 (+0.5)40.9 → 40.2 (−0.7)Nano Banana Pro58.3 → 58.2 (−0.1)49.4 → 49.3 (−0.1)
Emu3.542.9 → 42.5 (−0.4)35.4 → 35.6 (+0.2)GPT-Image-1.560.3 → 60.5 (+0.2)54.0 → 54.0 (0.0)
DreamOmni234.0 → 33.8 (−0.2)31.6 → 31.4 (−0.2)GPT-Image-261.6 → 62.0 (+0.4)58.2 → 57.9 (−0.3)
SenseNova-U149.0 → 48.3 (−0.7)43.4 → 42.7 (−0.7)BAGEL-SFT61.1 → 60.7 (−0.4)57.0 → 57.0 (0.0)

Task means change by at most 0.8 points, and rankings remain highly consistent (Spearman ρ: 0.993 / 1.00 for ROG / MCVI). Only Nano Banana Pro and SenseNova-U1.5 swap ranks 4 and 5 in ROG. The main model comparisons are stable to the choice of a valid reference GT.

  1. Scoring valid alternative views. We now fix the official GT and score the same farthest valid captured views as oracle outputs to quantify the mean penalty on valid alternatives. Original GT (Fixed Self) provides an ideal self-reference, with five reference-dependent metrics set to 100 and Quality retaining its measured value. Farthest Alternative reports measured scores. Quality differences reflect intrinsic image quality and are separate from reference-related penalties.
TaskCandidateDINO-ISemanticSpatialQualityRScoreTScoreOverall
ROGOriginal GT (Fixed Self)100.0100.0100.057.8100.0100.093.0
ROGFarthest Alternative94.299.489.257.894.096.488.5
MCVIOriginal GT (Fixed Self)100.0100.0100.060.3100.0100.093.4
MCVIFarthest Alternative92.597.285.958.693.096.787.3

Even the most pose-distant valid alternatives among the retained candidates achieve Overall scores of 88.5 / 87.3 on ROG / MCVI, showing that the fixed-GT protocol preserves substantial credit for valid alternative views, although some reference-related penalties remain.

AuthorsOpenReview ↗

Response to Reviewer Lsi6 (Part 2/3)

W2 / Q2. Contributions Beyond GSI-Bench

The central added value of MindView is to test whether a model remains faithful to the same real scene jointly constrained by multiple input images when generating a different view.

The single-image input in GSI-Bench does not provide direct observations beyond the original field of view. Even when a target GT is available for evaluation, differences in the generated image may reflect plausible completion of unobserved regions. MindView, by contrast, can check scene content and relations that are clearly shown in other input views but are not preserved in the generated output. For Rotation, for example, if another input image already shows a cabinet and a window along a wall, but the generated image turns the cabinet into a table or changes their spatial relationship in the scene, this constitutes a consistency error that can be verified against the input evidence. This directly tests whether models can integrate cross-view observations and respect these scene constraints when generating a new view.

Transfer to spatial understanding provides additional support. We evaluated the official GSI-Bench finetuned BAGEL checkpoint under the same evaluation protocol. Both fine-tuning approaches use the BAGEL backbone and generation supervision only.

ModelSAT-v2 MicroSAT-v2 Both-orderMMSI-BenchMindCube-Tiny
Vanilla BAGEL58.027.331.034.7
GSI-Bench BAGEL59.328.029.934.8
MindView-BAGEL-SFT62.741.332.236.5

MindView outperforms Vanilla BAGEL on all three understanding benchmarks, and its gains on all four metrics are larger than those of GSI-Bench. In particular, on MMSI-Bench, MindView improves accuracy from 31.0% to 32.2%, whereas GSI-Bench decreases it to 29.9%. The more consistent gains across benchmarks in this comparison provide additional evidence for the training value of MindView's scene-level spatial generation tasks.

W3 / Q3. Effects of View Relevance, Count, and Order

We added input-order shuffling, duplicate inputs, and count-matched view selection to examine the roles of order, image count, and additional-view selection in the C10 decline.

Setting# ImagesBAGELBAGEL-SFTΔC6ΔC10
Fixed C66.040.064.40.0 / 0.0+1.0 / +1.2
Fixed C1010.039.063.2−1.0 / −1.20.0 / 0.0
C10 Shuffle10.039.162.9−0.9 / −1.5+0.1 / −0.3
C6 + 4 Duplicates10.039.664.1−0.4 / −0.3+0.6 / +0.9
Geometric Selection7.438.763.6−1.3 / −0.8−0.3 / +0.4
RandDrop7.439.664.1−0.5 / −0.3+0.5 / +0.9
MindCue7.441.165.1+1.1 / +0.8+2.1 / +1.9

Scores are task-macro averages of Overall, and # Images denotes the average number of input images. ΔC6 / ΔC10 denotes the score difference relative to C6 / C10. Each pair is reported in BAGEL / BAGEL-SFT order.

  1. Input-order shuffling. C10 Shuffle randomly permutes the same set of C10 inputs, keeping the anchor in its original position for tasks that have one. After shuffling, both models remain close to their original C10 scores and below their C6 baselines. This indicates that changing the input order alone does not resolve the performance drop in the tested setting.

  2. Duplicate inputs. C6 + 4 Duplicates preserves the positions of the C6 views within C10 and replaces the additional views with copies of four non-anchor core views. With the same total of ten input images, it outperforms C10 by 0.6 / 0.9 points, but still falls 0.4 / 0.3 points below C6. This indicates that the added content accounts for much of the C10 decline beyond image count alone, highlighting the importance of selecting useful supplementary evidence.

  3. Count-matched view selection. Geometric Selection, RandDrop, and MindCue retain the same number of additional views for each sample from the same C10 candidate pool, preserving all C6 views and the original relative order of all retained views. The added Geometric Selection control does not use the query and prioritizes views that add coverage of previously uncovered regions, overlap less with the core views, and have higher reconstruction confidence. MindCue still outperforms both controls, supporting the conclusion that selecting evidence according to the target requirements provides additional gains after controlling for image count.

Taken together, these controls suggest that the benefit of C10’s increased coverage is outweighed by the difficulty of integrating more images and interference from less task-relevant content. MindCue uses the query-derived target-view hypothesis and supporting evidence to prioritize useful supplementary views, limiting irrelevant context while achieving gains over C6 and count-matched alternatives.

AuthorsOpenReview ↗

Response to Reviewer Lsi6 (Part 3/3)

W4 / Q4. Reference Models for Spatial Understanding and Training Controls

We added reference models and a training-volume-matched control to assess the level of understanding performance and the source of transfer gains, respectively.

  1. Reference models. We include the spatial reasoning model SpaceR [1], the general understanding model LLaVA-OneVision-1.5 [2], the unified understanding and generation model Emu3.5 [3], and GSI-BAGEL-SFT [4] as a reference model trained on generation tasks. All follow the input and scoring protocol of Table 3.
ModelSAT-v2 MicroSAT-v2 Both-orderMMSI-BenchMindCube-Tiny
SpaceR (7B)57.744.728.931.7
LLaVA-OneVision-1.5 (8B)60.340.729.036.5
Emu3.5 (34B)61.736.730.745.7
GSI-BAGEL-SFT (7B)59.328.029.934.8
MindView-BAGEL-SFT (7B)62.741.332.236.5

MindView-BAGEL-SFT achieves the highest scores among the listed models on both SAT-v2 Micro and MMSI-Bench. Emu3.5 performs better on MindCube-Tiny. Its world-modeling training, which emphasizes cross-view scene consistency, may be better suited to the spatial layout inference required by this benchmark.

  1. Training-volume control. We fine-tune the same BAGEL using MACRO [5], a general multi-image generation dataset. The selected samples contain an average of 5.8 input images, close to MindView's 6. Both groups use 20,000 training samples and 10,000 steps, generation supervision only, and the same core training hyperparameters.
ModelSAT-v2 MicroSAT-v2 Both-orderMMSI-BenchMindCube-Tiny
Vanilla BAGEL58.027.331.034.7
MACRO-BAGEL-SFT60.031.329.234.5
MindView-BAGEL-SFT62.741.332.236.5

With training samples and updates matched, MindView outperforms MACRO on all four metrics, with gains of 10.0 and 3.0 percentage points on SAT-v2 Both-order and MMSI-Bench, respectively. MACRO improves on SAT-v2 but falls below Vanilla BAGEL on MMSI-Bench and MindCube-Tiny, while MindView improves on all three benchmarks. These results support the transfer value of MindView’s spatial generation tasks beyond the benefit of additional training alone.

[1] SpaceR: Reinforcing MLLMs in Video Spatial Reasoning. arXiv 2025.

[2] LLaVA-OneVision-1.5: Fully Open Framework for Democratized Multimodal Training. arXiv 2025.

[3] Emu3.5: Native Multimodal Models are World Learners. arXiv 2025.

[4] Exploring Spatial Intelligence from a Generative Perspective. CVPR 2026.

[5] MACRO: Advancing Multi-Reference Image Generation with Structured Long-Context Data. arXiv 2026.

Thanks to the authors for adding the experiments and for their response to my questions. I think these experiments and replies have addressed my previous concerns, and I will maintain my original decision of weak accept.

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

The paper introduces MindView, a benchmark for generating a requested view of an indoor scene from several images and a language instruction. Its 1,500 examples cover seven tasks, from explicit camera movements to viewpoints described through objects and spatial relations. The authors evaluate eleven generators and find that adding more input views can hurt performance. They also construct 20,000 training examples, show improvements from fine-tuning BAGEL, and propose MindCue to select additional input views.

Critical strengths
  1. MindView provides a useful benchmark for language-guided multi-view generation, with tasks spanning explicit camera control and semantic grounding. This task design enables a systematic evaluation of content fidelity and viewpoint consistency, revealing limitations in current generative models that simply adding more context views does not overcome. These findings highlight the need for better use of multi-view spatial information and offer guidance for future model training.
Critical weaknesses
  1. Single-target evaluation may penalize valid outputs in ROG and MCVI. The paper checks uniqueness among recorded frames, but other viewpoints may also satisfy the instruction. Since the VLM judge receives neither the instruction nor the input views, it cannot assess whether an alternative output fulfills the request.
  2. The experiments do not isolate the benefit of MindCue’s target-view hypothesis. RandDrop controls the number of selected views, but no query-independent geometry-based selector or ablation of the target-camera hypothesis is included. These comparisons are needed to distinguish the contribution of target-specific reasoning from general filtering of redundant or unreliable views.
Questions for authors
  1. How does MindCue convert meter-valued instructions into VGGT’s reconstruction coordinates without access to sensor depth or ground-truth poses?
  2. Could you compare MindCue with a count-matched, query-independent geometric selector and ablate the target-camera hypothesis to isolate its contribution?
3 comments
AuthorsOpenReview ↗

Response to Reviewer VrJi (Part 1/2)

W1. Single-GT Evaluation: Rationale and Robustness

We agree that a single GT may penalize valid alternative views, as noted in our Limitations. We therefore added two experiments: comparison with the Instruction-aware LLM Judge, using human ratings to assess the rationale for fixed-GT LLM scoring, and reference GT replacement to test the robustness of model comparisons.

  1. Comparison with the Instruction-aware LLM Judge. We selected 30 questions each from ROG and MCVI, with 5 models producing 300 outputs, and had the three experts from Table 5 (revised version) rate them. Both automatic judges use the same Qwen3-VL-30B-A3B-Instruct model. The original evaluation compares only the GT and candidate image. The Instruction-aware LLM Judge and human experts use the same scoring criteria and see the instruction, 6 input images, a GT reference example, and the candidate image. Values are reported in Spatial / Semantic order, and correlations are computed across the five model means.
Model / StatisticHumanInstruction-aware LLM JudgeOriginal LLM Judge (Ours)
BAGEL39.1 / 40.281.7 / 85.028.3 / 39.7
BAGEL-SFT77.5 / 74.992.0 / 93.749.3 / 61.3
GPT-Image-272.6 / 73.599.7 / 99.742.3 / 54.3
Nano Banana Pro65.0 / 62.695.7 / 95.742.0 / 50.7
SenseNova-U1.551.7 / 51.394.3 / 95.339.3 / 49.0
Pearson r vs. Human—0.7254 / 0.75570.9394 / 0.9386
Spearman ρ vs. Human—0.40 / 0.401.00 / 1.00

The Instruction-aware LLM Judge assigns consistently high scores (GPT-Image-2: 99.7 on both metrics), possibly rewarding plausible layouts without fully understanding the input scene or verifying all spatial constraints, both of which can be challenging for LLMs. The original LLM judge shows stronger agreement with human scores (Pearson r: 0.9394 / 0.9386 vs. 0.7254 / 0.7557) and rankings (Spearman ρ: 1.00 / 1.00 vs. 0.40 / 0.40), supporting a fixed GT as an explicit scoring reference for model-level comparisons.

  1. Reference GT replacement. We identified 65 ROG groups and 143 MCVI groups in the construction records, each with at least an alternative GT satisfying the instruction and image-clarity requirements. Holding the generated images from all 12 models fixed, we replaced each sample’s GT with the captured view having the largest combined pose difference from the original GT among its saved valid candidates. Overall is the mean of six metrics. Parentheses show the score after replacement minus the score before replacement.
ModelROGMCVIModelROGMCVI
BAGEL43.5 → 43.6 (+0.1)36.2 → 35.9 (−0.3)SenseNova-U1.558.3 → 58.9 (+0.6)51.3 → 50.9 (−0.4)
FLUX.251.9 → 51.8 (−0.1)46.3 → 46.3 (0.0)Nano Banana54.6 → 54.8 (+0.2)47.5 → 47.4 (−0.1)
Qwen-Image-Edit-251150.0 → 50.5 (+0.5)40.9 → 40.2 (−0.7)Nano Banana Pro58.3 → 58.2 (−0.1)49.4 → 49.3 (−0.1)
Emu3.542.9 → 42.5 (−0.4)35.4 → 35.6 (+0.2)GPT-Image-1.560.3 → 60.5 (+0.2)54.0 → 54.0 (0.0)
DreamOmni234.0 → 33.8 (−0.2)31.6 → 31.4 (−0.2)GPT-Image-261.6 → 62.0 (+0.4)58.2 → 57.9 (−0.3)
SenseNova-U149.0 → 48.3 (−0.7)43.4 → 42.7 (−0.7)BAGEL-SFT61.1 → 60.7 (−0.4)57.0 → 57.0 (0.0)

Each model’s task mean changes by at most 0.8 points. Model rankings remain highly consistent (Spearman ρ: 0.993 / 1.00 for ROG / MCVI), with only one pair of adjacent models swapping ranks in ROG. These results support the stability of the main model comparisons to the choice of a valid reference GT on the evaluated subsets. We have added these experiments to the revised manuscript (Appendix B.6, Tables 7–8).

AuthorsOpenReview ↗

Response to Reviewer VrJi (Part 2/2)

W2 / Q2. Geometric Selection and Target-View Hypothesis Ablations

Thank you for this valuable suggestion. We added a general geometric selection control (G) and target-view hypothesis ablations (H1–H3). Each strategy matches MindCue's number of additional views for every sample.

  1. Geometric selection control (G). G (Geometric Selection) uses neither the query nor a target-view hypothesis. It prioritizes views that cover regions not covered by the core views, overlap less with the core views, and have higher reconstruction confidence.

  2. Target-view hypothesis ablations (H1–H3). H1 / H3 use the mean camera pose of the core views and a randomly selected core camera pose, respectively, both with query-derived semantic support removed. H2 recomputes the hypothesis and supporting evidence using a mismatched query, while the generator still receives the original instruction.

The table below reports differences in Overall between the reproduced MindCue results and each control, all run in the same environment.

ComparisonBAGELBAGEL-SFT
MindCue − G+2.4+1.6
MindCue − H1+0.5+0.5
MindCue − H2+1.2+0.8
MindCue − H3+1.3+0.8

Compared with G, MindCue improves the two generators by 2.4 / 1.6 points, respectively, showing that with matched view counts and the original relative order preserved, selecting evidence for the target is more effective than this general geometric selection strategy. MindCue also outperforms H1–H3, further supporting that a target-view hypothesis and supporting evidence matched to the current query help select more effective additional views. We have added these experiments to the revised manuscript (Appendix E.5, Table 10).

Q1. Metric Scale Calibration

Thank you for pointing out this detail. We had included scale calibration as part of the post-processing of VGGT reconstruction, with a scale-recovery rationale similar to that of GS-Reasoner [1], but had not described it separately. We have now added the scale-calibration details and unit conversion procedure to Appendix E.1 of the revised manuscript.

MindCue uses the lightweight monocular metric geometry estimation model MoGe-2 (35M) [2] to estimate scale solely from the input RGB images. We estimate s (meters per VGGT reconstruction unit) by unweighted least-squares alignment of corresponding VGGT camera-space points and MoGe-2 metric points over valid pixels, fitting only a positive scale. The unit conversion follows the same proportional scaling principle as the geometric evaluation: if each VGGT unit corresponds to s meters, then d meters in the instruction corresponds to d/s reconstruction units. This conversion is used to construct the target-camera hypothesis and distance constraints, without requiring sensor depth or ground-truth camera poses from the test scene, or the target GT.

The model has approximately 2.8% as many parameters as VGGT. Calibration takes only 0.035 seconds per sample on average and is already included in VGGT reconstruction in Table 11 of the revised manuscript. On all 1,500 evaluation samples, replacing RGB-based calibration with calibration using the input views' ground-truth depth increases Overall only from 65.1 to 65.3, indicating that RGB-based scale estimation already achieves downstream performance close to ground-truth-depth calibration in the current evaluation.

[1] Reasoning in Space via Grounding in the World. ICLR 2026.

[2] MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details. NeurIPS 2025.

The additional experiments provided by the authors have addressed my concerns, and I will keep my score.

Reviewer_746bOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This is a benchmark for evaluating the ability to "generate a target view from partial multi-view observations, guided by a natural language query." It comprises 1,500 samples across 513 scenes, sourced from ScanNet, ScanNet++, and ARKitScenes. The tasks are divided into two categories: geometric tasks, which use text to specify translation, rotation, or a combination of both; and semantic tasks, which require the model to infer a target view based on referenced objects and their spatial relationships. The target views consist of actual captured images. Three key findings emerge: First, even the best-performing model among the 11 generators (GPT-Image-2) achieved a score of only 57.9, indicating significant room for improvement. Second, while increasing the context from 6 to 10 images improves coverage, it leads to a decline in the performance of all models. Third, fine-tuning the BAGEL model on 20,000 training samples yields a 24.3-point increase in its total score, with concurrent improvements observed across three external benchmarks for spatial understanding. The authors also introduce "MindCue," a training-free evidence selector capable of reversing the performance degradation caused by larger context sizes.

Critical strengths

The evaluation is robust and extensive. VLM-based assessments and human expert evaluations show a 92.4% agreement in ranking across 210 samples. For the VGGT pose evaluation, the authors conducted stability tests across different generators, finding a median deviation of only 0.71 degrees for rotation and 2.75 cm for position relative to the ground truth. During the construction phase, an LLM was used in conjunction with independent reviews by three experts, yielding a Krippendorff's α of 0.815.

Critical weaknesses
  1. The 24.3-point improvement seen in BAGEL-SFT appears to be a case of in-distribution adaptation. Since the training set and the benchmark are derived from the same three datasets and utilize identical instruction templates, this score likely reflects the model learning the task format rather than an actual gain in spatial intelligence. Out-of-distribution (OOD) evaluation is necessary to demonstrate true generalization capabilities. Furthermore, the phenomenon where generation facilitates understanding has already been established by the Omni-View [1]. This paper does not appear to address that prior work.
  2. The performance gains from MindCue are marginal, ranging from only 0.8 to 1.6 points, which is close to the level of statistical noise. The note in section B.5 stating that "no significance testing was performed", is this "defensive writing" by an AI?
  3. The test images used are drawn from publicly available datasets. How can it be guaranteed that the closed-source models have not encountered this data? If they have, is the model’s output derived from “inferring the viewpoint based on partial observations” or from “retrieving the viewpoint from its memorized pre-training corpus”?
  4. The writing is somewhat disorganized.

[1] Omni-View: Unlocking How Generation Facilitates Understanding in Unified 3D Model based on Multiview images. ICLR 2026.

Questions for authors

Qwen-VL is used to evaluate the output visual signals. However, Qwen-Image also uses the output from Qwen-VL as a condition. How does the evaluation process prevent Qwen-Image from "cheating"?

10 comments
AuthorsOpenReview ↗

Response to Reviewer 746b (Part 1/2)

W1. Generalization and Related Work

Thank you for raising this concern. Our SFT experiments were designed to test whether task-specific generation supervision improves performance on these spatial tasks. We agree that OOD evaluation across datasets and instruction formulations can further test the generalization of these gains, and have therefore added the following experiments.

  1. Cross-dataset generalization. We completely remove ScanNet data from fine-tuning and retrain using only ARKitScenes and ScanNet++, keeping 10,000 training steps and all other training settings unchanged. On the same 654 ScanNet evaluation samples, Avg remains 63.2, a 22.6-point improvement over the original BAGEL, retaining 90.4% of the in-distribution SFT gain.
ModelSFT dataTrans.Rot.Comp.ROGTFROCRMCVIAvg
BAGELNone44.139.837.841.741.044.835.340.6
BAGEL-SFT (paper)All three datasets68.769.562.161.968.273.655.665.6
BAGEL-SFT (OOD)ARKitScenes + ScanNet++70.4 (+26.3)60.0 (+20.2)58.0 (+20.2)64.1 (+22.4)63.1 (+22.1)69.7 (+24.9)56.7 (+21.4)63.2 (+22.6)
  1. Robustness to instruction rewriting. Our task-specific instructions are not restricted to a single fixed formulation. We incorporate phrasing diversity during dataset construction, for example by expressing the same spatial constraints in multiple ways in MCVI. To further test robustness, we use GPT-6.1-Sol to rewrite the shared instruction template, task-specific core instruction, or both for all 1,500 samples, preserving task meaning and spatial constraints. The paper’s checkpoint, input views, and inference settings remain unchanged, with no retraining. Avg drops by 0.3, 1.9, and 2.3 points, respectively. When both are rewritten, it still reaches 62.1, retaining 90.6% of the original fine-tuning gain over the original Vanilla BAGEL baseline.
CorePublic templateTrans.Rot.Comp.ROGTFROCRMCVIAvg
OriginalOriginal67.168.256.362.766.573.456.264.4
OriginalRewritten66.9 (−0.2)67.8 (−0.4)56.1 (−0.2)62.4 (−0.3)66.1 (−0.4)72.9 (−0.5)56.3 (+0.1)64.1 (−0.3)
RewrittenOriginal65.5 (−1.6)66.5 (−1.7)53.2 (−3.1)60.5 (−2.2)63.0 (−3.5)72.9 (−0.5)55.8 (−0.4)62.5 (−1.9)
RewrittenRewritten64.9 (−2.2)66.5 (−1.7)52.5 (−3.8)60.9 (−1.8)62.1 (−4.4)72.6 (−0.8)55.5 (−0.7)62.1 (−2.3)

Together, these results show that the gains transfer to a data source excluded from fine-tuning, while performance remains robust to changes in instruction wording, providing evidence beyond adaptation to the original data source and fixed templates.

  1. Distinction from Omni-View. We agree that this work has studied how generation facilitates understanding, and we have added a more direct comparison in the Introduction of the revised manuscript. Omni-View improves 3D understanding through joint training of understanding and generation. We further examine whether spatial generation training alone can transfer to understanding, without adding understanding supervision during fine-tuning. We retain the original BAGEL architecture and use only RGB target-view generation supervision, yet obtain gains on three understanding benchmarks. The training-volume-matched MACRO control (Section 4.4, Table 3, and Appendix D) further supports that these gains are related to the design of the spatial generation tasks.
AuthorsOpenReview ↗

Response to Reviewer 746b (Part 2/2)

W2. Statistical Reliability

Thank you for this question. We have added experiments showing that MindCue’s gains are reproducible across the tested generation seeds.

  1. Stability across random seeds. We evaluate all five strategies using the same five generation seeds (42–46) on both BAGEL and BAGEL-SFT, keeping model checkpoints and each strategy’s input views fixed. All entries report mean ± sample standard deviation of Overall.
ModelFixed C6 (6.0)Fixed C10 (10.0)CounterSel (8.6)RandDrop (7.4)MindCue (7.4)ΔC6ΔC10
BAGEL40.0 ± 0.339.2 ± 0.137.2 ± 0.439.5 ± 0.241.0 ± 0.3+1.0+1.8
BAGEL-SFT64.3 ± 0.263.1 ± 0.362.6 ± 0.364.1 ± 0.165.2 ± 0.2+0.9+2.1

MindCue achieves the highest mean Overall on both backbones. Its gains over Fixed C6 and Fixed C10 exceed the observed standard deviations of both MindCue and these baselines, supporting the stability of the improvements across the tested generation seeds.

  1. Clarification of Appendix B.5. We included the final sentence to clarify the scoring and statistical reporting conventions for the entire benchmark, rather than specifically for MindCue. Overall is the mean of six metrics and should not be equated with accuracy. Following the reporting approach of GenSpace [1] and GSI-Bench [2], we mainly compared scores under fixed settings and explicitly disclosed that no significance testing was performed, in accordance with the principles of the ICLR Code of Ethics. The additional multi-seed experiments further validate the stability of MindCue’s gains.

W3. Exposure to Evaluation Data during Pre-training

This is a shared compromise when evaluating closed-source models on public datasets. For example, GSI-Bench [2] uses ScanNet++ for its real-scene component, while MSE-Bench [3] uses MS-COCO and LAION-Aesthetics, and both evaluate closed-source generative models. Because these models do not disclose their complete training data, it is difficult to verify whether they have encountered these images. However, public availability does not establish that an image was used for training, and even prior exposure alone does not establish that a model solves the task through memorization. We have explicitly stated this risk in the paper’s limitations (Appendix G).

[1] GenSpace: Benchmarking Spatially-Aware Image Generation. NeurIPS 2025.

[2] Exploring Spatial Intelligence from a Generative Perspective. CVPR 2026.

[3] VINCIE: Unlocking In-context Image Editing from Video. ICLR 2026.

W4. Presentation and Organization

We appreciate this feedback. The MindView benchmark is our core contribution. We first establish its tasks and evaluation protocol, then use the benchmark to identify limitations in current models. Limited spatial generation motivates SFT to examine learning and transfer, while degradation with additional views motivates MindCue to explore query-relevant evidence selection. Both studies build on MindView’s findings and offer practical directions for the community. We have clarified this progression in the experimental overview and strengthened the transitions between the corresponding subsections in the revised manuscript. We would be happy to provide any further clarification during the discussion.

Question: Potential Bias of the Qwen-Based Evaluator

  1. Clarification of the evaluation setup. To clarify, the generator and evaluator operate independently and do not share intermediate features. Qwen-Image-Edit-2511 uses Qwen2.5-VL features for generation, while our evaluator is Qwen3-VL-30B-A3B-Instruct (see §3.3 and Appendix B.2 of the revised manuscript).

  2. Independent check for same-family bias. We nevertheless agree that potential same-family preferences should be tested. We therefore re-score the same 1,500 saved outputs from each of the 12 models using GPT-6-Sol, keeping the original scoring criteria, image preprocessing, and order unchanged. The table reports the equally weighted mean of Semantic, Spatial, and Quality, with tasks also weighted equally.

ModelQwenGPT-6-SolModelQwenGPT-6-Sol
BAGEL41.839.5SenseNova-U1.555.051.7
FLUX.256.451.2Nano Banana54.050.8
Qwen-Image-Edit-251152.448.3Nano Banana Pro60.955.2
Emu3.546.544.8GPT-Image-1.562.556.9
DreamOmni243.440.5GPT-Image-265.160.0
SenseNova-U150.047.7BAGEL-SFT63.357.6

Under both evaluators, Qwen-Image-Edit-2511 ranks 8th, with no clear ranking advantage from using an evaluator of the same family observed in this comparison. The rankings of all 12 models are also highly consistent (Spearman ρ = 0.993), supporting the robustness of the main comparisons to the choice of evaluator.

Reviewer_746bOpenReview ↗

Questions about the training process.

Could you explain what the training process is?

You have mentioned that "We retain the original BAGEL architecture and use only RGB target-view generation supervision." Does this mean that during training, MindCue enables gradients for both the understanding and generation branches of BAGEL, while using only the loss from the generation branch to supervise the understanding branch?

This raises a significant issue: since the understanding and generation branched of BAGEL are decoupled, the loss derived from the generation branch cannot be directly used to update the parameters of the understanding branch. Omni-View addresses this by incorporating features from the understanding branch into an additional output head on the generation branch, yet it appears that MindCue does not implement a similar mechanism let the gradient flows from loss of generation to understanding?

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their careful reading and specific suggestions. In response, we have added experiments, controls, and methodological details, and further clarified MindView’s contributions and conclusions.

We appreciate the reviewers’ recognition of the following strengths:

  • Meaningful and valuable benchmark design (Lsi6, VrJi). The seven tasks span explicit camera control and viewpoint inference from objects and spatial relations, revealing specific limitations in current generators beyond aggregate scores.

  • Rigorous construction and robust evaluation (746b, Lsi6). Real captured targets, careful ambiguity screening, and independent expert review support benchmark reliability, complemented by 92.4% human–judge ranking agreement and geometric stability tests.

  • Extensive experiments and consistently strong adaptation gains. 746b recognizes the breadth and robustness of the evaluation. Lsi6 highlights BAGEL-SFT’s improvements across all seven tasks and all 42 task–metric combinations, demonstrating the breadth and consistency of the adaptation gains.

The reviews also raised several common concerns, for which we provide additional evidence:

  1. Generalization. Removing all ScanNet data from fine-tuning preserves 90.4% of the original fine-tuning gain on MindView’s ScanNet evaluation subset. The model also retains 90.6% of the original fine-tuning gain when both the shared instruction template and task instruction are rewritten. These results support cross-dataset generalization and robustness to instruction rewriting.

  2. Evaluation reliability. We added human comparisons, valid alternative GTs, and independent re-evaluation with GPT-6-Sol to test the robustness of the main comparisons. Model rankings remain relatively stable within the scope tested, and we also quantify the average penalty that single-GT evaluation still imposes on valid alternative viewpoints.

  3. MindCue’s gains and their sources. Regarding these gains and their sources, replications across five seeds show stable improvements. With 10 input images in both conditions, duplicating existing views performs better than adding new views, and shuffling the order does not restore performance to the six-view baseline. Together with count-matched geometric selection and target-view hypothesis ablations, these results further support that MindCue’s target-guided selection of supplementary evidence yields gains beyond simple image-count controls and general geometric filtering.

  4. Reference models and controls for transfer to spatial understanding. We added representative understanding models, a GSI-Bench checkpoint, and a MACRO training control. With the number of training examples and steps matched, MindView outperforms MACRO on all three understanding benchmarks, further supporting the transfer value of scene-level generation training.

Detailed experimental settings, results, and explanations are provided in the responses below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.