acceptodds
Under review as a conference paper at ICLR 2027

PanoReward: Adaptive Multi-Criteria Reward Modeling for Panoramic Image Generation

Abstract

Panoramic image generation has become an important component of VR/AR, cinematography, and world modeling. However, progress in this area is constrained by the lack of a reliable reward model, which in turn caps achievable performance. We posit that reward modeling for panoramic generation should be inherently multi-criteria, explicitly integrating multiple evaluation dimensions and adopting adaptive, preference-conditioned weighting to reflect judgment standards across diverse task requirements. To mitigate this bottleneck, we propose PanoReward, a reward model that evaluates panoramic output along three complementary axes: (i) 3D consistency, (ii) polar visual quality, and (iii) image-text consistency. The model is trained on PanoReward-Data, which contains 943,074 task-profile training instances derived from 195,930 source pairs. In panoramic image generation settings, PanoReward exhibits a strong and consistent alignment with human preferences under different task requirements. In addition, PanoReward supports downstream workflows for panoramic generation, e.g., reward-driven post-training. We will release both PanoReward and the associated training dataset.

Primary area:
datasets and benchmarks
Keywords:
Panorama Generation, Reward Model

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_BXo8OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

PanoReward is a reward model for panoramic image generation that scores 3D consistency, polar visual quality and image-text consistency on Qwen3-VL, and fuses them with weights predicted from a task instruction. The paper also contributes PanoReward-Data (943K task-conditioned preferences built from human-refined dimension-level labels) and a human-annotated benchmark, PanoReward-Bench, and uses the reward to post-train DiT360.

Critical strengths
  • A real and underexplored problem. Panoramic generation lacks a dedicated reward model, and the trade-off between geometry, pole quality and text fidelity is real.

  • Interesting reversal-based supervision. Turning one set of dimension scores into labels under different priority profiles, and focusing on pairs whose winner flips with the task, is a clever way to force the model to actually read the instruction instead of learning a single notion of "better".

  • Large dataset with human-refined labels. About 196K source pairs with three dimension-level labels, plus a 10-annotator benchmark with reported agreement, would be very useful to the community if released.

Critical weaknesses
  • Missing the most direct baseline. Every panorama already has human-refined G/V/T labels, and all task-conditioned labels are a deterministic function of these scores and the (8, 4, 1) template, so they carry no extra information. A natural baseline is to train each head directly on its own dimension labels, and use the task-conditioned (reversal) labels only to train the fusion branch. Instead, the heads are supervised only through the fused reward, and without task-conditioned labels they collapse to nearly the same score (Spearman 0.96–0.99, line 842). It is unclear whether this indirect supervision helps or hurts, especially on G, the weakest dimension (0.705).

  • Possible model bias. Training and benchmark use the same four generators, and each dimension is dominated by one of them: Matrix3D wins 27/45 G cells and FLUX wins 34/45 T cells (line 698). The reward model may simply learn which generator is good at what, rather than judging the image itself. There is no experiment on unseen generators to show the reward model generalizes.

Questions for authors
  • Can you add a baseline where each head is trained directly on its own dimension labels and the fusion branch is trained on the task-conditioned labels on top? How does it compare on per-dimension accuracy (especially G) and flip-both?

  • Can you run an experiment free of this model bias, e.g., evaluate PanoReward on panoramas from generators unseen in training, such as Qwen-Image-2.1 or FLUX.3?

  • The weights (8,4,1) is chosen to maximize reversal pairs (line 228), not agreement with humans. Would another template such as (4, 2, 1) change the labels and the results (accuracy, flip-both) much?

  • The RL setup is almost entirely missing. The algorithm is only implied by a citation to DiffusionNFT (line 412). Please specify the algorithm, training prompts, steps, and key hyperparameters.

6 comments

Thanks for your insightful comments. We list your advice and questions followed by detailed answers, then argue that your concerns can be addressed.

Weakness1 & Question1. Missing the most direct baseline. Every panorama already has human-refined G/V/T labels, and all task-conditioned labels are a deterministic function of these scores and the (8, 4, 1) template, so they carry no extra information. A natural baseline is to train each head directly on its own dimension labels, and use the task-conditioned (reversal) labels only to train the fusion branch. Instead, the heads are supervised only through the fused reward, and without task-conditioned labels they collapse to nearly the same score (Spearman 0.96–0.99, line 842). It is unclear whether this indirect supervision helps or hurts, especially on G, the weakest dimension (0.705). Can you add a baseline where each head is trained directly on its own dimension labels and the fusion branch is trained on the task-conditioned labels on top? How does it compare on per-dimension accuracy (especially G) and flip-both.

Answer. We conducted the requested comparison and then examined the concern about representation collapse. The results show that directly supervising each head with its own dimension labels is less effective than our fused preference supervision, particularly for G and task-conditioned preference switching.

(1/2) Direct-head supervision baseline. We compare the following two settings:

  • PanoReward: The three quality heads are indirectly supervised through the task-conditioned fused preference loss.
  • Direct-head supervision: Human-refined dimension scores are used to construct dimension-level pairwise losses for the G/V/T heads. The Fusion Branch is trained with task-conditioned preference labels; the fused loss updates only the Fusion Branch, with the head outputs detached.

Both methods use the same backbone, data split, and training configuration. We compare pairwise accuracy on G/V/T, with particular attention to G, as well as overall accuracy and flip-both accuracy:

VariantGTVAllFlip-both
Direct-head supervision0.4950.8170.8220.7130.483
PanoReward (ours)0.7050.8570.7780.7850.867

The comparison directly answers the question raised by this baseline. Replacing the fused supervision with direct dimension-level supervision reduces overall accuracy from 0.785 to 0.713, G accuracy from 0.705 to 0.495, and flip-both accuracy from 0.867 to 0.483. Only V improves slightly (0.778 to 0.822). Thus, indirect supervision through the fused reward does not impair the learning of the individual quality dimensions; instead, it is more effective for learning discriminative dimension representations, particularly G, while supporting task-conditioned preference switching.


(2/2) Analysis of representation collapse. PanoReward itself does not exhibit the collapse described in the review. The pairwise Spearman correlations between the outputs of its three heads are -0.72, -0.60, and -0.02 (reported in Appendix F), indicating that the dimension representations are not significantly collapsed. The high correlations of 0.96–0.99 occur only in the model trained without task-conditioned labels, not in the complete PanoReward model.

We also note that G is a more challenging evaluation dimension, including for human annotators. Its inter-annotator agreement (Fleiss' κ = 0.511) and leave-one-out human-reference agreement (0.801) are the lowest among the three dimensions. This suggests that the relatively lower G accuracy is substantially related to the intrinsic difficulty of the task, rather than being caused by our training strategy. Consistent with this observation, G accuracy decreases further to 0.495 under direct-head supervision, whereas PanoReward reaches 0.705 with fused preference supervision. Taken together, these results support the effectiveness of our training strategy for G assessment.

Weakness2 & Question2. Possible model bias. Training and benchmark use the same four generators, and each dimension is dominated by one of them: Matrix3D wins 27/45 G cells and FLUX wins 34/45 T cells (line 698). The reward model may simply learn which generator is good at what, rather than judging the image itself. There is no experiment on unseen generators to show the reward model generalizes. Can you run an experiment free of this model bias, e.g., evaluate PanoReward on panoramas from generators unseen in training, such as Qwen-Image-2.1 or FLUX.3?

Answer. Thank you for suggesting this experiment. We generated panoramas for the same 45 scene descriptions in PanoReward-Bench using two generators that were never seen during training: FLUX.3 and Qwen-Image-2.1, producing 90 panoramas in total.

For clarity, we refer to images generated by the four generators used during training (DiT360, FLUX.2, HunyuanWorld-1, and Matrix3D) as in-domain images, and those generated by FLUX.3 and Qwen-Image-2.1 as out-of-domain images. We construct three groups of pairwise comparisons within the same scene:

  • Out-of-domain vs. out-of-domain: 45 pairs, comparing FLUX.3 with Qwen-Image-2.1;
  • Out-of-domain vs. in-domain: 360 pairs, comparing each out-of-domain image with the four in-domain images from the same scene;
  • In-domain vs. in-domain: 270 pairs, used as the reference group.

We manually annotated the out-of-domain images using the same procedure as for the in-domain images. The results are as follows.

1. The model changes its preference between two unseen generators when the task instruction changes.

For the FLUX.3 and Qwen-Image-2.1 panoramas from the same scene, changing the instruction from “3D consistency priority” to “text consistency priority” reverses PanoReward's preference in 93.3% (42/45) of the scenes. The directions of these reversals are also consistent with the respective strengths of the two generators.

2. Dimension-wise results.

DimensionOut-of-domain vs. out-of-domainOut-of-domain vs. in-domainIn-domain vs. in-domain
G, cross-view consistency0.5560.4360.596
V, polar visual quality0.7560.8580.900
T, text consistency0.9760.9210.889

For generators unseen during training, PanoReward's T judgments agree with the independent evaluator at a rate no lower than the in-domain reference, while V decreases only slightly. Performance on G drops substantially. We believe this relates to the greater difficulty of assessing cross-view consistency, as discussed in our response to the "Weakness1&Question1" you raised.

Question3. The weights (8,4,1) is chosen to maximize reversal pairs (line 228), not agreement with humans. Would another template such as (4, 2, 1) change the labels and the results (accuracy, flip-both) much?

Answer. Thank you for suggesting the alternative weight template. We agree that it is important to examine how the choice of weights affects both the constructed preference labels and the performance of the resulting reward model. The main reason for using (8, 4, 1) is that integer programming selects it to increase preference-sensitive and preference-reversal samples while satisfying the prescribed tie-rate and priority constraints.

We therefore evaluate the suggested (4, 2, 1) template in two stages.

1. Stability of preference-label construction.

We construct task-conditioned preference labels using both templates on the same 195,930 source image pairs:

Metric(8, 4, 1)(4, 2, 1)
Weighted-tie units0 (0%)37,948 (3.2%)
Preference-sensitive source pairs127,734 (65.2%)89,027 (45.4%)
Reversal-type counterfactual units242,428189,189
Agreement with human G/T judgments (180 trade-off pairs, label level)0.839 / 0.9060.783 / 0.856
Label-level flip-both0.7440.639

The (4, 2, 1) template does not substantially change the labels overall (98.3% agreement), but it introduces 3.2% weighted-tie samples and substantially reduces the available task-conditioned supervision (30% fewer preference-sensitive pairs and 22% fewer reversal samples). Its constructed task-preference labels also show lower agreement with human G/T judgments. Thus, (8, 4, 1) is selected because it satisfies the lower tie-rate requirement, rather than because the weights were chosen retrospectively to maximize test accuracy.

2. Performance of the trained models.

Under the same backbone, training split, and training configuration, we train PanoReward using each template and compare G/V/T pairwise accuracy, overall pairwise accuracy, and flip-both accuracy on 180 trade-off pairs:

VariantGTVAllFlip-both
Trained with (4, 2, 1) labels0.7140.7700.7670.7510.817
Trained with (8, 4, 1) labels (ours)0.7050.8570.7780.7850.867

3. Implications for task-conditioned preference modeling.

Our goal is not to claim that (8, 4, 1) is the only reasonable weight template. Rather, we aim to verify that task-conditioned reward modeling remains effective under different reasonable priority templates, and to assess whether its performance is overly sensitive to the specific choice of weights.

The results show that the model trained with (4, 2, 1) is slightly better on G (0.714 vs. 0.705), but performs worse on T, V, overall accuracy, and flip-both accuracy (0.770 vs. 0.857, 0.767 vs. 0.778, 0.751 vs. 0.785, and 0.817 vs. 0.867, respectively). Both templates enable the model to learn task-conditioned reward modeling. However, (8, 4, 1) performs better overall and supports more reliable task-conditioned preference switching; its advantage is not limited to a single dimension.

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

This paper presents PanoReward, a task-profile-aware reward model for panoramic image generation. The model predicts separate scores for 3D consistency, polar visual quality, and image-text consistency, and predicts fusion weights from task instructions. The authors also construct PanoReward-Data, a multidimensional preference dataset, and PanoReward-Bench, a human-annotated evaluation benchmark. They evaluate the reward model through offline preference prediction, instruction-following evaluation, and RL of a generative model.

Critical strengths
  1. The model uses separate heads to score 3D consistency, polar visual quality, and image-text consistency (G/V/T), and dynamically predicts their fusion weights from task instructions. Incorporating task preferences into reward modeling is a well-motivated design choice.

  2. The training data emphasizes preference-reversal samples and includes contra cases. These examples encourage the model to learn trade-offs across quality dimensions.

Critical weaknesses
  1. The authors select the weight configuration (8, 4, 1) by maximizing the number of preference-sensitive and preference-reversal samples. However, this criterion primarily concerns the composition of the training data, and its connection to human preferences remains unclear. Could the authors clarify how this selection criterion relates to human preferences?

  2. The motivation for the Gaussian reward parameterization seems reasonable, but its training benefit over directly predicting scalar scores remains unclear for this task. A controlled comparison with a standard Bradley–Terry (BT) objective using scalar rewards would help establish the benefit of this formulation.

  3. Table 4 shows the quality trade-offs under different task weights, but it does not clearly establish the additional value of learned fusion over simpler weighting schemes. Could the authors explain why the learned weighting mechanism is needed?

  4. Typos and formatting: Some citations should use \citep (e.g., Lines 62–63). In addition, the cross-reference in Lines 719–720 is unresolved and appears as “Section ??”.

Questions for authors

see weaknesses

6 comments

Thank you for your insightful and comprehensive feedback. We have carefully studied your comments and argue that your concerns can be addressed.

Weakness1. The authors select the weight configuration (8, 4, 1) by maximizing the number of preference-sensitive and preference-reversal samples. However, this criterion primarily concerns the composition of the training data, and its connection to human preferences remains unclear. Could the authors clarify how this selection criterion relates to human preferences?

Answer. Thank you for raising this important question. We are glad to explain the relationship between the weight-selection criterion and human preferences. Our motivation is that human judgments of panorama quality are inherently multidimensional: several related but potentially conflicting quality dimensions jointly determine the assessment, and their relative importance changes with the downstream task. Therefore, our goal is not to learn a single fixed human-preference score, but to develop a reward model that adapts its quality criteria to task requirements.

Specifically, our design is based on three considerations.

1. Human preferences are first represented through multidimensional quality judgments.

For each panorama, we obtain human-refined scores for 3D consistency (G), polar visual quality (V), and image-text consistency (T). These dimensions characterize spatial structure, visual perception, and semantic correspondence, respectively. They may involve trade-offs: for example, one panorama may have better spatial structure but weaker text-image consistency. We therefore retain the dimension-level human judgments instead of reducing them to a single overall score in advance.

2. Task-conditioned preferences adapt the relative importance of these human-evaluated dimensions.

The final preference depends on the task. For example, 3D scene modeling places greater emphasis on spatial consistency, whereas text-driven scene creation places greater emphasis on semantic and layout accuracy. We use different priority templates to convert the dimension-level human assessments into task-conditioned preference labels:

$ y_{\pi}=\operatorname{sign}\left[w_{\pi}^{\top}(s^{(A)}-s^{(B)})\right]. $

Here, s(A)s^{(A)} and s(B)s^{(B)} denote the human-refined G/V/T scores of two panoramas, and wπw_{\pi} denotes the quality-priority weights for the corresponding task. Importantly, (8, 4, 1) is not intended as a direct estimate of the strength of human preferences for the three dimensions; it is a computable rule for translating qualitative task priorities into quality trade-offs.

3. Preference reversals provide informative counterfactual supervision for learning task-dependent preferences.

This is the main reason we search for the weight template using integer programming. If the same image pair receives the same preference under every task condition, that pair provides limited information about how task instructions affect the final judgment. In contrast, when human assessments conflict across dimensions, the preference for the same pair may reverse as task priorities change. Such cases directly reveal how task requirements alter the relative contributions of different quality dimensions, providing more discriminative supervision for learning task-conditioned preference mappings.

We use integer programming to select a weight template that maximizes the number of preference-sensitive and preference-reversal samples while satisfying the dimension-priority and tie-rate constraints. Thus, (8, 4, 1) is selected to improve the effectiveness of counterfactual task-preference supervision constructed from human G/V/T assessments. It is neither a direct estimate of human preference strength nor a template selected by maximizing accuracy on PanoReward-Bench.


In summary, the connection to human preferences is established at two stages. First, task-conditioned labels are constructed from human-refined G/V/T assessments together with predefined task priorities. Second, we use independent human rankings to evaluate the trained model. On 180 G/T trade-off pairs in PanoReward-Bench, PanoReward achieves 0.867 flip-both accuracy, with 0.917 and 0.900 accuracy under G-priority and T-priority conditions, respectively. These results show that the learned task-conditioned preferences are consistent with independent human evaluations.

Weakness2. The motivation for the Gaussian reward parameterization seems reasonable, but its training benefit over directly predicting scalar scores remains unclear for this task. A controlled comparison with a standard Bradley–Terry (BT) objective using scalar rewards would help establish the benefit of this formulation.

Answer. Following your valuable suggestion, we conduct a controlled comparison in which the backbone, data, training configuration, and preference labels are kept unchanged. We replace Gaussian reward prediction with scalar reward prediction: each quality head outputs a scalar score, and the Fusion Branch produces a weighted scalar score in the same manner. The model is trained with the standard Bradley–Terry loss. The results are as follows:

VariantGTVAllFlip-both
Scalar reward + standard BT0.7240.8650.7330.7820.911
Gaussian reward (ours)0.7050.8570.7780.7850.867

The two parameterizations achieve comparable overall accuracy (0.782 vs. 0.785). The Gaussian formulation performs better on V (0.778 vs. 0.733), while scalar BT performs slightly better on G and flip-both accuracy.

These results also suggest that the main gains of our method come from the task-conditioned multidimensional modeling framework, rather than Gaussian parameterization alone.

Weakness3. Table 4 shows the quality trade-offs under different task weights, but it does not clearly establish the additional value of learned fusion over simpler weighting schemes. Could the authors explain why the learned weighting mechanism is needed.

Answer. We use learned fusion not merely to introduce an additional trainable module, but to enable the model to adjust the relative contributions of different quality dimensions according to natural-language task requirements, rather than relying on manually specified fixed weights.


(1/2) Motivation.

Different tasks evaluate panorama quality from different perspectives. Immersive viewing emphasizes polar visual quality; 3D scene modeling emphasizes cross-view structural consistency; and scene creation emphasizes correspondence between text and layout. With fixed aggregation weights, the model ranks candidates according to the same criteria across all tasks. Allowing the weights to vary with natural-language task instructions enables a single reward model to represent task-specific quality trade-offs, without manually defining a separate weight template for every task.


(2/2) Experiments.

We compare different aggregation strategies.

1. Learned fusion outperforms fixed-weight aggregation.

Under the same task-conditioned preference supervision, Table 8 (in our paper) compares the aggregation strategies:

AggregationGTVAll
Uniform0.3810.8650.7000.660
Fixed (8, 4, 1)0.6860.7220.7220.710
Learned fusion0.7050.8570.7780.785

Compared with fixed weights, learned fusion improves overall accuracy from 0.710 to 0.785. This indicates that predefined weights alone are insufficient to fully align the reward model with human preferences.

2. Learned fusion better captures task-conditioned preference reversals.

On 180 G/T trade-off pairs, fixed-weight aggregation achieves 0.822 flip-both accuracy, whereas learned fusion achieves 0.867. Thus, learned fusion not only improves ordinary quality ranking, but also more reliably switches preferences between candidate panoramas according to task requirements.

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

Paper summary: This paper studied the problem of Panoramic image generation by proposing a reward model named as PanoReward. This reward model evaluate the generated image from three aspects: 3D consistency, polar visual quality and image-text consistency. Along with this reward, a dataset (named as PanoReward-Data) is constructed for training. Experiments verified the effectiveness of the proposed reward.

Critical strengths

Strength:

  1. Overall, the presentation of this paper is good, and it is easy to follow.
  2. The constructed training dataset PanoReward-Data as well as the benchmark dataset will provide useful resources for Panoramic image generation task.
  3. The experiments demonstrate the effectiveness of PanoReward in preference prediction. Weakness:
Critical weaknesses
  1. My major concern is the novelty of this work. The overall framework of combining multiple quality dimensions with instruction-conditioned reward weighting appears intuitive. The paper should better clarify its methodological novelty compared with existing multi-objective reward models and instruction-conditioned preference learning methods.
  2. In my opinion, the major contrition of this work is the dataset. Hence the author needs to discuss whether there are already some similar dataset for this task, if yes, the author needs to compare with them.
  3. The paper should better explain what makes panoramic image generation fundamentally different from conventional image generation and what unique challenges arise in this setting. This discussion is important for motivating the proposed method and establishing a clear connection between the specific challenges of panoramic image generation and the design of PanoReward.
Questions for authors
  1. What makes panoramic image generation fundamentally different from conventional image generation and what unique challenges arise in this setting?
  2. Are there any existing dataset for this task? If yes, the authors need to compare.
6 comments

Thanks for your insightful comments. We list your advice and questions followed by detailed answers, then argue that your concerns can be addressed.

Weakness1. My major concern is the novelty of this work. The overall framework of combining multiple quality dimensions with instruction-conditioned reward weighting appears intuitive. The paper should better clarify its methodological novelty compared with existing multi-objective reward models and instruction-conditioned preference learning methods.

Answer. Thank you for raising the question of methodological novelty. We would like to clarify that PanoReward is not simply a multi-objective reward model with instruction-conditioned weights. Its contribution lies in identifying and modeling the quality dimensions that are specific to panoramic image generation, and in learning how their relative importance changes with downstream task requirements.

Specifically, our method makes three contributions.

1. Quality assessment for panoramic generation requires panorama-specific dimensions.

Existing image reward models mainly focus on general visual quality, aesthetics, and text-image consistency. These criteria alone are insufficient for panoramas, which additionally require cross-view 3D structural consistency and robustness to the distinctive visual distortions near the poles under equirectangular projection (ERP). We therefore introduce three complementary dimensions:

  • G: 3D consistency, measuring structural continuity across viewing directions;
  • V: Polar visual quality, assessing visual distortions in the Zenith/Nadir regions;
  • T: Image-text consistency, measuring the correspondence between generated content and the scene description.

2. The method decouples quality measurement from task preference.

Unlike methods that directly feed task instructions into every assessment head, PanoReward's three quality heads take only the panorama and scene description as input and produce relatively task-independent dimension scores. The task instruction is provided only to a separate Fusion Branch, which predicts the weights of the three dimensions according to the task-specific evaluation criteria. Thus, the measurement of each quality dimension remains stable across tasks, while the task condition controls only its relative contribution to the final reward. This design avoids unnecessary coupling between task preference and the underlying quality representations.

3. Counterfactual preference supervision teaches task-conditioned quality trade-offs.

This is the key aspect we would like to emphasize. Unlike a reward model trained with a fixed preference, PanoReward uses preference changes for the same image pair under different task priorities as supervision. PanoReward-Data contains 242,428 reversal-type counterfactual samples, in which the preferred panorama changes with the task instruction. The model must therefore learn not only the quality differences along individual dimensions, but also how task instructions change the final preference between two candidate panoramas.

The ablations on task-conditioned supervision and learned fusion verify the importance of this design. Removing task-conditioned preference labels and using fixed aggregation weights (8, 4, 1) reduces overall preference prediction accuracy from 0.785 to 0.604 and flip-both accuracy from 0.867 to 0.022. These results demonstrate that it can assess panorama quality while adapting preference rankings to task requirements.

Weakness2 & Question2. In my opinion, the major contrition of this work is the dataset. Hence the author needs to discuss whether there are already some similar dataset for this task, if yes, the author needs to compare with them. Are there any existing dataset for this task? If yes, the authors need to compare.

Answer. We agree that the relationship between PanoReward-Data and existing panorama quality assessment datasets should be made clearer. The data resource in this work serves two related but distinct purposes: (1) supporting multidimensional quality assessment and preference evaluation for panoramas, and (2) providing supervision for training a reward model for panoramic image generation. We compare the datasets from both perspectives.

1. Panorama quality assessment.

Several datasets have been proposed to assess AI-generated panoramas. AIGCOIQA2024 collects subjective ratings along Quality and Text-image Consistency. AIGC-PanoIQA evaluates panorama quality through dimensions including Visual Quality and 3D Consistency. These datasets provide valuable foundations for studying the visual quality and human perception of generated panoramas.

We further compare their evaluation dimensions and supervision formats with those of PanoReward:

DatasetVisual QualityText-image ConsistencyCross-view 3D ConsistencyTask-conditioned Preferences
AIGCOIQA2024✓✓✗✗
AIGC-PanoIQA✓✗✓✗
PanoReward-Data (ours)✓✓✓✓

The main distinction is that PanoReward jointly models polar visual quality, text-image consistency, and cross-view 3D consistency.

2. Reward learning for panoramic generation.

To the best of our knowledge, existing public datasets do not provide multidimensional pairwise preference supervision specifically designed for task-conditioned reward learning on panoramas. Existing panorama quality datasets mainly provide subjective quality scores for individual images. In contrast, this work investigates how such multidimensional assessments can be used to construct preference supervision that adapts to different task requirements.

Specifically, PanoReward-Data is built from 32,655 scene descriptions and four panorama generators, yielding 195,930 same-prompt source image pairs. Each panorama is assigned human-refined quality scores for G/V/T. Based on these annotations, we construct pairwise preference labels under six strict quality-priority settings and one overall evaluation condition, resulting in 943,074 training instances.

3. The key distinction from existing panorama quality datasets.

Existing datasets mainly provide subjective quality assessments for individual panoramas. PanoReward-Data further provides multidimensional comparisons between candidate panoramas from the same scene and converts these dimension-level assessments into task-conditioned pairwise preference labels. The key contribution is not simply the addition of another quality dimension; it is that the preference for the same image pair can change according to task priorities.


In summary, PanoReward-Data supports both multidimensional panorama quality assessment based on G/V/T annotations and the training of reward models that rank candidates according to natural-language task instructions.

Weakness3 & Question1. The paper should better explain what makes panoramic image generation fundamentally different from conventional image generation and what unique challenges arise in this setting. This discussion is important for motivating the proposed method and establishing a clear connection between the specific challenges of panoramic image generation and the design of PanoReward. What makes panoramic image generation fundamentally different from conventional image generation and what unique challenges arise in this setting?

Answer. Thank you for pointing out the need to better motivate the problem setting. Panoramic image generation differs from conventional image generation in that it must preserve not only local visual plausibility, but also the spatial continuity of a complete spherical scene, projection-related visual quality, and the immersive viewing experience. We clarify the resulting challenges below.

1. Distinctive visual distortions caused by spherical projection.

Panoramas are commonly represented by mapping spherical content to a 2D rectangular image using Equirectangular Projection (ERP). This mapping introduces latitude-dependent geometric distortion. In particular, regions near the poles (Zenith and Nadir) are prone to stretching, deformation, and structural artifacts. A familiar analogy is the distortion in a world map. Conventional image quality metrics may not adequately capture these panorama-specific polar distortions. Existing image generation models also have difficulty producing realistic textures in these regions.

2. Cross-view spatial consistency over a 360° scene.

Conventional image generation primarily concerns visual plausibility within a limited field of view. Panoramic generation must additionally ensure structural continuity across different viewing directions. A generated panorama may appear plausible in an individual local view while exhibiting structural breaks, duplicated objects, inconsistent perspective, or stitching artifacts between neighboring views. These issues directly affect the explorability and spatial credibility of the generated scene.

3. Different tasks require different quality trade-offs.

Panoramas can be used for immersive VR/AR viewing, text-driven scene creation, and 3D world modeling. These applications place different emphasis on quality: immersive viewing prioritizes visual quality and polar distortions; scene creation emphasizes correspondence between text descriptions and scene layout; and 3D scene exploration relies more heavily on cross-view structural consistency.

These challenges motivate the three evaluation dimensions in PanoReward: G measures cross-view spatial consistency, V assesses polar visual quality under ERP, and T evaluates semantic consistency between the text description and the generated scene. Furthermore, since different applications prioritize these dimensions differently, Instruction-conditioned Fusion dynamically adjusts their relative contributions according to the task.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.