- 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
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.
| Dimension | Out-of-domain vs. out-of-domain | Out-of-domain vs. in-domain | In-domain vs. in-domain |
|---|---|---|---|
| G, cross-view consistency | 0.556 | 0.436 | 0.596 |
| V, polar visual quality | 0.756 | 0.858 | 0.900 |
| T, text consistency | 0.976 | 0.921 | 0.889 |
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 units | 0 (0%) | 37,948 (3.2%) |
| Preference-sensitive source pairs | 127,734 (65.2%) | 89,027 (45.4%) |
| Reversal-type counterfactual units | 242,428 | 189,189 |
| Agreement with human G/T judgments (180 trade-off pairs, label level) | 0.839 / 0.906 | 0.783 / 0.856 |
| Label-level flip-both | 0.744 | 0.639 |
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:
| Variant | G | T | V | All | Flip-both |
|---|---|---|---|---|---|
| Trained with (4, 2, 1) labels | 0.714 | 0.770 | 0.767 | 0.751 | 0.817 |
| Trained with (8, 4, 1) labels (ours) | 0.705 | 0.857 | 0.778 | 0.785 | 0.867 |
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.
Thanks for your insightful comments. We list your advice and questions followed by detailed answers, then argue that your concerns can be addressed.
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:
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:
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.