- Rating
- 3(+1)
- Confidence
- 2(−1)
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Diversity Preference Ranker (DPR), a unified scoring framework for improving diversity in text-to-image generation while preserving image quality. The authors argue that existing methods typically rely on separate diversity and preference scorers, leading to additional computational overhead and limited interaction between the two objectives. They also suggest that independently encoding candidates may fail to fully capture their distinctiveness within a candidate pool.
To address these issues, DPR learns a shared, pool-aware visual representation through cross-candidate self-attention, from which both diversity and preference scores are computed. The framework combines spatial feature distillation, human preference supervision, and listwise selection distillation to guide candidate selection.
The authors evaluate DPR on four accelerated text-to-image generators, comparing it against SGI and other diversity-enhancing methods. The experiments include quantitative comparisons, diversity–preference trade-off analyses, ablation studies, and an extension to noise optimization. The results show modest improvements in diversity and reduced scoring overhead without generator-specific retraining. However, the gains are generally marginal, and text-image alignment, as measured by CLIPScore, consistently decreases compared to SGI.
Critical strengths
-
Clear problem formulation and well-motivated design. The paper addresses a meaningful problem with a clear formulation and a well-structured methodology. The overall framework is easy to follow, and the design choices are generally reasonable.
-
Well-written paper and solid experimental evaluation. The paper is clearly written and carefully organized, with detailed technical explanations. The experiments cover multiple generators and include useful ablation studies and analyses, providing reasonable support for the proposed method.
Critical weaknesses
-
Limited performance improvements. The improvements over existing methods are generally marginal, and some quality metrics even become worse. In particular, CLIPScore consistently decreases across all four generators, raising concerns about whether DPR achieves a better trade-off between diversity and image quality. These limited gains make the practical benefits of the method less convincing.
-
Limited methodological novelty. The main components, including shared representations, cross-candidate self-attention, and listwise distillation, are all relatively established techniques. While their combination is reasonable, I find the overall contribution somewhat incremental and not sufficiently novel for a top-tier conference like ICLR.
Questions for authors
-
Motivation and qualitative comparisons. I am still not fully convinced by the motivation for pool-aware feature encoding. Existing methods already consider candidate relationships through pairwise diversity scores, so I am not sure why incorporating pool context directly into feature representations is necessary. Also, I personally do not find the similarity cluster highlighted in Figure 1 particularly convincing, and I have similar concerns about some examples in the appendix. Could the authors provide a clearer and more detailed explanation, using the existing examples if possible, to demonstrate why the images selected by DPR are actually more diverse and better than those selected by the baseline, and what specific advantages pool-aware encoding provides over independent encoding?
-
Limited improvements and decreased CLIPScore. The performance gains seem relatively marginal, while CLIPScore consistently decreases across all four generators. I understand the authors' explanation that SGI directly uses CLIP similarity for selection, whereas DPR relies on learned human preference scores. However, I am still unsure whether DPR achieves a better overall diversity–quality trade-off, especially when text-image alignment is taken into account. Although Appendices A.10.1 and A.12 provide gradient analyses, it remains unclear how these findings relate to the observed changes in downstream quality and alignment. Could the authors further clarify this relationship and, if possible, demonstrate whether DPR retains a meaningful diversity advantage at comparable levels of text-image alignment?
-
Methodological novelty. The main components of DPR, including shared representations, cross-candidate attention, and listwise distillation, are all well-established techniques. Could the authors further clarify what they consider the key methodological novelty of this work, beyond combining these existing components, and why this particular combination is necessary?
These are the main concerns affecting my current assessment. I am open to increasing my score if the authors can convincingly address these questions, especially regarding the motivation, methodological novelty, and the diversity–quality trade-off.
10 comments
Reviewer X4Fp (2/4)
Diversity-only updates increase diversity and lower preference means; preference-only raises HPS. Combined DPR updates remain close to diversity-only outputs.
CLIP guidance improves all six means over original DPR and DivGen [14], supporting a better observed diversity–preference balance at this operating point.
The relevant gradients are initial-noise rewards (paper Table 13, A.13.1), distinct from shared-feature training gradients (Table 17, A.16).
We measure the weighted initial-noise gradient norms under these optimization settings.
Table R16. Initial weighted gradient norms (PixArt-DMD; first 50 COCO prompts; B=5).
| Gradient term | Median initial weighted norm |
|---|---|
| Diversity | 15.00 |
| Preference | 0.043 |
| CLIP | 0.034 |
The dominant diversity-gradient norm is consistent with the combined method’s diversity-only-like outputs in Table R15. This connects the diagnostic to observed updates, rather than inferring output quality from feature-space gradients.
Selected images also support scarce-data learning.
To test usefulness beyond generation metrics, we augment DTD training under scarce real labels and measure classification on real test images.
Table R17. DTD few-shot classification (10 splits).
| Method | 1-shot accuracy (%)↑ | 2-shot accuracy (%)↑ |
|---|---|---|
| Real-only | 20.47 ± 2.53 | 28.51 ± 1.74 |
| Random | 27.26 ± 1.00 | 32.26 ± 1.27 |
| SGI | 27.30 ± 0.90 | 32.55 ± 0.92 |
| DPR | 27.51 ± 1.15 | 32.96 ± 1.33 |
Values are mean ± standard deviation over 10 splits.
DPR yields the highest mean accuracy in both settings, exceeding SGI [13] by 0.21 and 0.41 percentage points. This connects selection to the practical problem of learning with limited real data.
W2 & Q3) Methodological contribution and component combination
We appreciate the reviewer’s feedback on novelty. DPR shifts the interaction between diversity and preference from score combination alone into representation learning. Its shared, pool-conditioned field learns both criteria and the ranking decisions they jointly support. Matched controls isolate the value of selection supervision and pool context (Tables R18–R19), while the component ablation supports their combined use (Table R20).
(1) A shift from separate scoring to learning a shared representation for joint selection.
SGI [13] uses separately pretrained representations to score diversity and preference, combining their outputs at selection time. DPR brings this interaction into representation learning: a shared, pool-conditioned feature field learns both criteria and the ranking decisions induced by their combined score.
Spatial and human-preference supervision teach the individual criteria; listwise distillation teaches how they work together when selecting the next image. Drawing on listwise learning-to-rank [3, 4], DPR trains the representation through the joint decisions it supports, beyond consolidating two scorers into a smaller network.
(2) Learning combined decisions improves diversity beyond criterion-level supervision.
The matched comparison changes only selection supervision; attention, criterion-level supervision, and training budget are fixed.
Table R18. Matched comparison of selection supervision (PixArt-DMD; 200 COCO prompts; B=5).
| M | Selection supervision | Vendi↑ | MSS↓ | LPIPS↑ | HPS↑ | Pick↑ | CLIP↑ |
|---|---|---|---|---|---|---|---|
| 30 | None | 3.8562 | 0.3927 | 0.5271 | 28.1336 | 22.0661 | 31.0149 |
| 30 | Pairwise | 3.8698 | 0.3904 | 0.5311 | 28.1302 | 22.0543 | 30.9764 |
| 30 | Listwise | 3.8702 | 0.3903 | 0.5280 | 28.1140 | 22.0503 | 30.9983 |
| 64 | None | 3.9065 | 0.3799 | 0.5380 | 28.1180 | 22.0579 | 30.9587 |
| 64 | Pairwise | 3.9206 | 0.3772 | 0.5414 | 28.0502 | 22.0330 | 30.9018 |
| 64 | Listwise | 3.9245 | 0.3768 | 0.5406 | 28.1028 | 22.0473 | 30.9416 |
| 128 | None | 3.9455 | 0.3707 | 0.5476 | 28.0822 | 22.0400 | 30.9561 |
| 128 | Pairwise | 3.9694 | 0.3646 | 0.5519 | 27.9905 | 22.0261 | 30.9445 |
| 128 | Listwise | 3.9728 | 0.3641 | 0.5506 | 28.0133 | 22.0319 | 30.9361 |
Reviewer X4Fp (3/4)
Bold marks the best mean within each pool size. These matched runs use one training seed and the same 60,000-step budget, separately from the original DPR checkpoint.
Both pairwise and listwise supervision improve all three diversity means over criterion-only supervision. Listwise leads Vendi and MSS; pairwise leads LPIPS. These results support learning combined selection decisions, with small reductions in preference and alignment means.
We next compare intra-image and cross-candidate encoding at matched parameter counts.
Table R19. Parameter-matched encoding comparison (PixArt-DMD; 200 prompts; M=30, B=5; 6.1M parameters per model).
| Metric | Intra-image attention | Cross-candidate DPR | Paired Δ 95% CI |
|---|---|---|---|
| Vendi↑ | 3.8676 | 3.8934 | [+0.00612, +0.04608] |
| MSS↓ | 0.3907 | 0.3852 | [−0.01002, −0.00105] |
| LPIPS↑ | 0.5284 | 0.5312 | [+0.00038, +0.00524] |
Cross-candidate encoding improves all three diversity means at matched capacity, supporting adaptation of the comparison features to the candidate pool.
(3) The integrated design achieves the highest mean Vendi and HPS.
The component ablation compares full DPR with the plain student and variants removing pool context or listwise supervision.
Table R20. Component ablation (selected variants from paper Table 3).
| Variant | Listwise supervision | Vendi↑ | HPS↑ | Pick↑ |
|---|---|---|---|---|
| Plain student | Off | 3.8487 | 28.0952 | 22.0547 |
| DPR w/o listwise supervision | Off | 3.8547 | 28.1082 | 22.0557 |
| Intra-image attention | On | 3.8676 | 28.0988 | 22.0616 |
| Full DPR | On | 3.8934 | 28.1227 | 22.0560 |
Full DPR achieves the highest mean Vendi and HPS, while intra-image attention leads PickScore. Together with the matched controls and representation-sharing ablation (paper Table 2), this supports learning pool-aware features and combined decisions within one scorer.
Q1) Role of pool-aware encoding and qualitative comparisons
We thank the reviewer for the comments on encoding and visual comparisons. Pairwise scoring compares features; pool-aware encoding adapts those features to the available alternatives. The matched-capacity comparison supports this distinction (Table R19), while Figs. 1 and 3 and the user study show what changes in the choices presented to users (Table R21).
(1) Pool context changes the features used for comparison.
An independently encoded image has the same features in every pool. Yet a side view is a distinctive alternative among front views and a redundant one among similar side views. DPR lets surrounding candidates shape the features before computing pairwise distances.
DLCM [5] and SetRank [6] similarly use candidate-list context to refine ranking representations. Table R19 compares cross-candidate DPR with intra-image attention at matched parameter counts.
At matched capacity, cross-candidate encoding improves all three diversity metrics, with paired intervals supporting each gain (Table R19).
The complete selector also shows larger mean diversity and preference advantages in larger pools (W1, Table R14).
(2) Users slightly favor the more varied alternatives.
With hidden labels and randomized presentation, participants judged diversity, faithfulness, and final-image selection preference.
Table R21. COCO user study (12 participants; 50 prompts; M=30, B=5).
| Criterion | DPR wins | Ties | SGI wins | DPR preference score |
|---|---|---|---|---|
| Perceived diversity | 245 (40.8%) | 129 (21.5%) | 226 (37.7%) | 51.6% |
| Prompt faithfulness | 190 (31.7%) | 248 (41.3%) | 162 (27.0%) | 52.3% |
| Final-image selection preference | 264 (44.0%) | 96 (16.0%) | 240 (40.0%) | 52.0% |
Wins receive 1, ties 0.5, and losses 0 in the preference score.
Participants slightly favored DPR in all three criteria, with perceived diversity consistent with the quantitative trend.
Original DPR selects fewer highlighted-cluster images in Fig. 1, offering circular foreground structures and elongated arrangements along the water. Fig. 3(a) varies subject scale and framing; Fig. 3(b) varies cake decoration and frosting texture.
We thank the reviewer again for the constructive feedback, which helped sharpen the methodological contribution and connect the scoring design to the observed results.
Reviewer X4Fp (1/4)
We thank Reviewer X4Fp for recognizing the clear problem formulation and solid evaluation. We clarify DPR’s shift from separate scoring to shared representation learning for joint selection, and address the diversity–alignment trade-off and the role of pool context.
We thank the reviewer for the comments on alignment and gradient analysis. (i) With the same CLIP preference term as SGI [13], DPR improves all reported means on DrawBench and all except LPIPS on COCO (Table R13). (ii) We argue that larger-pool selection reveals a stronger diversity–preference benefit at fixed final set size (Table R14). (iii) Actual noise updates connect the gradient measurements to output changes: CLIP guidance improves preference and alignment means over original DPR while retaining its diversity gains (Tables R15–R16).
For effect-size context, SGI reports a Vendi gain of +0.024 over NegToMe on SD3 (3.534→3.558) [13, Table 4], and CNO reports +0.0155 over CADS on SDXL (2.8864→2.9019) [7, Table 1]. DPR’s larger-pool gains are +0.0275 and +0.0210, versus +0.0099 at M=30. Although protocols differ, these examples place the increments in context; DPR additionally achieves them with a 6.1M scorer.
(1) Diversity benefits persist under the same text-aware preference term.
SGI [13] directly rewards CLIP similarity, introducing a selection bias toward higher CLIPScore; we therefore hold this preference term fixed and compare the frozen DPR diversity distances under identical pools and selection budgets.
Table R13. COCO and DrawBench alignment controls (PixArt-DMD; 200 prompts per dataset; M=30, B=5).
Bold marks the best mean within each dataset. Operating points were selected on each benchmark’s evaluation prompts.
CLIP+DPR improves all reported means on DrawBench and all except LPIPS on COCO, retaining diversity benefits at higher measured alignment.
(2) Larger pools strengthen the diversity–preference advantage.
Larger pools expose more alternatives at fixed B, making selection more consequential. We compare identical pools using diversity metrics and HPSv2, PickScore, and CLIPScore.
Table R14. PixArt-DMD pool-size comparison (COCO 200 prompts; B=5).
The Vendi advantage increases from +0.0099 at M=30 to +0.0275 and +0.0210 at M=64 and 128. The HPS advantage grows from +0.0321 to +0.0529 and +0.0960, alongside better MSS, LPIPS, and PickScore means.
(3) Actual noise updates connect gradient balance to output trade-offs.
We isolate diversity, preference, and combined objectives under identical initial noise and update budgets.
Table R15. Matched noise optimization (PixArt-DMD; COCO 100 prompts; B=5; identical initial noise; 10 updates).