acceptodds
Under review as a conference paper at ICLR 2027

When More Search Reduces Coverage: Population-Aware Diffusion Search

Abstract

More verifier-guided diffusion search need not improve the generated population. On class-conditional ImageNet, increasing branching beyond an initial gain raises Top- ADM precision while reducing recall and worsening FID. Retained paths become parents of later candidates, linking selection to future exploration. We investigate this feedback with retention-schedule interventions and introduce Moment-QUBO, a rule that balances verifier quality with matching the candidate population's first two Inception-feature moments. Applying it throughout search improves FID beyond replacing only the final selection, showing that early retention decisions matter beyond terminal reranking. Across two independently seeded SiT-XL/2 curves and a DiT-XL/2 transfer, its advantage over Top- widens with branching at matched denoiser evaluations. At the largest SiT budget, Moment-QUBO reduces FID by over 32% relative to Top- in both runs. Benefits depend on representation and operating point: CLIP-CMMD favors determinantal selection in the matched-precision SiT comparison, and ordinary DiT sampling offers lower FID and observed cost at lower ADM precision. These findings identify population retention as a design choice for mitigating adverse scaling in repeated diffusion search.

Primary area:
generative models
Keywords:
diffusion models, inference-time scaling, population-aware search, quadratic unconstrained binary optimization (QUBO)

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BKgZOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors observe on ImageNet that, as the branching factor increases, Top-K intermediate candidate selection in verifier-guided diffusion inference-time search leads to higher precision but lower recall and worse FID in the final output set. To mitigate this issue, the authors propose a set-based candidate selection objective: while selecting candidates with high verifier scores, it encourages the first- and second-order statistics of the selected subset in Inception feature space to match those of the entire current candidate population. The experiments show that the proposed method can improve the FID of the final output set relative to Top-K.

Critical strengths

none

Critical weaknesses
  1. The motivation for studying this search mechanism is insufficient.

The authors take fixed-width pruning at each round as given and investigate the effects of different candidate retention rules under this setting. However, this does not establish that repeated global pruning with a constant width of 50 is a reasonable or effective computational allocation strategy. The existing experiments do not explore alternative search-budget allocation strategies or compare against other search strategies. Meanwhile, Table 2 shows that ordinary DDIM25 sampling achieves an FID of 9.544 and recall of 0.369, compared with 12.162 and 0.238 for Moment-QUBO. The latter increases ADM precision from 0.912 to 0.935, approximately 2.3 percentage points, while increasing the observed inference time per 50 images from 23.96 to 64.78 seconds. Thus, this increase in precision comes at the cost of nearly three times the observed inference time, a higher FID, and substantially lower recall. It therefore remains unclear whether this quality–cost trade-off is practically favorable relative to ordinary sampling. Mitigating degradation relative to Top-K within the prescribed search framework is therefore insufficient to establish the practical utility of the framework itself.

Moreover, practical generation scenarios often involve generating a single target image under a given condition. The search method used in the paper requires maintaining K=50K=50 trajectories. Thus, for DiT at B=2B=2, a full search population requires approximately 50×2×(10+23×2)=560050\times2\times(10+23\times2)=5600 image-level denoising network evaluations, compared with 50 evaluations for generating a single image using ordinary DDIM sampling—a roughly 112-fold increase in cumulative denoising computation. The use cases that justify this additional computation remain unclear.

  1. The proposed method is primarily driven by the evaluation metric, which limits the interpretation of the results.

It constructs an FID-motivated proxy by matching the first- and second-order moments of Inception features, while FID is also the central metric used to demonstrate its effectiveness. This makes it difficult to distinguish whether the gains reflect broader improvements in generation quality or primarily improved distribution matching in Inception space. The authors’ experiments also support this concern: for example, in Table 27, DPP performs better on CLIP-CMMD. Evaluations directly related to other downstream task requirements are currently lacking, leaving it unclear whether the gains correspond to broader improvements in generation quality.

  1. The manuscript is not self-contained.

The manuscript retains a substantial amount of internal experimental and automated workflow language, including terms describing conditional termination or acceptance, such as “gate passes” and “image gates,” as well as various manifests, source hashes, and generated files. The final “Submission checks” section, in particular, shows traces of automated checking. Moreover, the paper repeatedly refers to materials that are not visible in the manuscript, and the authors have not submitted the relevant materials.

Questions for authors
  1. What justifies the use of a fixed-width, repeated global-pruning strategy, given that ordinary DDIM sampling achieves better FID and recall at substantially lower cost in Table 2? What intended application or output requirement makes this trade-off worthwhile?

  2. Given that DPP achieves better CLIP-CMMD at closely matched precision in Table 27, more evidence is needed to establish that the proposed method provides broader improvements in generation quality beyond improved distribution matching in Inception space.

  3. The authors should compare the proposed repeated fixed-width global pruning strategy with alternative search or budget-allocation strategies under matched computational budgets. Such comparisons are necessary to justify the choice of this particular search design.

3 comments

Thank you for the detailed comments. We have clarified the output requirement and added comparisons that separate the value of search from the value of the retention rule.

W1 / Q1: Motivation for fixed-width repeated pruning

The application we consider is delivering several condition-following candidates under the same prompt. The SDXL study provides direct task evidence: dynamic Moment obtains 63.50% GenEval macro success versus 54.78% for DDIM168 at equal denoiser NFE, a difference of 8.72 percentage points (conditional 95% CI [5.18, 12.38]). This measures the complete search strategy relative to ordinary sampling.

The benefit attributable to changing retention is tested separately against dynamic Top-K under the same search settings. Moment increases LPIPS from 0.189 to 0.398 with the same 649/1,024 passing images; their GenEval difference is unresolved. This supports using a population-aware rule to reduce concentration when search is used for condition-following generation.

For DiT, DDIM25 at CFG 3 obtains FID 9.544 and recall 0.369 in 23.96 s per 50 outputs, compared with 12.162, 0.238 and 64.78 s for dynamic Moment. ADM precision is 0.91232 versus 0.93484. Sections 6, 7.1 and 8 now present both the favorable condition-following results and the cases where simpler sampling gives better coverage and latency.

The 112-fold computation comparison

Our DiT search returns 50 images per condition. At B=2 it requires 5,600 image-level NFE for the complete population, averaging 112 NFE per delivered image. Ordinary DDIM25 uses 50 NFE per image under the same convention of counting both classifier-free-guidance rows. The ratio is therefore 2.24 when all 50 outputs are delivered. This is an average population cost, not the cost of running the search to obtain only one image.

If only one output is required and the full search population is generated, the approximately 112-fold comparison with a single ordinary image is applicable. Our experiments do not establish that this single-output use is worthwhile. The revised introduction explicitly states the multi-output setting, and the SDXL study uses a smaller set of eight images per prompt.

W2 / Q2: Gains beyond Inception-space distribution matching

The SDXL experiment evaluates retention outside Inception space: it uses DINOv2 features, online CLIP similarity and GenEval condition-satisfaction judgments. Relative to dynamic Top-K, Moment increases all-output LPIPS by 0.2088 (conditional 95% CI [0.1840, 0.2330]). Among 83 prompts with at least two passing outputs under both policies, LPIPS among passing images increases by 0.1901 ([0.1588, 0.2211]).

Against dynamic DPP, all-output ImageReward favors Moment by 0.0553 ([0.0152, 0.0992]), while the LPIPS difference is unresolved. LPIPS measures perceptual separation, and ImageReward is a learned preference score; neither directly measures semantic coverage.

In the independent matched-precision SiT comparison, Moment improves FID from 7.516 to 6.429 and ADM recall from 0.3922 to 0.4125, while CLIP-CMMD favors DPP: 0.33535 versus 0.34791. The Moment-minus-DPP CLIP-CMMD difference is +0.01256 ([0.00910, 0.01584]). Section 7.1 therefore limits that advantage to Inception-space metrics, and Appendix K.3 separately reports the SDXL task and perceptual outcomes.

W3: Internal workflow language and unavailable material

We have removed internal workflow terminology and the submission-check narrative, and made the method and evaluation protocols explicit in the manuscript.

Section 3 gives a complete search round: expand the retained parents, obtain predicted-clean images and features, form the current-candidate reference, jointly select K states, and retain them as the next parents. The objective and greedy-plus-swap solver are specified directly, with derivations in the appendix.

Section 4 and Table 1 identify the early policy producing each final candidate pool and the selector applied to it. The full-class, mixed-feature crossover and the smaller same-feature diagnostic are distinguished. Appendix K.3 gives the SDXL time grid, re-noising, DDIM updates, clean previews, feature preprocessing, prompt split and NFE accounting. Section 8 reports end-to-end timing separately from denoiser NFE.

We have prepared the revised PDF and an anonymous supplementary archive to accompany the response. The archive contains selection code, generation and evaluation entry points, configurations, prompt splits, tests and recorded numerical results. README.md covers installation and minimal use, REPRODUCTION.md maps experiments to commands, and EVALUATION.md gives metric environments and conventions.

External model weights and evaluation assets must be obtained separately; full image archives and large activation caches are not bundled. Package checks cover CPU tests and recorded-result calculations, rather than a clean-machine regeneration of every experiment. These requirements are stated in the documentation.

Q3: Alternative search and budget-allocation strategies

Table 2 compares longer trajectories, independent generation followed by selection, and repeated branching at 5,600 NFE per 50 DiT outputs. DDIM56 at CFG 1.5 achieves FID 2.222 and recall 0.599; IID112 plus Moment achieves 4.321 and 0.484; dynamic Moment achieves 12.162 and 0.238. Their precisions are 0.802, 0.866 and 0.935. Top-K and DPP are included for independent and dynamic selection. The resulting trade-off favors simpler allocations on FID and recall at these settings.

Table 3 compares DDIM168, IID32 DDIM42 plus selection and dynamic SDXL search at 2,688 NFE per eight images. IID Moment has greater LPIPS, while dynamic Moment has higher ImageReward; their GenEval difference is unresolved. The selected IID policies differ from dynamic Moment by approximately 0.011 in mean development CLIP score.

Controlled timing with the same latent microbatch gives 182.92 s for DDIM168, 197.98 s for IID Moment and 257.27 s for dynamic Moment. This study uses two development prompts, each measured twice per method (four measurements per method).

Appendices K.2 and K.4 additionally vary population width and the rounds using Moment. Appendix K.5 tests a small shared-descendant allocation comparison: feedback changes final group count by +0.0625 relative to fixed branching, with an interval spanning zero, and provides no demonstrated improvement over static allocation.

Thank you for the rebuttal, but I still have some concerns.

  • The case for repeated fixed-width pruning remains unclear. The rebuttal acknowledges that there are settings where “simpler sampling gives better coverage and latency.” At the same denoiser NFE, both longer independent trajectories and independent generation followed by selection achieve lower FID and higher recall than dynamic Moment in the DiT experiments (Table 2). It remains unclear which application would benefit enough from the higher ADM precision to justify the loss of coverage and the added latency observed against DDIM25 (Appendix Table 43).

  • The multi-image results weaken the proposed use case. A full DiT search at B=2 uses about 112×112\times as much denoising compute as DDIM25 if only one image is delivered, or $2.24\times$$ per image on average if all 50 are delivered. The value of delivering 50 images depends on how many satisfy the condition and offer meaningfully different choices. In the SDXL experiment, dynamic Moment produces at least two GenEval-passing images with pairwise LPIPS above 0.4 on 60.9% of prompts, compared with 75.0% for independent sampling + Moment at the same denoiser NFE and 73.4% for the lower-cost DDIM40 (Appendix Table 52). Nearly 40% of prompts fail to yield even one qualifying pair under dynamic Moment. Both simpler methods have higher observed delivery rates at all three exploratory thresholds. These results leave the practical benefit of repeated search unclear even in the proposed multi-image setting.

  • Evidence beyond Inception space remains limited. At closely matched ADM precision in the SiT comparison, Moment achieves lower FID and higher ADM recall, while DPP achieves lower CLIP-CMMD (Table 4). This leaves open whether the gains measured in Inception space translate into broader improvements in semantic coverage or overall generation quality.

Therefore, I maintain the original score.

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

This paper aims to address the problem that increasing branching in diffusion inference-time search can improve ADM precision while reducing coverage. It proposes a population-aware search method that preserves the first- and second-order feature moments of the candidate population, improving FID compared with conventional Top-K search. Experiments on ImageNet with SiT and DiT models demonstrate its effectiveness in mitigating coverage degradation, with additional comparisons against diversity-aware selection methods such as DPP.

Critical strengths
  1. The paper provides extensive experiments to evaluate the proposed method, as presented in both the main text and appendices. These include experiments across different model architectures, ablation studies, and evaluations under different inference compute budgets. These thorough experiments provide a better understanding of the proposed method and strengthen the reliability of the paper.

  2. The paper addresses the problem of coverage degradation caused by search algorithms in diffusion generation. The authors demonstrate the importance of this problem and show that their proposed method can improve FID with only a small additional computational cost compared with Top-K search, making the approach more applicable in practice.

Critical weaknesses
  1. As shown in Figure 4, compared with DPP, Moment-QUBO achieves limited improvements in FID and still can not mitigate the trend of increasing FID as NFE grows. These limited performance gains weaken the overall contribution of the proposed method.

  2. Diversity-aware selection in diffusion inference has already been explored in prior work, such as DPP-based methods. The proposed method primarily extends existing diversity-aware selection frameworks with a moment-matching objective, leaving its novelty limited.

Questions for authors
  1. Prior work such as Group Inference has already explored joint quality-diversity optimization and intermediate candidate pruning. Could the authors clarify the main methodological novelty of Moment-QUBO beyond introducing a different population-level selection objective?
  2. Section 6.1 shows that Moment-QUBO outperforms DPP at matched ADM precision, while Appendix C.12 reports nearly overlapping FID–recall frontiers on a fixed candidate pool. How do the authors explain this difference, and what does it suggest about the source of Moment-QUBO's advantage?
  3. The main experiments fix the population size at K = 50, although population coverage is central to the proposed method. Could the authors clarify how the choice of K may affect the observed coverage degradation and the relative advantage of Moment-QUBO over Top-K?
3 comments

Thank you for the positive assessment of the experiments and the low incremental cost of changing retention.

W1: Improvement over DPP and the branching trend

At closely matched ADM precision, the independent SiT comparison gives a 14.5% FID reduction over Inception-DPP: 6.429 versus 7.516, with a precision difference of +0.00016. The FID difference is −1.087 (conditional 95% CI [−1.522, −0.713]); recall increases by 0.0203 ([0.0105, 0.0298]). The conditional precision interval lies within the stated ±0.005 matching margin. The calibrated DiT comparison gives a 6.0% FID reduction, from 12.941 to 12.162, at a precision difference of +0.00004.

On SiT-original at B=4, uncalibrated Inception-DPP has lower FID and higher recall, but precision 0.777 versus 0.847 for Moment. Section 7.1 presents this result before the calibrated comparisons. It also reports that CLIP-CMMD favors DPP on SiT, limiting the favorable comparison to Inception-space metrics.

The branching trend itself remains: Moment mitigates deterioration relative to Top-K but does not reverse it. Section 5 and Appendix E state this explicitly.

W2 / Q1: Novelty beyond Group Inference

The methodological distinction is the current-candidate moment target; the empirical contribution is separating early-policy effects from terminal reranking. Moment penalizes changes to the statistics of the population being pruned, balancing that discrepancy against verifier quality. Pairwise repulsion directly rewards separation within the selected set. The candidate reference is refreshed after each expansion.

Joint selection and intermediate pruning are established ideas, which we credit explicitly. Section 7.2 tests the objective difference using Group-Inference-style repulsion under the same features, verifier, search framework and solver. Independent evaluations with 50K images per policy yield FID 7.496 for repulsion and 6.568 for Moment, at a precision difference of −0.00016. This controls the objective within our framework rather than reproducing the full Group Inference pipeline.

Mean-only matching achieves FID 6.653 and captures most of the gain. Second moments provide a clearer recall benefit than FID benefit, as stated in the revised abstract.

Q2: Overlapping fixed-pool frontiers versus different dynamic results

A fixed-pool comparison changes selection from completed images. Dynamic retention also changes which parents generate later candidates. Table 1 tests this distinction by applying both final selectors to each of two final-round pools.

Let D and M denote the DPP and Moment policies; the first letter specifies the early policy and the second the final selector. Each cell delivers 50K class-balanced images at the same generation budget: DD has FID 7.172, DM 6.773, MD 6.850 and MM 6.455.

With final DPP fixed, changing the early policy lowers FID by 0.323; with final Moment fixed, it lowers FID by 0.318. The approximate conditional 95% intervals are [−0.592, −0.053] and [−0.567, −0.076]. Differences use unrounded metrics. Replacing the final selector within a fixed pool gives a separate FID change of about −0.40, accompanied by a precision shift.

In the full-class crossover, the D policy uses DINO features and M uses Inception features, so this is a complete-policy effect. Appendix K.4 adds a same-Inception crossover on 16 development classes with the same FID direction; that smaller diagnostic does not provide an independent full-class confirmation.

There is also a reference-target distinction: the fixed-pool sweep in current Appendix C.11 uses global real-reference moments, whereas dynamic retention uses current-candidate moments. Thus the frontier comparison alone cannot attribute the dynamic advantage to early retention. The crossed intervention is the relevant evidence for that conclusion.

Q3: Effect of population size K

Appendix K.2 varies K over 10, 25 and 50 at B=2 and B=4. Each setting delivers 50 images per class using five, two or one independent parent groups, respectively. The same 128 classes are evaluated, with matched denoiser NFE between selectors within each K/B pair.

Moment has lower FID and higher recall than Top-K in all six pairs. At B=4, diagnostic FIDs for Moment versus Top-K are 76.53 versus 80.47 at K=10, 78.99 versus 82.86 at K=25, and 80.75 versus 84.31 at K=50. Both methods deteriorate from B=2 to B=4 at every tested width.

The retention effect therefore extends beyond K=50. These class-restricted FIDs are not numerically comparable with full ImageNet-50K FID. Changing K also changes the number of independent groups and their ancestry structure, so the experiment measures width sensitivity under a fixed output count rather than identifying an optimal K or isolating solver size.

Thank you for the detailed clarifications. The response addresses most of my empirical concerns and well clarifies the distinction between fixed-pool selection and dynamic retention. While my concern about the limited methodological novelty remains, I will maintain my positive rating. I have no further questions.

Thank you for your thoughtful review and for taking the time to consider our responses. We appreciate your constructive feedback.

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

This paper investigates the degradation of population coverage in repeated verifier-guided diffusion search. The authors observe that increasing the branching factor can improve sample precision while reducing recall and worsening FID. To address this issue, they propose Moment-QUBO, a population-aware selection method that balances verifier scores with first- and second-moment matching in Inception feature space. Experiments on class-conditional ImageNet generation using SiT and DiT demonstrate that the proposed method improves FID and recall over conventional Top-K selection, particularly at larger branching factors.

Critical strengths
  1. The paper identifies an interesting phenomenon in repeated diffusion search: increasing the search budget does not necessarily improve the generated population and may instead reduce its coverage.

  2. The proposed method is relatively simple and does not require retraining the underlying diffusion model.

  3. The experimental evaluation is reasonably comprehensive, including different branching factors, independent random seeds, comparisons with alternative selection methods, and ablation studies. The experiments also provide evidence that intermediate selection decisions influence subsequent exploration..

Critical weaknesses
  1. Limited motivation and practical significance. The paper focuses on mitigating population coverage degradation caused by repeated verifier-guided diffusion search. However, the practical necessity of such a computationally expensive search framework is not sufficiently established. In the DiT experiments, ordinary DDIM sampling achieves better FID (9.544 vs. 12.162) and recall (0.369 vs. 0.238) than Moment-QUBO, while requiring substantially less inference time (23.96 s vs. 64.78 s per 50 images). Although Moment-QUBO achieves higher precision, it remains unclear whether this improvement justifies the additional computational cost. Consequently, the practical significance of improving population retention within this search framework is questionable.

  2. Insufficient comparison with simpler inference-time scaling strategies. The paper does not adequately establish why repeated branching search should be preferred over simply increasing the number of diffusion sampling steps or generating more independent samples. A systematic comparison under matched computational budgets would be important for demonstrating that the proposed approach provides meaningful benefits beyond these simpler alternatives.

  3. Limited methodological novelty. The proposed method essentially combines verifier-based quality selection with population-level diversity regularization. Similar quality-diversity trade-offs have already been explored in prior work, including DPP-based selection and Group Inference. Furthermore, the ablation results indicate that first-moment matching already accounts for most of the FID improvement, leaving the necessity of the full first- and second-moment formulation insufficiently justified.

  4. Limited generalizability of the experimental evaluation. The experiments primarily focus on class-conditional ImageNet generation using SiT and DiT. It remains unclear whether the observed coverage degradation and the benefits of Moment-QUBO generalize to modern text-to-image generation or other conditional generation tasks. Moreover, the method optimizes population statistics in Inception feature space, while some alternative representation-based metrics favor competing selection methods, suggesting that its advantages may depend on the evaluation representation.

Questions for authors
  1. What is the practical motivation for repeated verifier-guided diffusion search? The DiT experiments show that ordinary DDIM sampling achieves better FID and recall at substantially lower computational cost than Moment-QUBO. Although the proposed method improves precision, it remains unclear in which practical scenarios this improvement would justify the additional inference cost. Could the authors clarify the intended applications and the advantages of the underlying search framework?

  2. Why should repeated branching search be preferred over simpler inference-time scaling strategies? Increasing the number of denoising steps or generating additional independent samples are straightforward alternatives for utilizing more inference compute. Could the authors provide comparisons under matched computational budgets to demonstrate that repeated search with Moment-QUBO offers meaningful advantages over these alternatives?

  3. What is the necessity of second-moment matching in the proposed method? The ablation results suggest that first-moment matching already accounts for most of the FID improvement, while the additional benefit of second-moment matching appears relatively limited. Could the authors further clarify its contribution and whether the added complexity is justified?

  4. How well does the proposed method generalize beyond the current experimental setting? The evaluation primarily focuses on class-conditional ImageNet generation using SiT and DiT. Could the authors investigate whether similar coverage degradation occurs in modern text-to-image generation models and whether the proposed method provides consistent improvements across different feature representations and evaluation metrics?

2 comments

Thank you for recognizing the coverage-degradation phenomenon and the evidence that intermediate retention affects subsequent exploration.

W1 / Q1: Practical motivation and computational cost

Our intended application is generating several condition-following candidates for a user to choose from. The revised SDXL study separates two questions: whether search helps condition satisfaction, and whether Moment improves the population retained by search.

On 128 held-out prompts, dynamic Moment achieves 63.50% GenEval macro success versus 54.78% for DDIM168 at equal denoiser NFE, a gain of 8.72 percentage points (conditional 95% CI [5.18, 12.38]). This is an end-to-end search-policy comparison. To isolate the choice of retention rule within search, dynamic Moment and Top-K use the same search settings: Moment increases all-output LPIPS from 0.189 to 0.398, with the same 649/1,024 GenEval-passing images. The GenEval difference between these dynamic policies is unresolved. Thus the specific benefit of Moment is greater perceptual separation at the observed passing-image count, rather than the entire gain over ordinary sampling.

The DiT trade-off you identify remains: ordinary DDIM25 at CFG 3 has better FID, recall and latency than dynamic Moment, with ADM precision lower by 2.25 percentage points. Section 7.1 and Section 8 report the quality and latency comparisons; Appendix J, Table 43 gives the complete ordinary-DDIM25 comparison.

W2 / Q2: Longer trajectories and independent candidates

Table 2 and Appendix K.1 compare DiT policies at 5,600 image-level NFE per class, delivering 50 images each:

  • DDIM56, CFG 1.5: FID 2.222, precision 0.802, recall 0.599.
  • DDIM56, CFG 3: FID 9.892, precision 0.915, recall 0.372.
  • IID112 DDIM25 + Moment: FID 4.321, precision 0.866, recall 0.484.
  • Dynamic B=2 Moment: FID 12.162, precision 0.935, recall 0.238.

Top-K and DPP are also evaluated on the shared independent candidate bank and in dynamic search. The simpler allocations attain better FID and recall; repeated search reaches higher ADM precision.

Table 3 makes the corresponding SDXL comparison at 2,688 NFE per eight outputs. IID Moment achieves greater LPIPS than dynamic Moment (0.625 vs. 0.398), whereas dynamic Moment has higher ImageReward (0.979 vs. 0.862); their GenEval difference is unresolved. The selected IID policies differ from dynamic Moment by approximately 0.011 in mean development CLIP score.

Section 8 reports controlled SDXL delivery times of 257.27 s for dynamic Moment, 197.98 s for IID Moment and 182.92 s for DDIM168. These measurements include decoding and online features, which denoiser NFE excludes.

W3: Relation to existing quality–diversity methods

Our contribution combines a current-candidate moment target with experiments separating early retention from terminal reranking. At each round, Moment-QUBO penalizes changes to the first and second feature moments of the expanded candidate population while rewarding verifier quality. Pairwise repulsion instead rewards separation within the retained set. The reference is refreshed after each expansion, making the selection target specific to the population being pruned.

Joint quality–diversity selection and intermediate pruning are established components, and the revision credits Group Inference and particle-based methods explicitly. Section 7.2 tests the objective difference using a Group-Inference-style quadratic cosine-distance repulsion control inside the same four-round SiT search. Features, verifier and greedy-plus-swap solver are shared; the repulsion weight is calibrated on development classes to match the first-round retained verifier score.

In an independent evaluation with 50K images per policy, Moment achieves FID 6.568 versus 7.496 for repulsion. The difference is −0.928 (conditional 95% CI [−1.217, −0.656]); recall increases by 0.0106 ([0.0014, 0.0196]), with a precision difference of −0.00016. This is an objective-level control within our framework; the complete Group Inference pipeline also differs in candidate evolution, verifier and optimization procedure.

Section 4 tests the second contribution. Crossing early policies with final selectors shows that early retention affects final FID even when terminal selection is fixed. The full-class crossover compares policies with different feature representations; Appendix K.4 separately reports the smaller same-feature development comparison.

Q3: Contribution of second moments

The independent ablation identifies mean matching as the main source of the FID gain and second-moment matching as an additional source of recall. Mean-only matching achieves FID 6.653 and full Moment 6.568; the difference is −0.085, with conditional 95% CI [−0.255, 0.121]. The additional FID improvement is therefore unresolved.

Adding second moments increases recall by 0.0131 ([0.0048, 0.0213]). Precision point estimates are 0.84560 for mean-only matching and 0.84610 for full Moment, so the recall gain occurs without a lower precision point estimate. Tables 30–31 provide the complete comparison.

We now state this division of contributions in the abstract and Section 7.2. Mean-only matching is a strong simpler alternative. Both variants are quadratic in the selection variables and use the same fixed-cardinality solver.

W4 / Q4: Generalization beyond ImageNet and Inception features

The SDXL experiment in Section 6.2 extends the retention comparison to text-to-image generation using DINOv2 retention features, CLIP text–image scores and GenEval task evaluation. It produces eight 1024×1024 images per prompt. The 32 development prompts are separate from the 128 held-out prompts spanning six GenEval task types. Appendix K.3 specifies the search schedule, preprocessing, selector weights and evaluation protocol.

The all-output LPIPS gain over dynamic Top-K is 0.2088 (conditional 95% CI [0.1840, 0.2330]). Among the 83 prompts with at least two passing images under both policies, LPIPS among passing images increases by 0.1901 ([0.1588, 0.2211]).

Against dynamic DPP, Moment increases all-output ImageReward by 0.0553 ([0.0152, 0.0992]), while the LPIPS difference is unresolved. An exploratory valid-pair analysis also favors Moment over dynamic Top-K, but ordinary and independent sampling have higher delivery rates. This outcome was defined after inspecting the generated outputs.

In the independent matched-precision SiT comparison, Moment lowers Inception-space FID from 7.516 to 6.429 and raises ADM recall from 0.3922 to 0.4125. CLIP-CMMD favors DPP, at 0.33535 versus 0.34791. The DiT CLIP-CMMD difference interval spans zero. Section 7.1 reports these outcomes together.

These SDXL results extend the retention comparison beyond ImageNet and Inception features. SDXL was evaluated at B=4; the branching-budget trend was tested on ImageNet.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.