acceptodds
Under review as a conference paper at ICLR 2027

Prior Subspace Recomposition: A Unified Prior Initialization Framework for Conditional Generation

Abstract

Initial noise plays a crucial yet underexplored role in continuous conditional generative models. While existing golden-noise methods demonstrate that carefully selected or optimized noises can improve generation quality, they are often tailored to specific generation regimes or predefined quality objectives, limiting their adaptability across tasks and evaluation criteria. In this paper, we propose Prior Subspace Recomposition (PSR), a unified framework that reformulates noise initialization as metric-induced prior discovery followed by condition-aware anchor recomposition. Given a downstream evaluation suite, PSR first empirically identifies a metric-induced superior region in the Gaussian prior space and distills it into a reusable set of semantic-agnostic prior anchors. To handle complex conditions, PSR further estimates concept-conditioned anchor fitness and converts multi-dimensional metric responses into relative anchor preferences, which supervise a lightweight local ranker based on first-step denoising velocity fields. At inference time, local ranking predictions are resolved into a global anchor ordering through graph-based optimization, and the top-ranked anchors define a condition-specific local recomposition region on the Gaussian prior shell. The final initialization is obtained via geometry-preserving spherical averaging, integrating complementary anchor priors while preserving the prior distribution geometry. Extensive experiments across text-to-image, text-to-video, and image-to-video generation demonstrate that PSR consistently improves generation quality without modifying the underlying generative backbones.

Primary area:
generative models
Keywords:
Conditional Generation, Prior Initialization, HodgeRank

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_Ly5EOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Prior Subspace Recomposition (PSR), a framework for improving conditional generation through noise initialization. The method first discovers a set of high-quality noise anchors using downstream evaluation metrics. It then trains a lightweight ranker to predict the compatibility between these anchors and semantic conditions. During inference, the predicted preferences are aggregated into a global ranking, and the top anchors are combined to obtain the final initialization.

The method is evaluated on text-to-image, text-to-video, and image-to-video generation using several pretrained backbones. The experiments show consistent improvements over random initialization without modifying the underlying generative models.

Overall, I lean toward acceptance. The paper is clearly presented, the methodology is well explained, and the experiments provide reasonable support for the proposed approach. While I have some concerns about the computational overhead, I do not see any major technical issues.

I should also note that I am not an expert in this particular research area, and my knowledge of the related literature is limited. I am therefore not fully confident in judging the novelty of the contribution relative to prior work. My recommendation is mainly based on the clarity of the presentation and the experimental evidence, and I assign a relatively low confidence score.

Critical strengths
  1. The paper is clearly written and well organized. Although I am not an expert in this specific research direction, I found the motivation and methodology easy to follow. The overall presentation is quite complete, and the technical details are sufficiently explained.

  2. The experimental evaluation is fairly comprehensive, covering different generation tasks, backbones, and benchmarks. The reported improvements are consistent, and the ablation studies help explain the contributions of different components. Overall, I find the experimental results encouraging.

Critical weaknesses

My main concern is the additional computational cost. The method requires several offline stages, including anchor discovery and ranker training, and also introduces extra inference overhead. Although the authors provide an efficiency analysis, the practical trade-off between generation quality and computational cost could be discussed further.

Apart from this, I did not identify any major technical issues. However, my familiarity with the literature on noise initialization is limited, so I am less confident in assessing the contribution relative to existing work.

Questions for authors
  1. How does the additional inference overhead scale with the number of anchors and the model size? Is there a recommended configuration for practical applications?

  2. How sensitive is the method to the evaluation metrics used during anchor discovery? Would the selected anchors remain effective when evaluated using different metrics?

4 comments
AuthorsOpenReview ↗

Response to Weakness 1 and Question 1

For Weakness 1 and Question 1: Regarding Computational Cost, Scaling, and Practical Configuration

We thank the reviewer for emphasizing the practical trade-off between generation quality and computational cost. This trade-off is examined in Section 4.3 and Table 4 (right) of the main manuscript, with additional details in Appendix A.3, Tables 8–9. We clarify below how the additional computation translates into generation-quality improvements and how the anchor budget can be adjusted for practical use.

Offline construction and reuse. On Wan2.1-1.3B using eight NVIDIA H200 GPUs, prior anchor discovery, concept–anchor supervision construction, and ranker training take approximately 3, 46, and 13 hours, respectively. These costs are incurred once for a given backbone and evaluation configuration. The resulting anchor pool and ranker are then reused across prompts, so repeated generation does not require repeating the offline pipeline. The construction cost can therefore be amortized over subsequent use within the same configuration.

Practical benefit of the additional computation. The additional inference cost allows the smaller backbone to approach the average generation quality of a substantially larger model. As reported in Appendix A.3, Wan2.1-1.3B with PSR achieves a VMBench average of 62.55, comparable to 63.35 for Wan2.2-5B with random initialization, while requiring approximately 115 seconds rather than 150 seconds per video. Under the reported model-specific generation configurations, this corresponds to a 0.80-point score difference and approximately 23.3% lower inference latency. Thus, the practical benefit is not merely an improvement over the same model’s random initialization: PSR also brings the smaller backbone closer to the larger model’s aggregate benchmark performance.

The existing results further demonstrate that the initialization budget is adjustable:

ConfigurationAnchors KKVMBench Avg.Additional latencyTotal inference latency
Wan2.1-1.3B—60.820 s91 s
Wan2.1-1.3B+PSR2562.5113 s104 s
Wan2.1-1.3B+PSR5062.5524 s115 s
Wan2.2-5B—63.350 s150 s

Reducing the anchor pool from 50 to 25 decreases the additional initialization latency by 45.8%, while the average score decreases by only 0.04 points. The 25-anchor configuration retains approximately 97.7% of the average-score improvement achieved with 50 anchors, providing a lower-latency operating point with nearly unchanged aggregate quality.

Scaling with anchor count and backbone inference cost. PSR’s main inference overhead comes from extracting prompt-conditioned first-step velocity features. The unconditional features associated with fixed anchors are precomputed and cached, while the prompt-conditioned features are evaluated in batches across anchors. The overhead is therefore more naturally understood relative to the backbone’s own inference time, rather than as a fixed delay or a function of parameter count alone. Both ordinary generation and PSR feature extraction use forward evaluations of the same backbone. Increasing the anchor count increases the feature-extraction workload, whereas reducing it lowers the additional cost. The exact latency ratio also depends on batching efficiency and the available hardware.

Recommended practical configuration. For practical deployment, we recommend setting the anchor budget relative to the full sampling-step count, rather than using a fixed number of anchors for every generator. An initial heuristic is to use approximately half as many anchors as sampling steps, together with Top-2 recomposition under the cached and batched implementation. Table 4 (right) illustrates the practical value of this reduced-budget configuration: with K=25K=25, PSR adds 13 seconds to the 91-second generation baseline, corresponding to 14.3% overhead, while achieving 62.51 compared with 62.55 at K=50K=50. Thus, K=25K=25 is the recommended latency-sensitive configuration for the evaluated setting.

AuthorsOpenReview ↗

Response to Question 2

For Question 2: Regarding Sensitivity to Evaluation Metrics and Cross-Metric Transfer

We thank the reviewer for this question. PSR is metric-adaptive rather than metric-invariant: the evaluation suite defines the objective for anchor discovery and ranking supervision, so changing that objective can affect which anchors are preferred. However, this does not imply that the resulting benefits are restricted to the metrics used during offline construction. The manuscript already investigates this question through two complementary experiments.

Transfer to non-optimized video-quality dimensions. In Appendix A.2, Table 5, we restrict both anchor discovery and semantic-conditioned supervision on Wan2.2-5B to TCS and MSS only, while evaluating generation using all five VMBench metrics. The common prompt is also revised accordingly to emphasize temporal coherence and motion smoothness.

ConfigurationPASOISTCSMSSCASAvg.
Random initialization14.8668.6595.2179.4658.5863.35
PSR, TCS+MSS only21.3069.8795.7283.1163.5566.71
PSR, full metric suite24.8470.4396.3184.0063.3267.78

Although PAS, OIS, and CAS are excluded from anchor scoring and ranking supervision, they improve from 14.86 to 21.30, 68.65 to 69.87, and 58.58 to 63.55, respectively. The overall average increases from 63.35 to 66.71. This provides direct evidence that the benefits of the selected anchors and the resulting PSR pipeline extend to quality dimensions not used as offline supervision signals.

Using the full metric suite further improves the average score to 67.78, although the two-metric configuration retains a slightly higher CAS. Thus, the choice of objective affects the balance of improvements across dimensions, but the benefits are not confined to the explicitly optimized metrics.

Transfer to additional image-quality evaluators. The SDXL experiment in Table 4 (left) and Appendix A.4, Table 10 provides complementary evidence of cross-metric transfer. Both anchor discovery and semantic-conditioned ranking supervision use HPSv2 exclusively, while performance is evaluated using HPSv2, ImageReward, and CLIPScore. The results are reproduced below for convenience.

MethodHPSv2ImageRewardCLIPScore
Random initialization29.7196.630.8734
NPNet (Golden Noise)29.8898.810.8813
PSR33.23108.530.8991

PSR improves not only HPSv2 but also the two additional evaluation metrics: ImageReward increases from 96.63 to 108.53, and CLIPScore increases from 0.8734 to 0.8991, relative to random initialization. PSR also exceeds the reported NPNet scores on both metrics. These results indicate that the benefits of HPSv2-guided construction extend beyond the evaluator used to construct the anchor pool and supervise the ranker.

Together with the TCS+MSS-only video experiment, these results provide evidence of cross-metric transfer in both image and video generation. Our claim is not that one anchor pool is universally optimal for every evaluator. Rather, metric-guided construction can yield benefits beyond its offline objectives, while the same PSR procedure can be re-instantiated when a different evaluation objective is desired.

AuthorsOpenReview ↗

Follow-up Regarding Our Rebuttal for Submission #23595

Dear Reviewer Ly5E,

We hope this message finds you well. Thank you again for the time and thoughtful feedback you have devoted to our work.

We have posted a detailed response addressing each of your comments. Specifically, we clarified offline construction and inference costs, contextualized the quality–latency trade-off relative to a larger backbone, explained how overhead depends on anchor count and backbone inference cost, recommended a latency-sensitive configuration, and highlighted evidence of metric adaptability and cross-metric transfer in both image and video generation.

We would greatly appreciate hearing whether these clarifications and additional results have addressed your concerns. Should any questions remain, we would be pleased to provide further information during the author–reviewer discussion period.

Best regards,

The Authors of Submission #23595

Dear Reviewer,

The discussion period is nearing its end, and we haven’t yet seen a reply from you. We encourage you to respond to the authors’ comments, especially if you have further questions about points that may affect your assessment.

Thank you for taking the time to engage in the discussion.

Best,

Your AC

Reviewer_h8Y3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces Prior Subspace Recomposition (PSR), a unified prior initialization framework for continuous conditional generation. The authors formulates noise initialization as a metric-induced prior discovery and condition-aware recomposition problem. Given an evaluation suite, PSR finds a set of metric-induces anchors in the Gaussian prior. Then, it learns subject conditioned anchor preferences using first-step denoising velocities. During inference, PSR local anchor preferences are aggregated into global graph that sits on the dimensional Gaussian shell. PSR shows improvements on Text2Image and Text2Video on multiple base models and benchmarks.

Critical strengths
  1. Framework & Applicability. The paper presents an interesting approach to improving generation quality through the initial noise. The same framework is applied across different generation tasks, models, and evaluation metrics, which makes the contribution practically relevant.
Critical weaknesses
  1. Limited analysis of the contribution of prior subspace discovery. While the paper provides ablations for different anchor selection and recomposition strategies, it is still unclear how much of the improvement comes from the initial metric-induced anchor discovery itself. For example, Appendix Table 11 shows that using the highest-scoring anchor under the common prompt performs slightly worse than random initialization. I believe an important missing baseline is to apply the same learned ranking and recomposition pipeline to an anchor pool consisting of randomly selected noise vectors.

  2. Limited comparisons with existing noise-initialization methods. Although the paper compares against NPNet (Zhou et al., ICCV 2025) on SDXL, the main T2V and I2V experiments primarily compare against random initialization. I believe the experimental evaluation would benefit from comparisons with other relevant noise-initialization methods, such as FreeInit (Wu et al., ECCV 2024) and FastInit (Bai et al., 2025) for video generation, and InitNO (Guo et al., CVPR 2024) or FIND (Chen et al., ACM MM 2024) for image generation.

Questions for authors
  1. Can the authors provide the experiments described in W1?
  2. Can the authors provide visual examples of generation using PSR with the same prompt and different initial seeds? This will help determine if PSR hurts the model diversity capabilities.

These experiments can affect my score.

5 comments
AuthorsOpenReview ↗

Response to Weakness 1, and Question 1 and 2

For Weakness 1 and Question 1: Regarding the contribution of prior subspace discovery

We thank the reviewers for highlighting the need to isolate the contribution of prior subspace discovery. To directly address its effectiveness, we conduct a complementary experiments on Wan2.1-1.3B evaluated on all 1,050 VMBench prompts, using random anchor pools of the same size, K=50K=50.

We construct a random anchor pool by sampling 50 noises from the Gaussian prior without initial prior subspace discovery, and then rerun the remaining PSR pipeline. The downstream local ranking, aggregation, and Top-2 recomposition with α=0.3\alpha=0.3 remain unchanged. This comparison changes the pool construction strategy while keeping the subsequent training and inference procedure fixed.

MethodPASOISTCSMSSCASAvg.
Random11.46±0.0667.27±0.1993.77±0.0373.42±0.2358.18±0.0960.82
Random+PSR11.5468.3194.0174.2759.6161.55
Metric-selected+PSR11.6069.8194.0775.0762.1862.55

Effectiveness of initial anchor-discovery stage. Under the same downstream PSR procedure, replacing the random pool with the metric-selected pool increases the average score from 61.55 to 62.55. Improvements are observed on all metrics. Thus, Common-Prompt-based anchor selection provides an additional benefit to the complete pipeline.

For Question 2: Regarding generation diversity under PSR

We thank the reviewer for raising this important question.

We first clarify the source of variation in our generation pipeline. The flow-matching and rectified-flow backbones used in our experiments, including FLUX and Wan, generate samples by integrating a deterministic ODE from an initial noise state. For a fixed conditioning input and sampling configuration, the initial noise therefore determines the generation trajectory; variation arises from changing the initialization. Consequently, fixing these inputs produces a fixed output, whereas varying the initial noise provides different starting points for generation. Thus, deterministic sampling and generation diversity are compatible: variation originates from the initial conditions rather than stochastic perturbations along a fixed sampling trajectory.

In the default PSR configuration, the final initialization is determined by the selected anchor pair and the fixed recomposition coefficient α=0.3\alpha=0.3. Consequently, once these choices are fixed, repeated generation for the same prompt produces the same result. This configuration prioritizes quality-oriented initialization rather than repeated independent sampling from the original Gaussian prior.

Importantly, PSR supports controllable generation diversity beyond its fixed default configuration. For a given prompt and its ranked anchor pool, alternative initializations can be constructed by varying both the number MM of participating anchors and their recomposition weights. Changing MM introduces different anchor directions, while adjusting the weights provides continuous variation within the resulting family of initializations. These alternatives can also retain comparable generation quality in the evaluated settings (e.g., Table 11 reports VMBench averages of 62.55 and 62.48 for Top-2 and Top-3, respectively).

Thus, PSR provides a mechanism for producing diverse outputs under the same prompt: variation is introduced through anchor recomposition rather than independent sampling from the full Gaussian prior. Different anchor counts and weight configurations can be explored at inference time while reusing the same backbone, anchor pool, and trained ranker, without additional anchor discovery or training. The fixed Top-2 configuration with α=0.3\alpha=0.3 used in our main experiments is therefore a particular operating point for quality evaluation, rather than a structural restriction that limits PSR to a single possible output for each prompt. Visual examples illustrating same-prompt variation under different PSR recomposition settings are provided in Fig. 4 of the revised manuscript.

AuthorsOpenReview ↗

Response to Weakness 2

For Weakness 2: Regarding comparisons with existing noise-initialization methods

We thank the reviewers for emphasizing the importance of comparisons with existing noise-initialization methods.

Additional comparison with InitNO [1]. We conduct an additional experiment following the SD1.4/Pick-a-Pic evaluation setting in Appendix C and Table 7 of NPNet [2]. Specifically, we directly reuse both the SDXL-trained ranker and the anchors discovered on SDXL in our Table 4, without training or anchor rediscovery on SD1.4. The recomposition coefficient remains fixed at α=0.3\alpha=0.3. For transparency, the Standard, InitNO and NPNet results are taken from Table 7 of NPNet, while the PSR results are newly evaluated.

MethodPickScoreHPSv2AESImageReward
Standard (Random)19.1719.495.4575−122.73
InitNO [1]16.5014.475.3116−205.66
NPNet [2]19.2519.545.5151−95.89
PSR20.1025.025.5853-8.41

PSR exceeds the reported Standard, InitNO and NPNet baselines on all four metrics.

Additional comparison with NoiseQuery [3]. We additionally evaluate PSR following the SD1.5/DrawBench setting reported in Table 2 of NoiseQuery. We follow the reported evaluation protocol and rerun the PSR pipeline. The recomposition coefficient is fixed at α=0.3\alpha=0.3.

MethodImageRewardPickScoreHPSv2CLIPScore
SD1.5+random†0.0421.1124.5730.90
SD1.5+NoiseQuery†0.0821.1625.0231.41
SD1.5+PSR0.145721.279425.807131.5494

PSR achieves higher scores than the reported random and NoiseQuery baselines on all metrics.

We further clarify three points.

First, the experimental design follows the specific motivation stated in our Introduction that existing initial-noise methods are generally developed for a particular generation regime and a predefined quality objective. PSR addresses this mechanism-level limitation through a unified metric-induced discovery, semantic-conditioned routing, and recomposition framework that can be instantiated across T2I, T2V, and I2V tasks and under different evaluation metrics. Accordingly, our experiments on heterogeneous backbones and benchmarks are designed to test this framework-level generality. Random initialization is the standard inference procedure, and therefore provides the natural reference for measuring the benefit introduced by PSR.

Second, this comparison protocol is also consistent with established practice in the initial-noise literature. Random initialization is also the principal reference used by prior initialization methods such as Golden Noise and FreeInit [4], because the central question is whether the proposed initialization strategy improves the standard sampling pipeline, rather than whether every previously proposed noise optimizer can be reproduced on every newly released backbone. Golden Noise explicitly adopts standard sampling as its main baseline, while FreeInit evaluates each video generator by comparing its original inference procedure against the same backbone augmented with FreeInit. Moreover, generative backbones and quality benchmarks evolve rapidly; restricting the evaluation to older model–benchmark pairs may provide limited evidence regarding whether an initialization framework remains effective on current generation systems.

Last but not least, reproducing multiple task-specific initialization methods across all settings would be computationally disproportionate, particularly because large-scale video generation and evaluation are highly time- and resource-intensive. Moreover, many prior methods rely on model-specific training data and procedures, making direct transfer to new backbones nontrivial. For example, FastInit [5] constructs approximately 150,000 prompt–noise pairs separately for AnimateDiff and ModelScopeT2V and trains its VNPNet on these backbone-specific datasets; extending it to Wan or LTX would therefore require regenerating a comparable paired dataset and retraining the model. Furthermore, the FastInit paper states that its code and dataset "will be released", but no official implementation was publicly available, making faithful reproduction on additional video backbones particularly difficult. Consequently, porting such methods across modern backbones would require substantial data generation, training, and implementation effort, while the resulting adaptations might no longer constitute faithful comparisons with the original methods.

Reference:

[1] InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization.

[2] Golden Noise for Diffusion Models: A Learning Framework.

[3] The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation, ICCV, 2025.

[4] FreeInit: Bridging Initialization Gap in Video Diffusion Models.

[5] FastInit: Fast Noise Initialization for Temporally Consistent Video Generation.

AuthorsOpenReview ↗

Follow-up Regarding Our Rebuttal for Submission #23595

Dear Reviewer h8Y3,

We hope this message finds you well. Thank you again for the time and thoughtful feedback you have devoted to our work.

We have posted a detailed response addressing each of your comments. Specifically, we isolated the contribution of prior subspace discovery through a matched-size random-pool comparison, clarified controllable generation diversity with same-prompt visual examples, expanded comparisons with InitNO, NPNet, and NoiseQuery, and explained the evaluation scope and practical constraints of reproducing task-specific baselines across heterogeneous backbones.

We would greatly appreciate hearing whether these clarifications and additional results have addressed your concerns. Should any questions remain, we would be pleased to provide further information during the author–reviewer discussion period.

Best regards,

The Authors of Submission #23595

Reviewer_7WiPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Prior Subspace Recomposition (PSR) for noise initialization in conditional generation. It selects reusable noise anchors using evaluation metrics, learns prompt-conditioned rankings from first-step velocity features, and combines the highest-ranked anchors. The proposed method is evaluated on text-to-image, text-to-video, and image-to-video generation without modifying the generative backbones.

Critical strengths

The experiments cover multiple tasks and backbones, and the results support the practical value of the approach.

Critical weaknesses
  1. The benefit of the initial anchor-discovery stage is not isolated. Table 11 compares selection strategies within the discovered pool, but does not compare against randomly selected anchors using the same supervision, ranking, and recomposition pipeline. The results therefore leave unclear whether metric-based discovery is necessary. Likewise, the graph aggregation is not compared with simpler aggregation rules using the same local predictions.

  2. The comparison with closely related noise-library methods is incomplete. NoiseQuery [R1] already constructs a reusable noise library and retrieves initializations through feature matching. The paper’s Sim-Top1 baseline instead retrieves a subject and uses that subject’s preferred anchor. A direct comparison would help establish the additional value of PSR’s learned ranking and recomposition.

  3. Equation (8) normalizes the combined vector to norm d\sqrt{d}, whereas the stated implementation uses 1−α2a(1)+αa(2)\sqrt{1-\alpha^2}a_{(1)}+\alpha a_{(2)} without explicit normalization. These expressions are not equivalent. The latter may be a reasonable approximation for nearly orthogonal high-dimensional anchors, but the actual implementation and the effect of this approximation need clarification.

[R1] R. Wang, H. Huang, Y. Zhu, O. Russakovsky and Y. Wu, "The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation," 2025 IEEE/CVF International Conference on Computer Vision (ICCV), Honolulu, HI, USA, 2025, pp. 17618-17628.

Questions for authors
  1. How does PSR perform with 50 randomly selected anchors under otherwise identical training and inference? How much additional benefit comes from graph aggregation compared with a simpler aggregation rule?

  2. How does PSR compare with direct feature-based noise retrieval under a shared image-generation backbone and comparable candidate and computational budgets?

  3. Which recomposition formula generated the reported results? Does explicit normalization materially change either the noise norms or generation quality?

5 comments
AuthorsOpenReview ↗

Response to Weakness 1 and 2, and Question 1 and 2

For Weakness 1 and Question 1: Regarding the effectiveness of initial anchor-discovery stage

We thank the reviewers for highlighting the need to isolate the contribution of offline anchor discovery. To address its effectiveness, we conduct a complementary experiments on Wan2.1-1.3B evaluated on all 1,050 VMBench prompts, using random anchor pools of the same size, K=50K=50.

We construct a random anchor pool by sampling 50 noises from the Gaussian prior without initial anchor-discovery stage, and then rerun the remaining PSR pipeline. The downstream local ranking, aggregation, and Top-2 recomposition with α=0.3\alpha=0.3 remain unchanged. This comparison changes the pool construction strategy while keeping the subsequent training and inference procedure fixed.

MethodPASOISTCSMSSCASAvg.
Random11.46±0.0667.27±0.1993.77±0.0373.42±0.2358.18±0.0960.82
Random+PSR11.5468.3194.0174.2759.6161.55
Metric-selected+PSR11.6069.8194.0775.0762.1862.55

Effectiveness of initial anchor-discovery stage. Under the same downstream PSR procedure, replacing the random pool with the metric-selected pool increases the average score from 61.55 to 62.55. Improvements are observed on all metrics. Thus, Common-Prompt-based anchor selection provides an additional benefit to the complete pipeline.


For Weakness 1 and Question 1: Regarding comparison with simpler aggregation

We thank the reviewer for suggesting a comparison with simpler aggregation rules. To evaluate a simpler alternative to graph-based aggregation, we introduce a hierarchical naive aggregation, which uses the proposed ranker but replaces graph-based global ranking with iterative hierarchical selection.

Specifically, for each test prompt, we partition the 50 candidate anchors into 10 non-overlapping subsets of five anchors, matching the local ranking length τ=5\tau=5. Each subset is ranked independently using the proposed ranker, and only its highest-scoring anchor is retained, reducing the candidate pool from 50 to 10. The remaining 10 anchors are then partitioned into two non-overlapping subsets of five and ranked again. Retaining the highest-scoring anchor from each subset yields the final two anchors, resulting in a 50 ⁣→ ⁣10 ⁣→ ⁣250\!\rightarrow\!10\!\rightarrow\!2 selection procedure. These two surviving anchors are recomposed with α=0.3\alpha=0.3, as in the main experiments. We evaluate this baseline on Wan2.1-1.3B using all 1050 VMBench prompts.

MethodPASOISTCSMSSCASAvg.
Random11.4667.2793.7773.4258.1860.82
PSR (Hierarchical aggregation, 50→10→250 \rightarrow 10 \rightarrow 2)10.4269.4694.1774.6461.7062.08
PSR (Graph aggregation)11.6069.8194.0775.0762.1862.55

The hierarchical baseline improves the average score from 60.82 to 62.08, demonstrating that a simpler selection procedure already benefits from the learned anchor preferences. Nevertheless, graph aggregation further increases the average score to 62.55, providing an additional gain of approximately 0.47 points.

This improvement is consistent with the motivation for global preference aggregation. Hierarchical aggregation makes intermediate elimination decisions, potentially discarding anchors before their preferences can be considered more broadly. In contrast, graph-based aggregation integrates the available local comparisons into a global score vector before selecting the final anchors. The results suggest that this global treatment helps retain useful candidates and improves the final recomposed initialization, supporting the benefit of graph aggregation over the simpler hierarchical alternative.


For Weakness 2 and Question 2: Regarding comparison with NoiseQuery

We thank the reviewer for highlighting this closely related work. We additionally evaluate PSR following the SD1.5/DrawBench setting reported in Table 2 of NoiseQuery.

We follow the reported DrawBench evaluation protocol and rerun the complete PSR pipeline. The generative backbone remains frozen, and the recomposition coefficient is fixed at α=0.3\alpha=0.3. Performance is evaluated using ImageReward, PickScore, HPSv2, and CLIPScore.

MethodImageRewardPickScoreHPSv2CLIPScore
SD1.5+random†0.0421.1124.5730.90
SD1.5+NoiseQuery†0.0821.1625.0231.41
SD1.5+PSR0.145721.279425.807131.5494

PSR achieves higher scores than the reported random-initialization and NoiseQuery baselines on all four metrics. These results extend our evaluation beyond the simplified Sim-Top1/Top2 baselines to a closely related noise-library method.

The distinction is that NoiseQuery selects a stored initialization through feature matching, whereas PSR evaluates anchor compatibility using prompt-conditioned first-step backbone responses, learns relative anchor preferences, and recomposes the selected anchors.

AuthorsOpenReview ↗

Response to Weakness 3 and Question 3

For Weakness 3 and Question 3: Regarding the recomposition formula and the normalization approximation.

We thank the reviewer for pointing out the distinction between the normalized formulation and its implementation. The implementation combines the two selected anchors without explicit normalization. For Top-2 recomposition, the two expressions are

zinit⋆=d,β1a(1)+β2a(2)∣β1a(1)+β2a(2)∣2,z_{\mathrm{init}}^\star = \sqrt d\\, \frac{\beta_1 a_{(1)}+\beta_2 a_{(2)}} {\\|\beta_1 a_{(1)}+\beta_2 a_{(2)}\\|_2},

z~init⋆=1−α2,a(1)+αa(2),α∈[0,1].\widetilde z_{\mathrm{init}}^\star = \sqrt{1-\alpha^2}\\,a_{(1)} + \alpha a_{(2)}, \qquad \alpha\in[0,1].

The first expression fixes the norm to the square root of the latent dimension, rather than the latent dimension itself.


Equivalence conditions.

Setting

β1=1−α21−α2+α,β2=α1−α2+α\beta_1 = \frac{\sqrt{1-\alpha^2}}{\sqrt{1-\alpha^2}+\alpha}, \qquad \beta_2 = \frac{\alpha}{\sqrt{1-\alpha^2}+\alpha}

makes the two expressions identical in direction, with any difference arising solely from radial normalization. A sufficient condition for exact equivalence is

∣a(1)∣22=∣a(2)∣22=d,⟨a(1),a(2)⟩=0.\\|a_{(1)}\\|_2^2=\\|a_{(2)}\\|_2^2=d, \qquad \langle a_{(1)},a_{(2)}\rangle=0.

Under these conditions,

∣z~init⋆∣22=(1−α2)d+α2d=d,\\|\widetilde z_{\mathrm{init}}^\star\\|_2^2 = (1-\alpha^2)d+\alpha^2d=d,

so explicit normalization leaves the combination unchanged. We do not assume that the selected anchors satisfy these conditions exactly; instead, we examine their geometry and directly measure the resulting approximation error.


Empirical verification of anchor norms and near-orthogonality.

Since metric-based selection may alter the distribution of the retained anchors, we measure their geometry directly rather than relying solely on properties of independent Gaussian samples.

For the Wan2.1-1.3B Top-50 pool, each anchor has dimensionality

d=16×21×104×60=2,096,640,d≈1447.978.d=16\times21\times104\times60=2{,}096{,}640, \qquad \sqrt d\approx1447.978.

Across all 50 anchors, the measured normalized norms satisfy

0.998851≤∣a(m)∣2d≤1.001084.0.998851 \leq \frac{\\|a_{(m)}\\|_2}{\sqrt d} \leq 1.001084.

The maximum relative deviation from the target radius is therefore approximately 0.115%, confirming that the selected anchors remain close to the target norm.

We additionally compute the cosine similarity between every pair of distinct anchors:

ρu,w=⟨au,aw⟩∣au∣2∣aw∣2,u≠w.\rho_{u,w} = \frac{\langle a_u,a_w\rangle} {\\|a_u\\|_2\\|a_w\\|_2}, \qquad u\neq w.

The maximum and mean absolute cosine similarities are

max⁡u≠w∣ρu,w∣=0.002671,Mean⁡u≠w∣ρu,w∣=0.000536.\max_{u\neq w}|\rho_{u,w}|=0.002671, \qquad \operatorname{Mean}_{u\neq w}|\rho_{u,w}|=0.000536.

Together, these measurements support both approximate equal-norm concentration around the target radius and approximate pairwise orthogonality.


Direct measurement of the normalization error.

To account for both imperfect orthogonality and deviations in anchor norms, we directly measure the relative Euclidean discrepancy between the normalized and unnormalized initializations:

Erel=∣z~init⋆−zinit⋆∣2∣zinit⋆∣2=∣∣z~init⋆∣2d−1∣.E_{\mathrm{rel}} = \frac{ \\|\widetilde z_{\mathrm{init}}^\star-z_{\mathrm{init}}^\star\\|_2 }{ \\|z_{\mathrm{init}}^\star\\|_2 } = \left| \frac{\\|\widetilde z_{\mathrm{init}}^\star\\|_2}{\sqrt d}-1 \right|.

The equality follows from their identical directions and holds for any nonzero combination, without assuming independent Gaussian anchors.

For the Wan2.1 Top-50 pool, we evaluate this error over all 2,450 ordered pairs of distinct anchors. The numerical evaluation uses

z~init⋆=0.91,a(1)+0.3,a(2),\widetilde z_{\mathrm{init}}^\star = \sqrt{0.91}\\,a_{(1)} + 0.3\\,a_{(2)},

corresponding to an implementation argument α=0.3\alpha=0.3, which is the coefficient of the second anchor; its square is 0.09.

StatisticRelative error
Mean0.03896%
95th percentile0.09647%
Maximum0.18413%

Thus, explicit normalization changes the radial scale by less than 0.2% for every evaluated anchor pair. These results support a close geometric approximation for the measured anchor pool and mixing coefficient, rather than unconditional algebraic equivalence.


End-to-end comparison with explicit normalization. We therefore additionally evaluate the variant that explicitly applies Eq. (8), using the original Top-50 anchor pool and the velocity-based ranker.

ImplementationPASOISTCSMSSCASAvg.
Default11.6069.8194.0775.0762.1862.55
Eq. (8)11.2769.5594.1774.5062.2062.34

The explicitly normalized variant achieves an average VMBench score of 62.34, compared with 62.55 for the default implementation, a difference of only 0.21 points. Together, the geometric measurements and end-to-end comparison show that the default implementation closely approximates the radial scale of Eq. (8).

AuthorsOpenReview ↗

Follow-up Regarding Our Rebuttal for Submission #23595

Dear Reviewer 7WiP,

We hope this message finds you well. Thank you again for the time and thoughtful feedback you have devoted to our work.

We have posted a detailed response addressing each of your comments. Specifically, we isolated the contribution of offline anchor discovery, assessed graph aggregation against a simpler hierarchical alternative, extended baseline comparisons to NoiseQuery, clarified the equivalence conditions between the recomposition formula and its implementation, and quantified the geometric discrepancy and downstream performance differences associated with explicit normalization.

We would greatly appreciate hearing whether these clarifications and additional results have addressed your concerns. Should any questions remain, we would be pleased to provide further information during the author–reviewer discussion period.

Best regards,

The Authors of Submission #23595

Reviewer_9ciROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the optimization of initial noise in conditional generative models and proposes a unified framework, Prior Subspace Recomposition, to improve generation quality without modifying the parameters of pretrained models. The method first identifies high-quality noise samples based on evaluation metrics to construct a reusable anchor pool. It then uses generation quality under different semantic conditions as supervision to train a local ranker based on first-step denoising velocity fields. Finally, graph-based optimization is employed to obtain a global ranking of anchors, and the top-ranked anchors are recomposed through spherical blending to produce condition-aware initial noise. The authors evaluate the effectiveness of PSR across multiple generative models on T2I, T2V, and I2V tasks, and further analyze its performance improvements, generalization ability, and computational overhead through comparisons with existing methods and ablation studies.

Critical strengths
  1. Unified and Flexible Framework. The paper presents a unified initial noise optimization framework applicable to T2I, T2V, and I2V tasks without modifying pretrained generative models. Its metric-driven anchor discovery and condition-aware ranking design offer flexibility across different generation tasks and evaluation objectives.
  2. Consistent Empirical Improvements. The proposed method demonstrates improvements over random initialization across multiple generative backbones and benchmarks, with particularly notable gains on Wan2.2-5B. The comparison with NPNet and additional ablation studies further support the effectiveness of the proposed framework.
Critical weaknesses
  1. Lack of Fair Comparisons with Strong Baselines. The paper primarily compares the proposed method against random initialization, without direct comparisons with representative methods such as InitNO (CVPR 2024) and FreeInit (ECCV 2024) on applicable tasks.
  2. Insufficient Evidence for the Effectiveness of Offline Anchor Discovery. The authors select the Top-K noise samples using a single task-specific Common Prompt but do not demonstrate that these anchors offer better generation potential for unseen prompts. The paper lacks an analysis of noise-quality correlations across prompts and a controlled comparison with randomly selected anchor pools, making it difficult to establish the effectiveness of this core offline selection strategy. Lack of Direct Validation of the Semantic-Conditioned Anchor Ranking Mechanism. While the paper demonstrates improvements from condition-aware ranking, it does not directly analyze how ground-truth anchor preferences vary across subjects or quantify the agreement between predicted rankings and actual generation-quality rankings. Furthermore, a controlled comparison with a ranker using only anchor and text features is missing, leaving the necessity of first-step velocity features insufficiently justified.
Questions for authors
  1. Baseline Comparisons: Please include comparisons with more representative baseline methods.
  2. Offline Anchor Discovery: Does the metric-selected Anchor Pool outperform a randomly selected pool of the same size when using the identical ranking and recomposition pipeline? What evidence supports the transferability of Common Prompt-based anchor selection to unseen prompts?
  3. Semantic-Conditioned Anchor Ranking: Can the authors provide direct quantitative evidence that anchor preferences are subject-dependent and that the learned ranker accurately predicts generation-quality rankings on unseen prompts (e.g., using Spearman or Kendall correlations)? Additionally, how does the velocity-based ranker compare with an otherwise identical ranker using only anchor and text features?
7 comments
AuthorsOpenReview ↗

Response to Weakness 1 and 2, and Question 1 and 2

For Weakness 1 and Question 1: Regarding baseline comparisons

We thank the reviewers for emphasizing the importance of comparisons with representative baselines. We would first like to clarify that Table 4 of the main manuscript already includes a comparison with a representative method, NPNet (Golden Noise, ICCV 2025) [1] on SDXL. In that experiment, PSR achieves higher HPSv2, ImageReward, and CLIPScore scores than NPNet. Thus, our evaluation is not restricted to random-initialization baselines.

Additional comparison with InitNO [2]. To address the request for more representative baselines, we conduct an additional experiment following the SD1.4/Pick-a-Pic evaluation setting in Appendix C and Table 7 of NPNet. We report the same four metrics: PickScore, HPSv2, Aesthetic Score (AES), and ImageReward. Specifically, we directly reuse both the SDXL-trained ranker and the anchor pool discovered on SDXL in Table 4, without training or anchor rediscovery on SD1.4. The recomposition coefficient remains fixed at α=0.3\alpha=0.3. For transparency, the Standard, InitNO, and NPNet results are taken from Table 7 of NPNet, while the PSR results are newly evaluated.

MethodPickScoreHPSv2AESImageReward
Standard (Random)19.1719.495.4575−122.73
InitNO16.5014.475.3116−205.66
NPNet19.2519.545.5151−95.89
PSR (SDXL → SD1.4)20.1025.025.5853-8.41

Under this evaluation setting, PSR exceeds the reported Standard, InitNO and NPNet on all four metrics.

Additional comparison with NoiseQuery [3]. We additionally evaluate PSR following the SD1.5/DrawBench setting reported in Table 2 of NoiseQuery. We follow the reported evaluation protocol and rerun the PSR pipeline. The backbone remains frozen, and the recomposition coefficient is fixed at α=0.3\alpha=0.3.

MethodImageRewardPickScoreHPSv2CLIPScore
SD1.5+random†0.0421.1124.5730.90
SD1.5+NoiseQuery†0.0821.1625.0231.41
SD1.5+PSR0.145721.279425.807131.5494

PSR also achieves higher scores than the reported random-initialization and NoiseQuery on all metrics.

Reference:

[1] Golden Noise for Diffusion Models: A Learning Framework, ICCV, 2025.

[2] InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization, CVPR, 2024.

[3] The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation, ICCV, 2025.


For Weakness 2 and Question 2: Regarding the effectiveness and transferablity of offline anchor discovery

We thank the reviewers for highlighting the need to isolate the contribution of offline anchor discovery. To address both its effectiveness and transferability, we conduct two complementary experiments on Wan2.1-1.3B evaluated on all 1050 VMBench prompts, using anchor pools of the same size, K=50K=50.

  1. Does the metric-selected pool outperform a random pool under the same PSR pipeline?

    We construct a random anchor pool by sampling 50 noises from the Gaussian prior without Offline Anchor Discovery, and then rerun the remaining PSR pipeline. The downstream local ranking, aggregation, and Top-2 recomposition with α=0.3\alpha=0.3 remain unchanged. This comparison changes the pool construction strategy while keeping the subsequent training and inference procedure fixed.

  2. Do the selected anchors remain useful without learned selection?

    To assess the selected pool independently of the ranker, we additionally sample one anchor uniformly from the metric-selected pool for each evaluation prompt and directly use it for generation. We repeat the full benchmark four times. We report the mean and standard deviation of benchmark-level scores across these runs.

MethodPASOISTCSMSSCASAvg.
Random11.46±0.0667.27±0.1993.77±0.0373.42±0.2358.18±0.0960.82
Metric-selected11.10±0.2267.70±0.0993.86±0.2674.10±0.6059.43±0.2861.24
Random pool+PSR11.5468.3194.0174.2759.6161.55
Metric-selected+PSR11.6069.8194.0775.0762.1862.55

Effectiveness of Offline Anchor Discovery. Under the same downstream PSR procedure, replacing the random pool with the metric-selected pool increases the average score from 61.55 to 62.55. Improvements are observed on all metrics. Thus, Common-Prompt-based anchor selection provides an additional benefit to the complete pipeline.

Transferablity of Offline Anchor Discovery. Uniform sampling from the metric-selected pool achieves an average score of 61.24, compared with 60.82 for random initialization. Since this experiment uses neither learned ranking nor recomposition, the improvement provides evidence that the selected pool remains useful on unseen prompts.

Together, offline selection produces a candidate pool that improves average performance beyond the calibration condition and benefits the otherwise unchanged PSR pipeline.

AuthorsOpenReview ↗

Response to Weakness 2 and Question 3

For Weakness 2 and Question 3: Regarding ranker using only anchor and text features

We thank the reviewer for raising this question. Our use of first-step velocity features is motivated by the distinction between providing the conditioning text and exposing the target generator's response to that condition. Both the velocity-based ranker and an anchor-and-text variant receive textual information; the difference is whether the ranking representation explicitly incorporates the conditional dynamics of the generative backbone.

For a fixed anchor, its raw noise coordinates remain unchanged across prompts and do not explicitly describe how the target generator will process that anchor under a particular condition. In contrast, the first-step velocity is obtained by evaluating the target backbone at the initial sampling state. It therefore reflects the generator's initial response to the specific anchor–condition pair. This provides a backbone-informed, condition-dependent representation, rather than requiring anchor suitability to be inferred directly from raw noise coordinates and text embeddings. Consequently, this representation is intended to reduce the burden on the lightweight ranker by making the interaction between the anchor, the condition, and the target backbone explicit.

Thus, first-step velocity features serve as a generator-informed inductive bias for anchor preference learning and explicitly incorporating the backbone's response provides a more direct representation of the compatibility that PSR aims to predict.

Direct comparison with anchor-and-text features. To examine this design choice empirically, we additionally evaluate an Anchor+Text variant, which supplies the raw anchor noise and textual features to the ranker instead of first-step velocity features and textual features. The results on Wan2.1-1.3B evaluated by VMBench are shown below.

Ranker inputsPASOISTCSMSSCASAvg.
Anchor+Text12.5367.2393.3272.2860.1161.09
Velocity+Text11.6069.8194.0775.0762.1862.55

Using first-step velocity features increases the average score from 61.09 to 62.55. Notably, Anchor+Text performs comparably to uniformly sampling a single anchor from the metric-selected pool without ranking or recomposition (61.09 vs. 61.24). Thus, in this setting, the anchor-and-text variant provides little observed additional benefit over simply using the offline-selected pool. In contrast, the velocity-based ranker yields a higher average score, supporting the value of explicitly incorporating the target generator’s conditional response into anchor preference prediction.

For Weakness 2 and Question 3: Regarding the validation of the semantic-conditioned anchor ranking mechanism

We thank the reviewer for requesting direct validation of the learned anchor rankings. To complement the downstream generation results, we add a ranking-agreement evaluation over 100 complete ranking instances, each covering all 50 anchors.

We compare the proposed velocity-based ranker with both an Anchor+Text variant and a random-ranking baseline. Spearman and Kendall correlations measure agreement over the complete 50-anchor ordering, while IoU@10 and Top-10 overlap quantify the recovery of high-quality candidates.

Ranking-agreement metricRandom rankingAnchor+TextVelocity+Text
Spearman correlation0.00220.05600.4209
Kendall correlation0.00140.03900.2526
IoU@100.11550.12970.3736
Top-10 overlap count1.99672.235.36

Agreement with actual generation-quality rankings. The random-ranking baseline exhibits near-zero rank correlations and a Top-10 overlap close to the chance expectation of two anchors. Anchor+Text provides only weak agreement with the reference rankings, whereas Velocity+Text achieves substantially stronger alignment across all four measures, both over the complete anchor pool and among the highest-quality candidates. Together with the downstream generation results, these comparisons directly support the benefit of incorporating the target generator’s conditional response into anchor preference learning.

Identification of high-quality anchors. The predicted Top-10 by Velocity contains an average of 5.36 anchors from the reference Top-10, corresponding to 53.6% recovery, with a mean IoU@10 of 0.3736. For comparison, randomly selecting 10 anchors from a pool of 50 has an expected overlap of only 2 anchors with a fixed reference Top-10 set. These results indicate that the ranker identifies a useful subset of high-quality candidates, even though the complete predicted ordering is not perfectly aligned with the empirical quality ranking.

AuthorsOpenReview ↗

Follow-up Regarding Our Rebuttal for Submission #23595

Dear Reviewer 9ciR,

We hope this message finds you well. Thank you again for the time and thoughtful feedback you have devoted to our work.

We have posted a detailed response addressing each of your comments. Specifically, we expanded comparisons with representative noise-initialization methods, isolated the contribution of offline anchor discovery, examined the transferability of metric-selected anchors, assessed the benefit of first-step velocity features against an anchor-and-text baseline, and quantified the agreement between predicted anchor rankings and empirical generation-quality rankings.

We would greatly appreciate hearing whether these clarifications and additional results have addressed your concerns. Should any questions remain, we would be pleased to provide further information during the author–reviewer discussion period.

Best regards,

The Authors of Submission #23595

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Manuscript Revisions

We thank the Area Chair and reviewers for their constructive feedback.

The revised manuscript adds a discussion of generation diversity with same-prompt visual examples (Fig. 4) and controlled experiments on the effectiveness and transferability of offline anchor discovery, including random-pool comparisons and uniform sampling from the metric-selected pool. Newly added text is highlighted in blue.

Additional baseline comparisons, ranking and recomposition ablations, and clarifications regarding computational cost and cross-metric transfer are provided in our reviewer-specific responses.

AuthorsOpenReview ↗

Summary of Revisions and Additional Evidence

We thank the Area Chair and reviewers for their thoughtful feedback. We have conducted additional experiments and analyses to address the concerns regarding baseline comparisons, offline anchor discovery, semantic-conditioned ranking, recomposition, and practical efficiency. We summarize our responses below.

Baseline comparisons and generality (Reviewers 9ciR, h8Y3). We expanded comparisons with representative noise-initialization methods, including InitNO, NPNet (Golden Noise), and NoiseQuery, across SD1.4, SD1.5, and SDXL.

Effectiveness of prior discovery and ranking (Reviewers 9ciR, h8Y3). We isolated the contribution of metric-induced anchor discovery through matched-size random-pool comparisons and uniform anchor sampling. We further compared velocity-based ranking against Anchor+Text features, quantified ranking agreement using Spearman, Kendall, and Top-10 overlap, and evaluated graph aggregation against a simpler hierarchical alternative.

Recomposition and generation diversity (Reviewer h8Y3). We clarified the relationship between the normalized formulation and its implementation through geometric analysis and end-to-end comparisons. We also added a discussion of controllable generation diversity, accompanied by same-prompt visual examples in Fig. 4 of the revised manuscript.

Efficiency and metric adaptability (Reviewer 7WiP). We clarified the offline and inference costs, quantified the quality–latency trade-off under different anchor budgets, and provided practical configuration recommendations. We also consolidated cross-metric transfer evidence from both image and video generation.

We have provided detailed point-by-point responses and additional experimental evidence addressing the concerns raised. We also followed up to invite further discussion; however, we have not yet received further feedback or clarification requests from the reviewers. We would welcome any additional questions or suggestions during the remaining discussion period and hope the revised manuscript and supplementary evidence adequately address the concerns raised.

AuthorsOpenReview ↗

Thank You for the Updated Assessment

Dear Reviewer 7WiP,

Thank you very much for re-evaluating our work and for your encouraging feedback. We are pleased that our clarifications and additional results have addressed your concerns.

We sincerely appreciate the time and expertise you have devoted to reviewing our paper. Your constructive comments have helped us clarify the scope of our claims, strengthen the empirical evaluation, and improve the overall presentation. We will incorporate these clarifications and additional results into the revised manuscript.

Best regards,

The Authors of Submission #23595

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.