Pith. sign in

REVIEW 3 major objections 5 minor 28 references

By restricting initial-noise updates to a fixed-radius, low-frequency sphere, MoNO restores per-prompt diversity in few-step distilled diffusion models without auxiliary quality-control losses.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-07-31 23:27 UTC pith:PYRA37VB

load-bearing objection MoNO is a genuine advance for inference-time noise optimization: a hard low-frequency affine sphere with correct Riemannian updates, though the 'quality by construction' claim is softer than advertised and the diversity metrics are partly circular. the 3 major comments →

arxiv 2607.23937 v1 pith:PYRA37VB submitted 2026-07-27 cs.CV

Manifold-Constrained Noise Optimization for Diverse Diffusion Sampling

Qitan Shi , Cheng Jin , Ziyuan Liu , Yuantao Gu This is my paper
classification cs.CV
keywords diffusion modelsdiversity collapsenoise optimizationRiemannian manifoldfew-step distillationtext-to-image generationinference-time samplinglow-frequency constraint
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that the per-prompt diversity collapse of few-step distilled diffusion models can be reversed at inference time by optimizing the initial noise, with no retraining and no auxiliary quality-control objectives. Its central move is to treat the admissible noise set as a geometric object: the Gaussian prior fixes the noise norm, and empirical evidence indicates that low-frequency components are the quality-stabilizing handle, so updates are confined to an affine low-frequency sphere. On that manifold, Riemannian gradient descent minimizes the projection of the new image's visual feature onto the subspace spanned by previous samples, forcing each new sample into an uncovered direction. The method is training-free, converges in one to four iterations, and keeps image quality comparable to i.i.d. sampling while substantially increasing measured diversity.

Core claim

Constraining noise optimization to a manifold that enforces prior likelihood and frequency stability by construction turns diversity recovery into a few-step, regularization-free procedure. The feasible set is the intersection of the fixed-norm sphere of the initial Gaussian draw and the affine subspace that leaves high-frequency Fourier components untouched, so only the low-frequency part moves on a radius-fixed sphere. The diversity objective is sequential: for each new noise, the feature of its predicted clean image (from a single-step proxy predictor) is projected onto the orthonormal basis of previous generations' features, and the squared projection length is minimized. Riemannian grad

What carries the argument

The central object is the affine low-frequency sphere: the set of noise vectors that share the initial noise's Euclidean norm and its high-frequency Fourier components, so only the low-frequency component can move, on a fixed-radius sphere inside that subspace. The method equips this set with the inherited Euclidean metric and uses Riemannian gradient descent whose steps are closed-form geodesic rotations, so feasibility is preserved exactly without any regularization term. The driving objective is the sequential feature-subspace suppression loss — the squared norm of the projection of the new image's frozen-extractor feature onto the orthonormal basis of previous images' features. Gradients

Load-bearing premise

Restricting all noise movement to low-frequency components preserves generation quality and still leaves enough freedom to reach visually distinct modes; if a model's collapse is driven by high-frequency variation, or if freezing high frequencies blocks access to legitimate outputs, the method's quality guarantee and diversity gains both fail.

What would settle it

On a distilled generator whose sampler injects stochasticity mainly into high-frequency latent channels, MoNO should show little or no diversity gain because those channels are frozen; observing a large gain on such a model would contradict the claim that the low-frequency restriction is the operative quality-stabilizing mechanism. A second check: after removing only the frequency constraint while keeping the norm constraint, quality should drop or diversity gains should vanish, isolating the frequency constraint as the cause.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • MoNO eliminates auxiliary reward models, soft regularization, and anchoring losses from noise-optimization-based diversity methods; all fidelity constraints are hard geometric ones.
  • Closed-form geodesic updates allow large steps while staying feasible, producing measurable diversity gains in a single iteration and near-saturated gains in four, in contrast to prior methods whose conservative updates barely move the samples.
  • The method is training-free and plug-and-play: it works on 4-step FLUX.2 [klein] 4B, single-step SDXL-Turbo, and single-step DMD2 without changing model weights or the sampler, and its memory cost is independent of the number of generated samples.
  • The low-frequency cutoff and the rotation angle provide a simple quality–diversity dial, replacing the fragile balancing of loss weights with two interpretable hyperparameters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the low-frequency-stability premise holds across architectures, the same affine-sphere constraint set could serve as a universal feasible region for other noise-space objectives in diffusion models, such as prompt adaptation or style control, since it hard-codes prior likelihood and high-frequency robustness.
  • A natural testable extension is an adaptive cutoff that grows with the number of generated samples; the fixed ρ=0.15 may cap the reachable diversity once the low-frequency modes within a prompt are exhausted.
  • Because the diversity objective is purely visual-feature-based and ignores the text condition, a variant that combines the projection loss with a prompt-alignment term could steer the added diversity toward semantically distinct but on-prompt variations.
  • The paper's evidence that low-frequency updates are less disruptive than full-frequency ones is demonstrated on one model with one prompt; a systematic study across resolutions, model families, and sampler schedules would either establish the manifold as a universal handle or reveal regimes where frozen high-frequency components block legitimate modes.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This paper introduces MoNO, a training-free inference-time method for improving per-prompt diversity in few-step distilled text-to-image diffusion models. The method optimizes each newly sampled initial noise so that the DINOv2 patch features of the generated image have minimal projection onto the subspace spanned by previously generated images. The optimization is constrained to an 'affine low-frequency sphere': the norm of the low-frequency component is preserved, and the high-frequency component is frozen. Riemannian gradient descent with a closed-form exponential map keeps every iterate on this manifold, eliminating the need for auxiliary quality regularizers. Experiments on GenEval and DPG-Bench with FLUX.2, SDXL-Turbo, and DMD2 report consistent diversity gains over i.i.d. sampling and several baselines, with small quality drops.

Significance. If the empirical premise holds across models and resolutions, MoNO is a useful, simple, and memory-efficient plug-in for recovering per-prompt diversity in few-step distilled generators. The Riemannian derivation (Appendix A) is correct and clearly presented; the method is genuinely training-free, converges in very few iterations, and the proxy-predictor ablation quantifies a 2.37x speedup. The main strengths are the explicit geometric constraints and the sequential feature-subspace formulation. The two load-bearing concerns are the partial circularity between the optimization objective and some evaluation metrics, and the incomplete empirical support for the claim that the low-frequency sphere is universally quality-stabilizing. These are addressable with additional experiments and rephrasing, so the contribution is potentially publishable, but not in its current form.

major comments (3)
  1. [Appendix B.1/B.2, Table 1] The diversity objective is computed on DINOv2 patch features (Appendix B.2), while the Vendi Score and DreamSim in Table 1 are computed with facebook/dinov2-base and dino_vitb16 respectively (Appendix B.1). Thus the reported increases in DreamSim and Vendi Score are partially by construction: the optimizer is explicitly maximizing alignment with the same representation used for evaluation. LPIPS (VGG) also improves, so the effect is not wholly an artifact, but the main quantitative diversity claim needs to be re-measured with a non-DINO evaluation feature extractor (e.g., CLIP- or VGG-based Vendi, or a held-out DINO checkpoint), and the paper should report the DINO-vs-non-DINO gap. As written, the quantitative evidence for 'consistently improves diversity' is partially circular.
  2. [Frequency constraint, Fig. 3(b), Table 2] The central quality guarantee rests on the empirical premise that low-frequency updates are quality-stabilizing and high-frequency updates are artifact-prone. The evidence is a single illustrative prompt (Fig. 3b) and the Table 2 ablation, which compares Norm-only against Norm+Freq but not a High-freq-only variant (fix low-freq, optimize high-freq). This cannot establish that freezing high frequencies, rather than merely constraining to a low-dimensional subspace, is the stabilizing handle. Please add the missing High-freq-only ablation and characterize the quality behavior across models/resolutions (e.g., quality drop vs. cutoff rho), or soften the 'by construction' claim. Table 1 already shows consistent small quality drops (FLUX HPSv2 0.313->0.294, CLIP-IQA 0.630->0.610; SDXL-Turbo CLIPScore 0.330->0.326; DMD2 CLIPScore 0.333->0.327), so quality preservation is approximate, not exact.
  3. [Abstract/Conclusion and Fig. 8] The abstract and conclusion state 'maintaining image quality' and 'converges in far fewer iterations.' Table 1 shows small but systematic quality decreases under CLIPScore and often HPSv2/CLIP-IQA, and Fig. 8 reveals a clear quality-diversity trade-off. 'Converges' is never formally defined; Fig. 5 is qualitative. Please report confidence intervals or significance tests for the quality differences, state the empirical quality budget explicitly (e.g., 'within a pre-specified tolerance of i.i.d. quality'), and avoid the term 'converges' unless a stopping/convergence criterion is defined.
minor comments (5)
  1. [Appendix B.2, low-pass mask] The mask formula is garbled: it should be an indicator of sqrt(omega_x^2 + omega_y^2)/sqrt(2) <= rho (normalized frequency), not 'sqrt(omega_x^2+omega_y^2) sqrt(1/2)'. Please fix the typesetting.
  2. [Fig. 3(b)] The illustrative example lacks axis labels and does not state the update magnitude, norm, or frequency cutoff. Please specify these details so the reader can gauge whether the comparison is fair.
  3. [Algorithm 1 / B.2] The rotation angle phi is the tuned hyperparameter, but the raw Riemannian step size alpha is not reported. Please report both, or the conversion phi = alpha ||h|| / r, for reproducibility.
  4. [Table 3] The full-sampler variant is evaluated only on FLUX.2. Please either add the same comparison for the single-step models or explicitly acknowledge the scope. Also define what the 2.37x speedup measures (wall-clock time, backward time, or total optimization time).
  5. [General] No code availability or reproducible pipeline statement is included. For a journal version, please add a statement or public code release; the method is simple enough that this would materially strengthen the paper.

Circularity Check

0 steps flagged

No significant circularity: MoNO's derivation is self-contained; its empirical premises and metric overlaps do not reduce to the inputs.

full rationale

I traced the claimed derivation chain: the sequential diversity loss is a projection objective in DINOv2 patch space (Appendix B.2), the admissible set is defined by hard constraints M(z0) = {z : ||z|| = ||z0||, P_H z = P_H z0} (Method, 'Intersection manifold'), and the Riemannian gradient/exponential map are derived from first principles on that affine sphere (Appendix A). No load-bearing step is defined in terms of the quantities it claims to predict, and no fitted parameter is renamed as a prediction. The low-frequency stability premise is empirical, supported by Fig. 3b and the Table 2 ablation, and it is an assumption about the model's noise sensitivity rather than a self-imported theorem; its possible under-support across models is a correctness risk, not a circularity. The paper cites external preprints (Falck et al. 2025; Harrington et al. 2026; Jeon and Lee 2026) for frequency sensitivity, not prior work by the current authors, so there is no self-citation chain. The only potential concern is that the diversity loss uses DINOv2 patch features while Vendi Score uses DINOv2 features and DreamSim uses a DINO backbone. However, the loss is patch-wise and sequential, whereas Vendi is set-level and DreamSim uses DINOv1 ViT-B/16; moreover LPIPS (VGG backbone) is fully independent and also improves, and quality metrics (CLIPScore, CLIP-IQA, HPSv2) provide external grounding. Thus any metric/objective overlap is partial and not a by-construction reduction. The experimental validation against GenEval, DPG-Bench, multiple distilled generators, and prior baselines makes this a self-contained inference-time method with no circular derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The method relies on standard Riemannian geometry (sphere tangent space, geodesic exponential map) and several empirical domain assumptions: low-frequency noise updates are quality-stabilizing, DINOv2 features capture the perceptual diversity one wants to spread out, and the one-step clean prediction A_θ gives a trustworthy gradient signal. Free parameters (rotation angle, cutoff, iteration count, feature extractor) are chosen by hand and swept on GenEval, so they contribute to the diversity-quality trade-off rather than being derived.

free parameters (4)
  • rotation angle φ per iteration = 20° (4-iter) / 30° (1-iter)
    Tuned on GenEval; directly controls the diversity-quality trade-off (Fig. 8).
  • low-frequency cutoff ρ = 0.15
    Binary circular Fourier mask cutoff; chosen by hand and swept on GenEval (Fig. 8).
  • optimization iterations K = 4 (main), 1 (variant)
    Chosen budget; the 'converges in far fewer iterations' claim rests on this choice without a matched-compute comparison.
  • feature extractor and resolution = DINOv2-base at 256×256
    Design choice; the diversity loss and the Vendi evaluation share this extractor, creating partial circularity.
axioms (5)
  • standard math Isotropic Gaussian prior density depends only on ∥z∥2; moving on a fixed-radius sphere preserves prior likelihood exactly.
    Used to justify the norm constraint in the 'Norm constraint' section. Standard concentration-of-measure fact (Vershynin 2020).
  • domain assumption Low-frequency components of the initial noise are a more stable, quality-preserving subspace to optimize than full-frequency updates.
    Load-bearing empirical premise based on Fig. 3(b) and citations (Falck et al. 2025; Harrington et al. 2026; Jeon and Lee 2026). Not a theorem.
  • domain assumption DINOv2 patch-feature projection onto the subspace spanned by previous samples is a faithful proxy for perceptual per-prompt diversity.
    Used in the diversity loss ('Practical Instantiation'). Partially validated by LPIPS/Vendi metrics, but not derived.
  • domain assumption The one-step clean-image prediction A_θ(z,c) provides gradients whose optimization signal matches the full sampler G_θ.
    Ablated in Table 3: proxy is slightly worse on CLIP-IQA/HPSv2 than full trajectory, so the assumption holds only approximately.
  • standard math Projection onto tangent space and exponential map on an affine sphere under the Euclidean metric follow standard Riemannian optimization.
    Used in 'Riemannian Noise Update' and Appendix A (Absil et al. 2008; Boumal 2023).

pith-pipeline@v1.3.0-alltime-deepseek · 16168 in / 12335 out tokens · 114093 ms · 2026-07-31T23:27:42.081096+00:00 · methodology

0 comments
read the original abstract

Few-step distilled diffusion models generate high-quality images quickly, but often lose per-prompt diversity, producing near-identical samples across random seeds. Optimizing the initial noise at inference time offers an appealing way to recover this diversity, yet existing methods directly update the initial noise in an unconstrained Euclidean space, ignoring both the geometry of the Gaussian prior and the model's sensitivity to noise frequencies. They therefore introduce auxiliary quality-control objectives to maintain generation fidelity, adding compute and weighting hyperparameters while still requiring conservative updates to prevent degradation. In this work, we propose MoNO, a training-free method that performs Manifold-constrained Noise Optimization on a low-dimensional, quality-stabilizing noise manifold. MoNO sequentially optimizes each new initial noise so that its predicted visual feature complements previous generations, while Riemannian updates on an affine low-frequency sphere preserve prior likelihood and fix unstable high-frequency components by construction. This enables large geodesic steps, removes the need for auxiliary quality-control objectives, and converges in far fewer iterations than prior noise-optimization methods. Experiments with multiple distilled text-to-image diffusion models show that MoNO consistently improves per-prompt diversity while maintaining image quality.

Figures

Figures reproduced from arXiv: 2607.23937 by Cheng Jin, Qitan Shi, Yuantao Gu, Ziyuan Liu.

Figure 1
Figure 1. Figure 1: Generations for the prompt “cat with sunglasses” with FLUX.2 [klein] 4B. (a) Under i.i.d. sampling, few-step distilled models collapse in diversity and produce visually similar images. (b) Our method changes only the initial noise, yielding more diverse generations. can be grouped into three families. Guidance methods modify the denoising dynamics at inference time, for example by perturbing the conditioni… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MoNO. Given a new initial noise [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effects of noise norm and update frequency on generation quality for the prompt [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of per-prompt generation diversity on FLUX.2 [klein] 4B and SDXL-Turbo for the prompt [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Optimization progress of three different noise [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the proxy predictor using FLUX.2 [klein] 4B. The first row shows images generated by the full [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Hyperparameter sweeps of MoNO over rotation angle, low-frequency cutoff, and optimization iterations. Each plot [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison between i.i.d. sampling and MoNO (1-iter) using FLUX.2 [klein] 4B for the prompt [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison between i.i.d. sampling and MoNO (1-iter) using SDXL-Turbo for the prompt [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison between i.i.d. sampling and MoNO (1-iter) using DMD2 for the prompt [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison between i.i.d. sampling and MoNO (4-iter) using FLUX.2 [klein] 4B for the prompt [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison between i.i.d. sampling and MoNO (4-iter) using SDXL-Turbo for the prompt [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison between i.i.d. sampling and MoNO (4-iter) using DMD2 for the prompt [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 12 linked inside Pith

  1. [1]

    European Conference on Computer Vision , pages=

    Adversarial diffusion distillation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  2. [2]

    2025 , howpublished=

    2025

  3. [3]

    Advances in neural information processing systems , volume=

    Improved distribution matching distillation for fast image synthesis , author=. Advances in neural information processing systems , volume=

  4. [4]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Distilling diversity and control in diffusion models , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  5. [5]

    arXiv preprint arXiv:2602.03139 , year=

    Diversity-Preserved Distribution Matching Distillation for Fast Visual Synthesis , author=. arXiv preprint arXiv:2602.03139 , year=

    Pith/arXiv arXiv

  6. [6]

    x-Distill: Breaking the Diversity, Quality, and Efficiency Barrier in Distribution Matching Distillation , author=

    1. x-Distill: Breaking the Diversity, Quality, and Efficiency Barrier in Distribution Matching Distillation , author=. arXiv preprint arXiv:2604.04018 , year=

    Pith/arXiv arXiv

  7. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  8. [8]

    arXiv preprint arXiv:2510.03813 , year=

    Diverse Text-to-Image Generation via Contrastive Noise Optimization , author=. arXiv preprint arXiv:2510.03813 , year=

    arXiv

  9. [9]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    Pith/arXiv arXiv

  10. [10]

    International Conference on Learning Representations , volume=

    Particle guidance: non-iid diverse sampling with diffusion models , author=. International Conference on Learning Representations , volume=

  11. [11]

    arXiv preprint arXiv:2605.06553 , year=

    Diverse Sampling in Diffusion Models with Marginal Preserving Particle Guidance , author=. arXiv preprint arXiv:2605.06553 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2508.15773 , year=

    Scaling group inference for diverse and high-quality generation , author=. arXiv preprint arXiv:2508.15773 , year=

    Pith/arXiv arXiv

  13. [13]

    International Conference on Learning Representations , volume=

    CADS: Unleashing the diversity of diffusion models through condition-annealed sampling , author=. International Conference on Learning Representations , volume=

  14. [14]

    University of California, Irvine , volume=

    High-dimensional probability , author=. University of California, Irvine , volume=

  15. [15]

    arXiv preprint arXiv:2505.11278 , year=

    A fourier space perspective on diffusion models , author=. arXiv preprint arXiv:2505.11278 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2605.07253 , year=

    LENS: Low-Frequency Eigen Noise Shaping for Efficient Diffusion Sampling , author=. arXiv preprint arXiv:2605.07253 , year=

    Pith/arXiv arXiv

  17. [17]

    2008 , publisher=

    Optimization algorithms on matrix manifolds , author=. 2008 , publisher=

    2008

  18. [18]

    2023 , publisher=

    An introduction to optimization on smooth manifolds , author=. 2023 , publisher=

    2023

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    Geneval: An object-focused framework for evaluating text-to-image alignment , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  21. [21]

    arXiv preprint arXiv:2306.09344 , year=

    Dreamsim: Learning new dimensions of human visual similarity using synthetic data , author=. arXiv preprint arXiv:2306.09344 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2210.02410 , year=

    The vendi score: A diversity evaluation metric for machine learning , author=. arXiv preprint arXiv:2210.02410 , year=

    Pith/arXiv arXiv

  23. [23]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    Clipscore: A reference-free evaluation metric for image captioning , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  24. [24]

    arXiv preprint arXiv:2306.09341 , year=

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis , author=. arXiv preprint arXiv:2306.09341 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2412.01339 , year=

    Negative token merging: Image-based adversarial feature guidance , author=. arXiv preprint arXiv:2412.01339 , year=

    Pith/arXiv arXiv

  26. [26]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  27. [27]

    arXiv preprint arXiv:2403.05135 , year=

    Ella: Equip diffusion models with llm for enhanced semantic alignment , author=. arXiv preprint arXiv:2403.05135 , year=

    Pith/arXiv arXiv

  28. [28]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Exploring clip for assessing the look and feel of images , author=. Proceedings of the AAAI conference on artificial intelligence , volume=