Pith. sign in

REVIEW 5 major objections 5 minor 53 references

This paper claims that text-to-image diversity can be decomposed along two orthogonal axes in CLIP space and actively expanded at inference time—without retraining—to produce more varied images, including prompt-independent background varia

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

2026-08-02 22:16 UTC pith:CTUZUWTL

load-bearing objection A clean geometric idea whose specific disentanglement claim outruns its evidence; the residual-axis identification and the proof both need work, but the paper deserves a serious round of review. the 5 major comments →

arxiv 2602.17200 v2 pith:CTUZUWTL submitted 2026-02-19 cs.CV

GASS: Geometry-Aware Spherical Sampling for Disentangled Diversity Enhancement in Text-to-Image Generation

Ye Zhu , Kaleb S. Newman , Johannes F. Lutzeyer , Adriana Romero-Soriano , Michal Drozdzal , Olga Russakovsky This is my paper
classification cs.CV
keywords text-to-image generationdiversity enhancementCLIP embedding geometryspherical samplingorthogonal decompositioninference-time guidancediffusion modelsflow models
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 asserts that the limited diversity of text-to-image models under a fixed prompt is not a single scalar problem: a batch of generated images varies along a direction aligned with the prompt (pose, layout, object identity) and along a perpendicular direction that captures prompt-independent attributes such as backgrounds and style. Working in the unit sphere of CLIP embeddings, it decomposes each image embedding into its projection onto the text embedding and its projection onto a dominant residual axis found by a Gram-Schmidt search, and defines a diversity score as the sum of projection spreads along these two axes. The proposed method, GASS, then expands the batch's spread along both axes during inference—perturbing target embeddings, re-normalizing, and using gradient updates from a frozen CLIP image encoder—and reports higher diversity scores with small quality/alignment loss across two frozen backbones and two benchmarks. If the claim holds, diversity becomes a controllable, training-free dial, and background variation can be increased without editing the prompt.

Core claim

The central claim is that the diversity of a generated image batch can be separated into prompt-dependent and prompt-independent components in CLIP space, and that actively widening the geometric spread along the corresponding two orthogonal directions at inference time increases diversity without retraining the generative model. For a fixed text prompt, let et be the normalized text embedding and uind an orthonormal vector found by maximizing the mean absolute projection of batch embeddings in the subspace orthogonal to et. GASS rewrites each image embedding as its projection onto et plus its projection onto uind plus a residual, applies uniform shifts to the two projections, re-normalizes

What carries the argument

The central object is the orthogonal pair (et, uind) inside CLIP's unit hypersphere. et is the text embedding that anchors prompt-aligned semantics; uind is a unit vector orthogonal to et selected from N=10 Gram-Schmidt candidate directions by maximizing the average absolute projection of the batch (Algorithm 1), intended to capture the dominant prompt-independent mode (background, style). The mechanism is the GASS expansion: each embedding is decomposed as (ei^T et)et + (ei^T uind)uind + ri, the two coefficients are shifted by uniform perturbations with ranges rdep and rind (Eqs. 5–6), and the re-normalized vectors serve as targets. The targets are enforced on the sampling trajectory by a b

Load-bearing premise

The load-bearing premise is that one dominant residual direction, found by a quick random search over ten Gram-Schmidt candidates, reliably captures the prompt-independent variety of the batch, and that widening the batch's spread in CLIP space transfers through pixel-space gradient updates into real image diversity without collapsing fidelity.

What would settle it

Generate a batch from a fixed prompt, remove the text-embedding component from each image embedding, and compute the principal component of the residual covariance. If that top eigenvector captures much more variance than the direction GASS selects with N=10 Gram-Schmidt candidates—or if expanding along GASS's direction leaves background/stylistic attributes unchanged while expanding along the top eigenvector changes them—then the claimed disentanglement and background-diversity control would be refuted.

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

If this is right

  • Diversity gains are available at inference time on frozen models: no retraining or backbone modification is required, so the method can be layered onto existing text-to-image systems.
  • The source of diversity becomes controllable: adjusting the expansion ranges along et and uind respectively amplifies prompt-dependent variation (pose, layout) or prompt-independent variation (backgrounds, styles).
  • The spherical spread score provides a reference-free, decomposition-based diversity measure that separates semantic from non-semantic variation; the paper shows real images score roughly 50% higher on it than images generated by the two tested backbones.
  • Sparse guidance suffices: applying the intervention over only about 15–20 sampling steps achieves most of the diversity gain, bounding the added per-batch cost to roughly 3.7 seconds versus 1.7 seconds in the reported setup.

Where Pith is reading between the lines

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

  • [Editorial inference] If the dominant-residual assumption holds more broadly, the same two-axis decomposition could serve as a training-free diversity and bias lever in other CLIP-conditioned generators beyond text-to-image sampling, such as image editing and personalization.
  • [Editorial inference] Because CLIP defines which attributes count as prompt-independent, uind is encoder-specific; replacing CLIP with another multimodal encoder would shift the factors that get diversified, providing a direct test of how much of the effect is geometric versus tied to CLIP's similarity metric.
  • [Editorial inference] The hypervolume guarantee suggests per-step spread gains could compound; a natural follow-up is to measure whether repeated GASS applications saturate as the batch approaches the high-density boundary of the CLIP manifold, which would bound the practical diversity ceiling.
  • [Editorial inference] The motivation about societal bias suggests a testable fairness extension: measuring whether expanding along uind systematically reduces stereotyped background or context associations in generated images.

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

5 major / 5 minor

Summary. The paper proposes GASS, an inference-time method intended to increase the diversity of frozen text-to-image models by operating on CLIP embeddings. It decomposes the diversity of a generated batch into a prompt-dependent component (projections onto the text embedding et) and a prompt-independent component (projections onto a single direction uind found by random search over N=10 Gram-Schmidt candidates), and defines SPP = Ddep + Dind. At selected sampling steps, GASS expands each image embedding along et and uind via random shifts, renormalizes, and uses CLIP gradients to move the predicted clean image toward the expanded target. Experiments on SD2.1 and SD3-M over ImageNet and DrawBench compare against PG, CADS, IG, and SPELL, reporting Density, Coverage, Vendi Score, ClipScore, ImageReward, and SPP.

Significance. If the residual-axis identification and the volume guarantee were established, GASS would be a useful controllable diversity intervention with modest quality cost, applicable to frozen U-Net and DiT backbones and distinct from entropy-only methods. The paper's strengths include experiments across two backbones and two benchmarks, non-cherry-picked qualitative comparisons, sparse intervention for reduced cost, and an explicit limitations section. However, the central identification mechanism is not validated, the proof of Proposition 4.1 is incomplete, and the headline metric is closely aligned with the method's objective, so the claimed disentanglement currently rests on unsupported premises.

major comments (5)
  1. [Sec. 3.2, Algo. 1, Eq. (3)] The claim that N=10 random Gram-Schmidt candidates robustly identify the dominant residual axis is not supported and is likely false in d≈768 CLIP space. A random unit tangent vector has typical absolute projection ≈ 1/√d ≈ 0.036; the maximum over 10 candidates is only ≈0.06, far below the alignment expected for a true dominant mode. Thus uind is probably an essentially arbitrary tangent direction, and Dind measures spread along that arbitrary axis rather than the prompt-independent mode. Please validate the selection by comparing uind with PCA/SVD of the batch residuals, reporting the fraction of residual variance captured, and providing a sensitivity analysis over N, or replace the random search with a deterministic PCA-based direction.
  2. [Appendix A, Prop. 4.1, Eqs. (13)-(19)] The proof of the expected hypervolume guarantee is incomplete. In Step 2.4, the claim that the cross terms in v^T(ΔG)v 'preserve non-negativity in expectation' is asserted without proof, and the phrase 'we align our perturbations with the high-variance basis directions' is circular with the unvalidated uind selection. Moreover, Eq. (19) moves the expectation inside the determinant, which is not valid in general. Additionally, Eq. (13) has an index error: the coefficient of et should be δdep_i − δdep_j, not δdep_i − δind_j. The proposition should be given a rigorous proof or explicitly downgraded to an empirical observation.
  3. [Sec. 5.2, Tables 2-3] SPP is the quantity GASS is designed to increase through Eqs. (5)-(7), so reporting SPP gains does not independently validate the method. The independent metrics show smaller and sometimes inconsistent improvements — e.g., SD3-M Vendi Score for GASS is 28.877 vs 28.456 for CADS, and SD2.1 Coverage is 0.603 vs 0.614 for PG. No error bars, repeated-seed statistics, or significance tests are reported anywhere. Please provide mean±std over multiple seeds and significance tests for VS, Coverage, and SPP.
  4. [Sec. 5.3, Table 4] The ablation differences are within the range one would expect from noise without significance testing. For example, rdep=0, rind=0.02 gives SPP 0.111 and VS 8.207, while the full setting gives SPP 0.114 and VS 8.212; these differences are tiny. The claim that expanding along both axes yields the best overall diversity gains is therefore not established. Also, the 'w/o Norm' variant has the highest VS (8.876) but the lowest SPP (0.123 vs 0.122? reported 0.123), which complicates the interpretation. Report variances and explain the trade-offs.
  5. [Algo. 3, Stage 2] There is an inconsistency between Algorithm 1 and Algorithm 3. Algo. 1 defines E_k = (1/B) Σ_i |e_i^T r_k|, but Algo. 3 computes E_k = |e_{i,0|t}^T r_k| without any averaging over the batch. This changes which candidate is selected as uind. The complete algorithm should match the definition in Eq. (3) and Algo. 1.
minor comments (5)
  1. [Eq. (13)] Index typo: the first perturbation in the et coefficient should be δdep_i − δdep_j, not δdep_i − δind_j.
  2. [Sec. 5.1] The implementation details refer to 'SD3.5-M' while the rest of the paper uses 'SD3-M' (SD3 Medium). Please make the naming consistent.
  3. [Eq. (3)] The selection criterion maximizes mean absolute projection rather than variance. Since the text says 'dominant residual basis' and 'captured energy,' please justify the L1-type criterion or switch to variance.
  4. [Appendix C.1] The baseline is inconsistently spelled 'CADs' and 'CADS'; use one spelling.
  5. [Fig. 4] The qualitative disentanglement claim (et→pose/layout, uind→background/style) would be stronger with a quantitative verification, e.g., attribute classifiers or caption-based analysis, because an arbitrary residual axis can also correlate with prompt-dependent factors.

Circularity Check

1 steps flagged

SPP gains are the GASS optimization objective by construction; independent metrics are smaller/mixed, so the circularity is substantial but not total.

specific steps
  1. fitted input called prediction [Sec. 4 (introduction), Sec. 4.1 Eq. 6, Sec. 4.2 Eq. 7; reported as SPP in Tab. 2-3]
    "Building on our geometric analysis above, we introduce GASS ... Formally, we aim to increase the diversity score SP P defined in Sec. 3.3, thereby pushing the newly generated set X ′ = {x ′ i}B i=1 to cover a wider spread of the manifold measured on the CLIP sphere. ... To align the generation with our diversity target ˜e after GASS guidance, we define a batch-wise loss LSPP that measures the alignment between the current estimated embedding and the updated target after geometric expansion: LSPP = sum_i (1− EI(ˆxi,0|t)⊤˜ei)."

    SPP is defined in Eq. 4 as Ddep + Dind, i.e., the sum of projection spreads along et and uind. Eq. 6 constructs the target embeddings etilde_i by adding uniform shifts δdep, δind along exactly these two axes, and Eq. 7-8 optimize the predicted clean images to maximize cosine similarity to etilde_i. Therefore any increase in SPP (and in its Ddep/Dind components) is the optimization objective itself, not an independent validation. Reporting SPP as evidence of 'effectiveness of disentangled diversity enhancement' is self-validating. Independent metrics (VS, Coverage, ImageReward, ClipScore) are not optimized directly and show smaller or mixed gains, so the circularity is partial rather than total.

full rationale

The core circular step is that GASS explicitly optimizes the same spread score (SPP) it later reports as a diversity improvement: Sec. 4 states the aim is to increase SPP, Eq. 6 expands embeddings along the SPP axes, and Eq. 7 minimizes alignment loss to those expanded targets. Hence the SPP/Dind/Ddep gains in Tabs. 2-3 are forced by construction. I did not flag the uind random-search choice (Algo. 1, N=10) as circular: it is a correctness/robustness concern, not a definitional reduction. Similarly, Proposition 4.1's proof has an unsupported step ('these cross terms preserve non-negativity in expectation'), but that is an unproven assumption rather than circularity. The paper does provide independent evidence: VS is an external entropy-based metric computed on Inception features and shows gains; ImageReward/ClipScore are not optimized and remain competitive. These keep the central claim partially independent, so the score is 6 rather than higher. Appendix D's candid failure-case discussion does not change the assessment.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The method rests on one fitted expansion-range pair and a hand-set candidate count N=10, plus the domain assumption that CLIP-space orthogonality to the text direction isolates prompt-independent variation. The invented entity uind is not independently validated. The theoretical proof adds an additional unsupported assumption that the Gram-matrix cross terms are non-negative in expectation.

free parameters (5)
  • rdep (expansion range along text axis et) = 0.02
    Chosen by ablation on DrawBench (Tab. 4); controls how much the projection onto the text direction is expanded.
  • rind (expansion range along residual axis uind) = 0.02
    Chosen by ablation on DrawBench (Tab. 4); controls how much the projection onto the prompt-independent direction is expanded.
  • N (number of candidate residual directions) = 10
    Set empirically in Sec. 3.2 with the statement that N=10 is sufficient; no sensitivity analysis is provided.
  • Number of GASS intervention steps t = 20
    Ablated in Tab. 4 (t=10, 15, 20); default set to 20 for main experiments.
  • Optimization hyperparameters (learning rate, max steps, early stopping) = lr=1e-4, max 60 steps, tolerance 5e-4, patience 4
    Standard optimization choices stated in Sec. 5.1; not varied in ablations.
axioms (5)
  • domain assumption CLIP embeddings are normalized to a unit hypersphere and inner-product alignment with the text embedding corresponds to semantic consistency.
    Section 3.2 adopts the CLIP embedding space as the metric space for disentangling prompt-dependent vs prompt-independent variation.
  • domain assumption Deep representation variation lies on a low-dimensional manifold, so a single dominant residual direction uind suffices to capture prompt-independent diversity.
    Invoked in Sec. 3.2 when simplifying the full (d-1)-dimensional residual subspace to one dominant basis vector.
  • ad hoc to paper Random search over N=10 Gram-Schmidt candidate directions robustly identifies the dominant residual axis.
    Sec. 3.2 asserts this empirically without providing a distribution or stability analysis; the central Dind metric depends on it.
  • domain assumption Steering the predicted clean image through CLIP gradients in pixel space transfers CLIP-space expansion into real image diversity without collapsing fidelity.
    Sec. 4.2 relies on frozen CLIP image encoder gradients; there is no decoder from CLIP space back to pixels, so this transfer is assumed.
  • standard math Linear algebra facts used in Appendix A: Gram determinant gives hypervolume; positive-semidefinite perturbation increases determinant; E[sqrt(X)] >= sqrt(E[X]).
    The first two are standard, but the Jensen inequality is applied in the wrong direction, making the proof invalid as written.
invented entities (1)
  • uind (dominant residual direction) no independent evidence
    purpose: A constructed unit vector orthogonal to the text embedding, intended to capture prompt-independent image variation such as background and style.
    uind is estimated from the current batch via random search; it has no external validation against attribute labels, held-out axes, or other falsifiable handles outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 18102 in / 12153 out tokens · 112231 ms · 2026-08-02T22:16:59.376992+00:00 · methodology

0 comments
read the original abstract

Despite high semantic alignment, modern text-to-image (T2I) generative models still struggle to synthesize diverse images from a given prompt. In this work, we enhance the T2I diversity through a geometric lens. Unlike most existing methods that rely primarily on entropy-based guidance to increase sample dissimilarity, we introduce Geometry-Aware Spherical Sampling (GASS) to enhance diversity by explicitly controlling both prompt-dependent and prompt-independent sources of variation. Specifically, we decompose the diversity measure in CLIP embeddings using two orthogonal directions: the text embedding, which captures semantic variation related to the prompt, and an identified orthogonal direction that captures prompt-independent variation (e.g., backgrounds). Based on this decomposition, GASS increases the geometric projection spread of generated image embeddings along both axes and guides the T2I sampling process via expanded predictions along the generation trajectory. Our experiments on different frozen T2I backbones (U-Net and DiT, diffusion and flow) and benchmarks demonstrate the effectiveness of disentangled diversity enhancement with minimal impact on image fidelity and semantic alignment.

Figures

Figures reproduced from arXiv: 2602.17200 by Adriana Romero-Soriano, Johannes F. Lutzeyer, Kaleb S. Newman, Michal Drozdzal, Olga Russakovsky, Ye Zhu.

Figure 1
Figure 1. Figure 1: Illustration of our geometric decomposition of sam￾ple diversity and GASS enhancement method in CLIP space. We decompose the diversity of generated image batches from T2I models in the CLIP hypersphere along two orthogonal axes: text embedding et (i.e., prompt-dependent) and our identified direction uind (i.e., prompt-independent). Our GASS method explicitly ex￾pands the geometric spread along both axes, t… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of our proposed Geometry-Aware Spherical Sampling (GASS) method. At the generation inference step t, original T2I sampling first estimates the predicted clean image xˆ0|t based on the intermediate noisy samples xt, and then predict the noise we should remove from xt to get xt−1. Our GASS alters the predicted clean image from xˆ0|t to xˆ ∗ 0|t through the geometric expansion (see Sec. 4.1) and … view at source ↗
Figure 3
Figure 3. Figure 3: Non-cherry-picked qualitative comparisons with other diversity enhancement methods on ImageNet (Russakovsky et al., 2015) and Drawbench (Saharia et al., 2022). Compared to other methods (i.e., PG (Corso et al., 2024), CADS (Sadat et al., 2024), IG (Kynka¨anniemi et al. ¨ , 2024), and SPELL (Kirchhof et al., 2025)), our proposed GASS generates images with both richer semantic variation (e.g., object poses a… view at source ↗
Figure 4
Figure 4. Figure 4: GASS controls the source of diversity by expanding the geometric spread along specified directions. Specially, GASS on prompt-dependent axis et diversifies images through variations via poses and layout, while expansion along prompt-independent direction uind changes attributes like background and styles. typically incur quality drops across different methods. In general, our approach achieves superior div… view at source ↗
Figure 5
Figure 5. Figure 5: GASS still introduces generated image diversity, even when provided with more complex text prompts. Expansion Range. In our proposed sampling method, we define expansion ranges via hyperparameters rdep and rind along et and uind, respectively. Ablation studies in Tab. 4 show that r = 0.02 achieves optimal trade-offs across dif￾ferent evaluation metrics. In addition, expanding along both axes yields the bes… view at source ↗
Figure 6
Figure 6. Figure 6: Additional non-cherry-picked qualitative results comparison with other methods on ImageNet with the example class ”goldfish”. D. Further Discussions D.1. Limitations While our proposed GASS method effectively enhances T2I diversity and explores the residual space beyond the given prompts, similar to other sampling-based post-training guidance methods, it incurs extra inference time compared to the original… view at source ↗
Figure 7
Figure 7. Figure 7: Additional non-cherry-picked qualitative results comparison with other methods on Drawbench. While the probability of both perturbations being simultaneously close to zero is very low, such cases do occur, which result in rather trivial modifications after GASS. Importantly, these isolated instances do not impact the overall batch-level diversity metrics, as diversity is measured across the entire batch ra… view at source ↗
Figure 8
Figure 8. Figure 8: Failure case analysis. (Left): A small number of images still resemble the original ones after GASS. (Right:) When the base models can’t generate accurate counts specified by the prompt, despite GASS introducing extra diversity, it is less likely to correct these inconsistencies by itself. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.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

53 extracted references · 4 linked inside Pith

  1. [1]

    Improving geo-diversity of generated images with contextualized vendi score guidance

    Askari Hemmat, R., Hall, M., Sun, A., Ross, C., Drozdzal, M., and Romero-Soriano, A. Improving geo-diversity of generated images with contextualized vendi score guidance. In ECCV, pp.\ 213--229. Springer, 2024

    2024

  2. [2]

    R., and Drozdzal, M

    Astolfi, P., Careil, M., Hall, M., Ma \ n as, O., Muckley, M., Verbeek, J., Soriano, A. R., and Drozdzal, M. Consistency-diversity-realism pareto fronts of conditional image generative models. arXiv preprint arXiv:2406.10429, 2024

    Pith/arXiv arXiv 2024

  3. [3]

    A., Krause, F., Neumayr, M., Stracke, N., Sevi, M., Hu, V

    Baumann, S. A., Krause, F., Neumayr, M., Stracke, N., Sevi, M., Hu, V. T., and Ommer, B. Continuous, subject-specific attribute control in t2i models by identifying semantic directions. In CVPR, pp.\ 13231--13241, 2025

    2025

  4. [4]

    Representation learning: A review and new perspectives

    Bengio, Y., Courville, A., and Vincent, P. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35 0 (8): 0 1798--1828, 2013

    2013

  5. [5]

    Entropy rectifying guidance for diffusion and flow models

    Berrada, T., Romero-Soriano, A., Drozdzal, M., Verbeek, J., and Alahari, K. Entropy rectifying guidance for diffusion and flow models. In NeurIPS, 2025

    2025

  6. [6]

    Diversity-rewarded cfg distillation

    Cideron, G., Agostinelli, A., Ferret, J., Girgin, S., Elie, R., Bachem, O., Perrin, S., and Ram \'e , A. Diversity-rewarded cfg distillation. arXiv preprint arXiv:2410.06084, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Corso, G., Xu, Y., De Bortoli, V., Barzilay, R., and Jaakkola, T. S. Particle guidance: non-iid diverse sampling with diffusion models. In ICLR, 2024

    2024

  8. [8]

    A., Hall, M., Verbeek, J., Romero-Soriano, A., and Drozdzal, M

    Dall'Asen, N., Zhang, X., Hemmat, R. A., Hall, M., Verbeek, J., Romero-Soriano, A., and Drozdzal, M. Increasing the utility of synthetic images through chamfer guidance. In NeurIPS, 2025

    2025

  9. [9]

    Imagenet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In CVPR, pp.\ 248--255. Ieee, 2009

    2009

  10. [10]

    Image generation diversity issues and how to tame them

    Dombrowski, M., Zhang, W., Cechnicka, S., Reynaud, H., and Kainz, B. Image generation diversity issues and how to tame them. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 3029--3039, June 2025

    2025

  11. [11]

    Scaling rectified flow transformers for high-resolution image synthesis

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024

    2024

  12. [12]

    and Dieng, A

    Friedman, D. and Dieng, A. B. The vendi score: A diversity evaluation metric for machine learning. Transactions on Machine Learning Research, 2023

    2023

  13. [13]

    Clipscore: A reference-free evaluation metric for image captioning

    Hessel, J., Holtzman, A., Forbes, M., Le Bras, R., and Choi, Y. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, pp.\ 7514--7528, 2021

    2021

  14. [14]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017

    2017

  15. [15]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  16. [16]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. NeurIPS, 2020

    2020

  17. [17]

    Horn, R. A. and Johnson, C. R. Matrix analysis. Cambridge university press, 2012

    2012

  18. [18]

    Sparke: Scalable prompt-aware diversity and novelty guidance in diffusion models via rke score

    Jalali, M., Haoyu, L., Gohari, A., and Farnia, F. Sparke: Scalable prompt-aware diversity and novelty guidance in diffusion models via rke score. In NeurIPS, 2025 a

    2025

  19. [19]

    Conditional vendi score: An information-theoretic approach to diversity evaluation of prompt-based generative models

    Jalali, M., Ospanov, A., Gohari, A., and Farnia, F. Conditional vendi score: An information-theoretic approach to diversity evaluation of prompt-based generative models. In CVPR, 2025 b

    2025

  20. [20]

    Shielded diffusion: Generating novel and diverse images using sparse repellency

    Kirchhof, M., Thornton, J., B \'e thune, L., Ablin, P., Ndiaye, E., et al. Shielded diffusion: Generating novel and diverse images using sparse repellency. In ICML, 2025

    2025

  21. [21]

    Improved precision and recall metric for assessing generative models

    Kynk \"a \"a nniemi, T., Karras, T., Laine, S., Lehtinen, J., and Aila, T. Improved precision and recall metric for assessing generative models. NeurIPS, 32, 2019

    2019

  22. [22]

    Applying guidance in a limited interval improves sample and distribution quality in diffusion models

    Kynk \"a \"a nniemi, T., Aittala, M., Karras, T., Laine, S., Aila, T., and Lehtinen, J. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. NeurIPS, 37: 0 122458--122483, 2024

    2024

  23. [23]

    J., Bj \"o rck, ., and Gander, W

    Leon, S. J., Bj \"o rck, ., and Gander, W. Gram-schmidt orthogonalization: 100 years and more. Numerical Linear Algebra with Applications, 20 0 (3): 0 492--532, 2013

    2013

  24. [24]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. In ICLR, 2023

    2023

  25. [25]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and Liu, Q. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023

    2023

  26. [26]

    Training diffusion models towards diverse image generation with reinforcement learning

    Miao, Z., Wang, J., Wang, Z., Yang, Z., Wang, L., Qiu, Q., and Liu, Z. Training diffusion models towards diverse image generation with reinforcement learning. In CVPR, pp.\ 10844--10853, 2024

    2024

  27. [27]

    F., Oh, S

    Naeem, M. F., Oh, S. J., Uh, Y., Choi, Y., and Yoo, J. Reliable fidelity and diversity metrics for generative models. In ICML, pp.\ 7176--7185. PMLR, 2020 a

    2020

  28. [28]

    F., Oh, S

    Naeem, M. F., Oh, S. J., Uh, Y., Choi, Y., and Yoo, J. Reliable fidelity and diversity metrics for generative models. In ICML, pp.\ 7176--7185. PMLR, 2020 b

    2020

  29. [29]

    and Nushi, B

    Naik, R. and Nushi, B. Social biases through the text-to-image generation lens. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pp.\ 786--808, 2023

    2023

  30. [30]

    and Mitter, S

    Narayanan, H. and Mitter, S. Sample complexity of testing the manifold hypothesis. NeurIPS, 23, 2010

    2010

  31. [31]

    and Farnia, F

    Ospanov, A. and Farnia, F. Do vendi scores converge with finite samples? truncated vendi score for finite-sample convergence guarantees. In The 41st Conference on Uncertainty in Artificial Intelligence, 2024

    2024

  32. [32]

    Towards a scalable reference-free evaluation of generative models

    Ospanov, A., Zhang, J., Jalali, M., Cao, X., Bogdanov, A., and Farnia, F. Towards a scalable reference-free evaluation of generative models. NeurIPS, 37: 0 120892--120927, 2024

    2024

  33. [33]

    Scendi score: Prompt-aware diversity evaluation via schur complement of clip embeddings

    Ospanov, A., Jalali, M., and Farnia, F. Scendi score: Prompt-aware diversity evaluation via schur complement of clip embeddings. In CVPR, pp.\ 16927--16937, 2025

    2025

  34. [34]

    J., Mohamed, S., and Lakshminarayanan, B

    Papamakarios, G., Nalisnick, E., Rezende, D. J., Mohamed, S., and Lakshminarayanan, B. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research, 22 0 (57): 0 1--64, 2021

    2021

  35. [35]

    Understanding the latent space of diffusion models through the lens of riemannian geometry

    Park, Y.-H., Kwon, M., Choi, J., Jo, J., and Uh, Y. Understanding the latent space of diffusion models through the lens of riemannian geometry. NeurIPS, 36: 0 24129--24142, 2023

    2023

  36. [36]

    Pasarkar, A. P. and Dieng, A. B. Cousins of the vendi score: A family of similarity-based diversity metrics for science and machine learning. In International Conference on Artificial Intelligence and Statistics, pp.\ 3808--3816. PMLR, 2024

    2024

  37. [37]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In ICCV, pp.\ 4195--4205, 2023

    2023

  38. [38]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In ICML, pp.\ 8748--8763, 2021

    2021

  39. [39]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In CVPR, pp.\ 10684--10695, 2022

    2022

  40. [40]

    U-net: Convolutional networks for biomedical image segmentation

    Ronneberger, O., Fischer, P., and Brox, T. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pp.\ 234--241. Springer, 2015

    2015

  41. [41]

    Imagenet large scale visual recognition challenge

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. Imagenet large scale visual recognition challenge. IJCV, 115 0 (3): 0 211--252, 2015

    2015

  42. [42]

    Sadat, S., Buhmann, J., Bradley, D., Hilliges, O., and Weber, R. M. Cads: Unleashing the diversity of diffusion models through condition-annealed sampling. In ICLR, 2024

    2024

  43. [43]

    L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E. L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022

    2022

  44. [44]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. In ICLR, 2021

    2021

  45. [45]

    Introduction to linear algebra

    Strang, G. Introduction to linear algebra. SIAM, 2022

    2022

  46. [46]

    Rethinking the inception architecture for computer vision

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna, Z. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 2818--2826, 2016

    2016

  47. [47]

    Survey of bias in text-to-image generation: Definition, evaluation, and mitigation

    Wan, Y., Subramonian, A., Ovalle, A., Lin, Z., Suvarna, A., Chance, C., Bansal, H., Pattichis, R., and Chang, K.-W. Survey of bias in text-to-image generation: Definition, evaluation, and mitigation. arXiv preprint arXiv:2404.01030, 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation

    Wang, R., Yang, Y., Qian, Z., Zhu, Y., and Wu, Y. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024

    2024

  49. [49]

    The silent assistant: Noisequery as implicit guidance for goal-driven image generation

    Wang, R., Huang, H., Zhu, Y., Russakovsky, O., and Wu, Y. The silent assistant: Noisequery as implicit guidance for goal-driven image generation. In ICCV, 2025

    2025

  50. [50]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., and Dong, Y. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurIPS, 36: 0 15903--15935, 2023

    2023

  51. [51]

    The intricate dance of prompt complexity, quality, diversity, and consistency in t2i models

    Zhang, X., Courville, A., Drozdzal, M., and Romero-Soriano, A. The intricate dance of prompt complexity, quality, diversity, and consistency in t2i models. arXiv preprint arXiv:2510.19557, 2025

    arXiv 2025

  52. [52]

    Boundary guided learning-free semantic control with diffusion models

    Zhu, Y., Wu, Y., Deng, Z., Russakovsky, O., and Yan, Y. Boundary guided learning-free semantic control with diffusion models. In NeurIPS, 2023

    2023

  53. [53]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...