Pith. sign in

REVIEW 4 major objections 5 minor 91 references

A three-layer probe of few-step diffusion models separates prediction type from distillation objective across 23 text-to-image checkpoints.

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 · grok-4.5

2026-07-12 03:45 UTC pith:E7QYZWD2

load-bearing objection A careful, estimator-matched diagnostic that cleanly separates prediction type from distillation objective inside a 23-model sweep; the RF latent band is real in the data, but the causal attribution still rests on an ad-hoc rule and thin non-UNet controls. the 4 major comments →

arxiv 2607.03256 v1 pith:E7QYZWD2 submitted 2026-07-03 cs.CV

A Decomposable Probe for Few-Step Diffusion Models: Prompt, Latent, and Score Selectivity across Backbone Families and Distillation Paradigms

Patrick Mu Haojie This is my paper
classification cs.CV
keywords few-step diffusiondistillationrectified flowselectivity probeBures Wassersteintext-to-imagelatent diffusionadversarial distillation
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.

Few-step distilled diffusion models cut text-to-image sampling from dozens of network evaluations to a handful, but the quality gap is usually reported as a single FID or CLIP number that cannot say which part of the conditioning response changed. This paper replaces that scalar with a training-free probe that injects controlled mean, variance, and scale perturbations at three sites—the prompt encoder, the denoiser input, and the denoiser output—and summarizes each cell by a bootstrap-median Bures W₂² selectivity ratio on Inception features. Under one matched estimator applied to five teachers and eighteen students spanning UNet, DiT, and MMDiT backbones and five distillation paradigms, the three layers track three separable factors: the prompt layer is a universal mean response (a sanity channel), the latent layer reads prediction type, and the score layer reads the distillation objective. The main result is that only rectified-flow models form a sustained elevated latent selectivity band, and that band survives pure adversarial distillation on both teachers and students; a matched epsilon-prediction T5 control rules out wide text conditioning as the cause. Secondary score-layer patterns, under narrower scope, flag adversarial students at four steps and reveal an early-strength spike for trajectory-rollout students on both UNet and DiT.

Core claim

Within this 23-model sweep the latent-layer selectivity ratio exceeds 1 across a sustained low-to-mid strength band only for rectified-flow backbones (SD3.5, FLUX), as both teachers and adversarially distilled students; no epsilon-prediction model forms that band. A T5-conditioned epsilon-prediction control (PixArt-α) does not reproduce the band, attributing the fingerprint to prediction type rather than wide conditioning, and the fingerprint survives ADD distillation. The score layer separately tracks distillation objective via a 4-step ADD-versus-rest contrast and a CI-separated early-strength spike on trajectory-rollout students.

What carries the argument

A decomposable layer-/mode-resolved probe: controlled mean, variance, and scale perturbations of six strengths injected at prompt encoder, denoiser input (latent), and denoiser output (score), summarized by the bootstrap-median Bures W₂² selectivity ratio R = amp_mean / max(amp_var, amp_scale) on Inception features under one matched estimator across all models.

Load-bearing premise

The claim that the sustained latent band isolates prediction type rests on one epsilon-prediction DiT control and no non-adversarial rectified-flow student, so residual architecture or conditioning factors could still drive the pattern.

What would settle it

An epsilon-prediction model that forms a sustained latent band (R_lo > 1 on at least three of the four strengths in {0.05, 0.1, 0.2, 0.3}, including one s ≥ 0.2) under the same matched estimator, or a non-ADD rectified-flow student that loses that band, would falsify the prediction-type detector.

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

If this is right

  • The latent-layer sustained band can serve as an empirical fingerprint of rectified-flow prediction type that survives pure adversarial distillation.
  • At fixed step count, score-layer ratios can distinguish adversarial-dominated distillation from other paradigms without training logs.
  • Trajectory-rollout objectives leave a detectable CI-separated early-strength score spike on both UNet and DiT.
  • Layer-resolved selectivity ratios can diagnose which axis of conditioning response changed after distillation, unlike single FID/CLIP scalars.
  • The released per-cell tables and matched estimator enable direct, CI-citable cross-model comparison under identical statistics.

Where Pith is reading between the lines

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

  • If the latent band truly isolates prediction type, it could audit released checkpoints whose training recipe is undisclosed without relying on architecture metadata.
  • The same per-layer readings could be used as an online training signal to steer distillation away from collapse on latent or score selectivity relative to the teacher.
  • A non-adversarial rectified-flow student would test whether the latent fingerprint is objective-invariant or only known to survive ADD.
  • An orthogonal diversity channel (for example Vendi on DINOv2 features) could separate sample-diversity collapse from the conditioning-response changes the current probe measures.

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

4 major / 5 minor

Summary. The paper proposes a training-free, decomposable probe for diffusion text-to-image models that injects controlled mean/variance/scale perturbations at three forward-pass layers (prompt encoder, denoiser input/latent, denoiser output/score) and summarises each cell by a bootstrap-median Bures W2^2 selectivity ratio on Inception-v3 features. Under one matched estimator on a 23-model sweep (five teachers, 18 few-step students; SDXL/SD1.5/SD3.5/PixArt-α/FLUX; UNet/DiT/MMDiT; five distillation paradigms), the three layers are reported to track three empirically separable factors: a universal prompt-mean response, prediction type (rectified-flow vs ε-prediction) on the latent layer, and distillation objective on the score layer. The main claim is that, within this sweep, a sustained low-to-mid latent band (R_lo>1 on a stated multi-strength rule) appears only for rectified-flow backbones (SD3.5, FLUX) as both teachers and ADD students; PixArt-α (T5, ε-prediction) is used to rule out wide-T5 conditioning alone. Two narrower score-layer findings are a 4-step ADD-vs-rest contrast on UNet families and a CI-separated early-strength spike on trajectory-rollout students (UNet and DiT). Per-cell CI tables and the estimator are released.

Significance. If the probe is reliable, it is a useful diagnostic instrument for a literature that still reports few-step quality almost exclusively via end-to-end FID/CLIP scalars. The matched estimator, bootstrap-median Bures ratios, explicit scoping of secondary findings, and public release of per-cell tables are real methodological strengths and make the empirical grid citable. The latent-layer RF fingerprint is an interesting, falsifiable empirical pattern even if its causal attribution remains incomplete. Downstream uses (recipe auditing, training-time signals) are correctly left open. The contribution is primarily an instrument plus carefully scoped readings rather than a mechanistic theory of flow matching or distillation; that is still valuable for cs.CV if the main separation is robust under the stated design.

major comments (4)
  1. §3.1 and §4.3: The main RF detector is the sustained-band rule (R_lo>1 on at least three of {0.05,0.1,0.2,0.3}, including one s≥0.2). This rule cleanly separates the four RF cases from all ε-prediction models in the released tables, but it is not derived from flow-matching geometry and appears tuned to the observed RF shape (vs isolated low-s excursions such as PixArt-LCM). Because the central claim is defined by this criterion, the paper needs either (i) a short sensitivity analysis over nearby band definitions / thresholds, or (ii) explicit language that the detector is an empirical fingerprint criterion chosen for this sweep, not a pre-specified or theory-derived test. Without that, the near-binary claim is harder to evaluate outside the current grid.
  2. §4.3, Table 3, Table 8, and §5.2: Attribution of the latent band to prediction type (rather than residual architecture/conditioning/training confounds) rests on a single ε-prediction DiT control (PixArt-α) that holds T5 fixed. RF cases are MMDiT / hybrid-DiT with different conditioning stacks, resolutions, VAEs, and training recipes; the 2×2 in Table 8 has no RF+CLIP-only cell and only one non-UNet ε cell. The PixArt control rules out wide-T5 alone, which is useful, but does not fully isolate prediction type. The main claim should either soften from “prediction type” to “RF backbone family as instantiated in this sweep” or add a dedicated limitations paragraph that lists the remaining confounds as first-order, not secondary.
  3. §4.3 and §5.1: Survival of the latent fingerprint is shown only for pure-ADD RF students (SD3.5-Turbo, FLUX-schnell); there is no non-ADD RF student. The body is careful (“survives ADD”), but the abstract and introduction still frame a more general “survives distillation” reading. Align abstract/intro with the precise §5.1 statement, and treat “no non-ADD RF student” as a load-bearing scope limit on the main result rather than a minor coverage gap.
  4. §3.1 Eq. (1) and Appendix A: The headline statistic is amp_mean / max(amp_var, amp_scale) under a Gaussian Bures W2^2 on Inception pool3. The bootstrap-median fix for plug-in bias is well motivated, but the paper does not show that the mean-vs-higher-moment ratio (as opposed to raw amplitudes, other feature spaces, or non-Gaussian OT) is the quantity that isolates prediction type. A short ablation—e.g., reporting whether the RF band survives under DINOv2 features, under amp_mean alone, or under a diagonal-covariance control already mentioned as biased—would make the instrument choice less free-parameter-like for the central claim.
minor comments (5)
  1. Figure 2 is information-dense; annotating the sustained-band RF rows and the s=0.5 ADD column more explicitly (or splitting latent vs score panels) would help readers verify the three patterns without the appendix tables.
  2. Several reported intervals collapse to a single two-decimal value (e.g., 0.66[0.66,0.66]). The footnote explains rounding, but stating n_resample and effective n per cell once in the main text (not only Appendix) would reduce the appearance of zero-width CIs.
  3. Table 1 / paradigm labels: progressive-adv vs ADD vs mixed are operationally clear in §2.1, but a one-line mapping from vendor checkpoint names to loss-family labels in the table caption would reduce cross-referencing.
  4. §4.6 prompt-collapse observation is correctly marked as non-law; consider moving it fully to appendix or discussion so it does not compete with the three structured findings.
  5. Notation: R_sel, R_point, R_lo, R_hi are introduced cleanly; keep the equality line R=1, the band criterion, and the R_lo≥2 visual marker visually distinct in all figures (the text already warns against conflating them).

Circularity Check

1 steps flagged

Mild post-hoc criterion design for the RF band detector; measurements themselves are independent of the distillation losses or fitted generative parameters.

specific steps
  1. other [§3.1 (Reading a cell) and §4.3 (Main Finding)]
    "the rectified-flow latent detector (Section 4.3) is a sustained-band criterion: a model/configuration passes when its latent-layer ratios have Rlo >1 on at least three of the four low-to-mid strengths s∈{0.05,0.1,0.2,0.3}, including at least one s≥0.2 ; this is what separates a genuine band from an isolated single-cell low-s excursion. … Within this sweep the latent layer is thus a near-binary detector of the prediction type (read as a sustained low-to-mid band)"

    The detector criterion is defined precisely so that the observed RF rows (SD3.5/FLUX teachers + ADD students) form a sustained band while ε-pred rows (including PixArt-α and the low-s PixArt-LCM excursion) do not. The “near-binary detector” claim therefore holds by the post-hoc choice of band shape rather than by an independent, pre-specified geometric prediction from flow-matching; the underlying W2^{2} ratios themselves are not forced by construction.

full rationale

The paper is an empirical instrument paper: a training-free probe injects fixed perturbations, computes bootstrap-median Bures W2^{2} selectivity ratios under one matched estimator, and reports patterns across a 23-model public-checkpoint sweep. No parameters are fitted to the target quantities and then re-presented as predictions; no uniqueness theorems or load-bearing results are imported via self-citation (author is listed as independent; references contain no prior works by the same author that force the claims). The three-layer “orthogonal factors” reading and the main RF claim are summaries of the observed heatmap and tables, not algebraic identities. The only mild circularity is that the sustained-band rule used to declare a “near-binary detector” (R_lo > 1 on ≥3 of four low-to-mid strengths including one s ≥ 0.2) is chosen after inspecting the same data so that the four RF cases pass and the ε-pred cases (including the PixArt-α control and the isolated PixArt-LCM excursion) fail; this is ordinary post-hoc fingerprint definition rather than a derivation that reduces by construction. Score 2 reflects that single non-load-bearing design choice; the ratios and CIs remain externally checkable from the released tables.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central empirical claims rest on a constructed selectivity statistic, a hand-chosen strength grid and band rule, and standard but strong assumptions that Inception-feature Gaussian Bures distances track the mechanistic factors of interest. No new physical entities are postulated; the invented objects are the probe and its decision criteria. Free parameters are operational thresholds and estimator settings, not physics constants.

free parameters (4)
  • perturbation strength grid s = {0.05,0.1,0.2,0.3,0.5,1.0}
    Six discrete strengths {0.05,0.1,0.2,0.3,0.5,1.0} define the entire response surface and which cells enter the RF band test.
  • sustained-band RF detector rule = ≥3 of {0.05,0.1,0.2,0.3} with R_lo>1, including s≥0.2
    Requires R_lo>1 on ≥3 of four low-to-mid strengths including one s≥0.2; this rule is the operational definition of the main claim and is not derived from theory.
  • bootstrap n_resample and CI level = 200 resamples; 90% percentile interval
    n_resample=200 and 5/95 percentiles define every reported interval and CI-disjoint contrast.
  • strong-selectivity visual marker R_lo≥2.0 = 2.0
    Used as a display threshold (not the detector); still a hand-chosen visual cutoff in the analysis narrative.
axioms (5)
  • domain assumption Inception-v3 pool3 features of generated images are adequately modeled as Gaussians so that closed-form Bures W2² measures the relevant output-distribution change.
    Section 3.1 and 2.5 adopt full Bures on d=2048 Inception features as the headline distance without independent validation that this geometry tracks prediction-type or distillation-objective mechanisms.
  • ad hoc to paper Mean vs max(variance,scale) amplitude ratio is a meaningful selectivity summary of conditioning response.
    Equation (1) defines R_sel; the paper treats large R as mean-dominant selectivity without a prior theoretical necessity for this particular ratio form.
  • domain assumption PixArt-alpha holds wide T5 conditioning fixed while flipping only prediction type relative to RF models sufficiently to isolate prediction type.
    Section 4.3 uses PixArt as the control; residual differences (pure DiT vs MMDiT/hybrid, training recipe, resolution/step defaults) are assumed not to drive the latent band.
  • standard math Bootstrap-median of the nonlinear Bures ratio is the correct point estimator for cross-model comparison at n≈1500–2000.
    Appendix A motivates bootstrap-median over plug-in because Bures is nonlinear in covariance; this is a standard statistical fix under the paper's cell sizes.
  • domain assumption Public teacher/student checkpoints and fixed samplers are representative enough that within-sweep regularities can be stated as factor readings (prediction type, distillation objective).
    Sections 4.1 and 5.1 scope claims to the sweep but still interpret layers as reading those factors rather than checkpoint idiosyncrasies.
invented entities (3)
  • decomposable layer/mode selectivity probe no independent evidence
    purpose: Replace end-to-end FID/CLIP with layer-resolved response measurements via forward hooks.
    Core instrument of the paper; defined operationally in Section 3, not previously standardized in the cited literature.
  • selectivity ratio R_sel = amp_mean / max(amp_var, amp_scale) no independent evidence
    purpose: Scalar summary of whether mean perturbations dominate higher-moment ones on a layer.
    Equation (1); the main findings are statements about this constructed ratio's pattern across models.
  • sustained low-to-mid latent band criterion for rectified-flow detection no independent evidence
    purpose: Operational near-binary classifier separating RF from epsilon-prediction models.
    Section 3.1 defines the band rule used as the main detector; it is a paper-specific decision procedure.

pith-pipeline@v1.1.0-grok45 · 27998 in / 3964 out tokens · 58858 ms · 2026-07-12T03:45:37.010347+00:00 · methodology

0 comments
read the original abstract

Few-step distilled diffusion students cut text-to-image inference from ~50 to 1-8 network evaluations, but the quality gap is usually summarised by a single FID/CLIP scalar that cannot say which axis of the conditioning response changed, nor whether a behaviour comes from the architecture, the distillation objective, or simply from being a diffusion model. We replace the scalar with a decomposable probe that injects controlled perturbations along three layers (prompt encoder, denoiser input, denoiser output) under three modes (mean, variance, scale) and six strengths, reporting a bootstrap-median Bures W2^2 selectivity ratio on Inception features. Under a single matched estimator across 23 models -- five teachers and 18 distilled students spanning five backbone families (SDXL, SD1.5, SD3.5, PixArt-alpha, FLUX), three architecture classes (UNet, DiT, MMDiT), and five distillation paradigms -- the three layers read three empirically separable factors: the prompt layer is a universal prompt-mean response (a sanity channel, not a discriminator), the latent layer reads the prediction type, and the score layer reads the distillation objective. Our main result: within this sweep, the latent layer is a near-binary detector of rectified-flow backbones. Its ratio exceeds 1 across a sustained low-to-mid band only for rectified-flow models (SD3.5, FLUX); no epsilon-prediction model qualifies. A matched epsilon-prediction control (PixArt-alpha) rules out wide-T5 conditioning, and the fingerprint survives adversarial (ADD) distillation as both teacher and student. Two secondary score-layer findings hold under narrower scopes: a canonical 4-step ADD-vs-rest contrast on the UNet families with a non-ADD baseline, and a CI-separated trajectory-rollout early-strength score spike on both UNet and DiT. All ratios are CI-citable under one estimator; we release the per-cell tables and the estimator.

Figures

Figures reproduced from arXiv: 2607.03256 by Patrick Mu Haojie.

Figure 1
Figure 1. Figure 1: The probe and its main reading. (A) Where and how the probe injects. The prompt embedding is perturbed once per generation at the text encoder; the latent and score tensors are perturbed at every one of the N denoise steps. Each layer is hit under three modes — mean (ht+δ µˆ ), variance (ht+δσεˆ ), and scale (ht(1+δ)) — across six strengths. The decoded image is embedded by Inception-v3 and each cell is su… view at source ↗
Figure 2
Figure 2. Figure 2: Master heatmap of the 23-model sweep, all three layers. Rows are model/configuration rows (teachers, then UNet students, then non-UNet students; FLUX.1-dev appears at both 28 and 50 steps); columns are the three layers × six strengths. Cell colour is log10 R on a diverging scale centred at R=1 (red > 1, blue < 1). Three patterns are visible at once: the latent block turns red only on the rectified-flow row… view at source ↗
Figure 3
Figure 3. Figure 3: Trajectory-rollout students spike on the score layer at low strength. Score-layer R vs strength s (log x). Solid coloured curves are trajectory-rollout students (LCM on UNet and DiT, Flash on SD1.5); their score ratio rises above R=1 at s≤0.1 and then falls. Grey dashed curves are the ADD / mixed / distribution-matching students, none of which have a CI-separated low-s peak above 1. PixArt-LCM (DiT) shows … 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

91 extracted references · 40 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Improved Distribution Matching Distillation for Fast Image Synthesis , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  2. [2]

    International Conference on Machine Learning (ICML) , year =

    Score Identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation , author =. International Conference on Machine Learning (ICML) , year =

  3. [3]

    International Conference on Machine Learning (ICML) , year =

    Input Perturbation Reduces Exposure Bias in Diffusion Models , author =. International Conference on Machine Learning (ICML) , year =

  4. [4]

    Elucidating the

    Yu, Meng and Sun, Lei and Zeng, Jianhao and Chu, Xiangxiang and Zhan, Kun , booktitle =. Elucidating the. 2026 , note =

    2026

  5. [5]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    One-step Diffusion with Distribution Matching Distillation , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  6. [6]

    International Conference on Machine Learning (ICML) , year =

    Consistency Models , author =. International Conference on Machine Learning (ICML) , year =

  7. [10]

    Ren, Yuxi and Xia, Xin and Lu, Yanzuo and Zhang, Jiacheng and Wu, Jie and Xie, Pan and Wang, Xing and Xiao, Xuefeng , journal =. Hyper-

  8. [11]

    European Conference on Computer Vision (ECCV) , year =

    Adversarial Diffusion Distillation , author =. European Conference on Computer Vision (ECCV) , year =

  9. [12]

    Lin, Shanchuan and Wang, Anran and Yang, Xiao , journal =

  10. [14]

    arXiv preprint arXiv:2406.01561 , year =

    Long and Short Guidance in Score Identity Distillation for One-Step Text-to-Image Generation , author =. arXiv preprint arXiv:2406.01561 , year =

    Pith/arXiv arXiv

  11. [15]

    arXiv preprint arXiv:2505.12674 , year =

    Few-Step Diffusion via Score Identity Distillation , author =. arXiv preprint arXiv:2505.12674 , year =

    Pith/arXiv arXiv

  12. [16]

    AAAI Conference on Artificial Intelligence (AAAI) , year =

    Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation , author =. AAAI Conference on Artificial Intelligence (AAAI) , year =

  13. [17]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Scalable Diffusion Models with Transformers , author =. IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  14. [18]

    International Conference on Learning Representations (ICLR) , year =

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author =. International Conference on Learning Representations (ICLR) , year =

  15. [19]

    International Conference on Learning Representations (ICLR) , year =

    Flow Matching for Generative Modeling , author =. International Conference on Learning Representations (ICLR) , year =

  16. [20]

    International Conference on Machine Learning (ICML) , year =

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author =. International Conference on Machine Learning (ICML) , year =

  17. [21]

    2024 , note =

    Chen, Junsong and Yu, Jincheng and Ge, Chongjian and Yao, Lewei and Xie, Enze and Wu, Yue and Wang, Zhongdao and Kwok, James and Luo, Ping and Lu, Huchuan and Li, Zhenguo , booktitle =. 2024 , note =

    2024

  18. [22]

    2024 , howpublished =

    2024

  19. [23]

    Chen, Junsong and Wu, Yue and Luo, Simian and Xie, Enze and Paul, Sayak and Luo, Ping and Zhao, Hang and Li, Zhenguo , journal =

  20. [24]

    2024 , howpublished =

    Stable Diffusion 3.5 Large Turbo (model card) , author =. 2024 , howpublished =

    2024

  21. [25]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    High-Resolution Image Synthesis with Latent Diffusion Models , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  22. [26]

    International Conference on Learning Representations (ICLR) , year =

    Podell, Dustin and English, Zion and Lacey, Kyle and Blattmann, Andreas and Dockhorn, Tim and M. International Conference on Learning Representations (ICLR) , year =

  23. [27]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Rethinking the Inception Architecture for Computer Vision , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  24. [28]

    Microsoft

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll. Microsoft. European Conference on Computer Vision (ECCV) , year =

  25. [29]

    International Conference on Machine Learning (ICML) , year =

    Learning Transferable Visual Models from Natural Language Supervision , author =. International Conference on Machine Learning (ICML) , year =

  26. [30]

    Journal of Machine Learning Research (JMLR) , year =

    Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer , author =. Journal of Machine Learning Research (JMLR) , year =

  27. [31]

    Transactions on Machine Learning Research (TMLR) , year =

    Oquab, Maxime and Darcet, Timoth. Transactions on Machine Learning Research (TMLR) , year =

  28. [32]

    Expositiones Mathematicae , volume =

    On the Bures--Wasserstein Distance between Positive Definite Matrices , author =. Expositiones Mathematicae , volume =. 2019 , note =

    2019

  29. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  30. [34]

    International Conference on Learning Representations (ICLR) , year =

    Sequence Level Training with Recurrent Neural Networks , author =. International Conference on Learning Representations (ICLR) , year =

  31. [35]

    International Conference on Learning Representations (ICLR) , year =

    Elucidating the Exposure Bias in Diffusion Models , author =. International Conference on Learning Representations (ICLR) , year =

  32. [36]

    Qin, You and Wang, Linqing and Fei, Hao and Zimmermann, Roger and Bo, Liefeng and Lu, Qinglin and Wang, Chunyu , journal =

  33. [37]

    2017 , note =

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , booktitle =. 2017 , note =

    2017

  34. [38]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Improved Precision and Recall Metric for Assessing Generative Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  35. [39]

    Transactions on Machine Learning Research , year =

    The Vendi Score: A Diversity Evaluation Metric for Machine Learning , author =. Transactions on Machine Learning Research , year =

  36. [40]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  37. [41]

    European Conference on Computer Vision (ECCV) , year =

    Self-Rectifying Diffusion Sampling with Perturbed-Attention Guidance , author =. European Conference on Computer Vision (ECCV) , year =

  38. [42]

    2024 , note =

    Si, Chenyang and Huang, Ziqi and Jiang, Yuming and Liu, Ziwei , booktitle =. 2024 , note =

    2024

  39. [43]

    International Conference on Machine Learning (ICML) , year =

    Align Your Steps: Optimizing Sampling Schedules in Diffusion Models , author =. International Conference on Machine Learning (ICML) , year =

  40. [44]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Guiding a Diffusion Model with a Bad Version of Itself , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  41. [45]

    Transactions on Machine Learning Research (TMLR) , year =

    Classifier-Free Guidance is a Predictor-Corrector , author =. Transactions on Machine Learning Research (TMLR) , year =

  42. [46]

    NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year =

    Classifier-Free Diffusion Guidance , author =. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year =

    2021

  43. [47]

    International Conference on Learning Representations (ICLR) , year =

    Progressive Distillation for Fast Sampling of Diffusion Models , author =. International Conference on Learning Representations (ICLR) , year =

  44. [48]

    International Conference on Learning Representations (ICLR) , year =

    Denoising Diffusion Implicit Models , author =. International Conference on Learning Representations (ICLR) , year =

  45. [49]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Denoising Diffusion Probabilistic Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  46. [50]

    Effectively Unbiased

    Chong, Min Jin and Forsyth, David , booktitle =. Effectively Unbiased. 2020 , note =

    2020

  47. [51]

    and Steinhardt, Jacob , booktitle =

    Gandelsman, Yossi and Efros, Alexei A. and Steinhardt, Jacob , booktitle =. Interpreting. 2024 , note =

    2024

  48. [52]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  49. [53]

    Transactions on Machine Learning Research (TMLR) , year =

    Scaling Autoregressive Models for Content-Rich Text-to-Image Generation , author =. Transactions on Machine Learning Research (TMLR) , year =

  50. [54]

    Self-rectifying diffusion sampling with perturbed-attention guidance

    Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. In European Conference on Computer Vision (ECCV), 2024. arXiv:2403.17377

    Pith/arXiv arXiv 2024

  51. [55]

    On the bures--wasserstein distance between positive definite matrices

    Rajendra Bhatia, Tanvi Jain, and Yongdo Lim. On the bures--wasserstein distance between positive definite matrices. Expositiones Mathematicae, 37 0 (2): 0 165--191, 2019. arXiv:1712.01504

    Pith/arXiv arXiv 2019

  52. [56]

    Black Forest Labs . FLUX.1 . https://github.com/black-forest-labs/flux, 2024 a . Rectified-flow text-to-image model; FLUX.1-dev and the distilled FLUX.1-schnell

    2024

  53. [57]

    FLUX.1 -schnell (model card)

    Black Forest Labs . FLUX.1 -schnell (model card). https://huggingface.co/black-forest-labs/FLUX.1-schnell, 2024 b . Rectified-flow transformer distilled with latent adversarial diffusion distillation; 1--4 step sampling

    2024

  54. [58]

    Classifier-free guidance is a predictor-corrector

    Arwen Bradley and Preetum Nakkiran. Classifier-free guidance is a predictor-corrector. Transactions on Machine Learning Research (TMLR), 2025. arXiv:2408.09000; NeurIPS 2024 M3L Workshop

    Pith/arXiv arXiv 2025

  55. [59]

    Flash diffusion: Accelerating any conditional diffusion model for few steps image generation

    Cl \'e ment Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. In AAAI Conference on Artificial Intelligence (AAAI), 2025. arXiv:2406.02347

    Pith/arXiv arXiv 2025

  56. [60]

    PixArt- : Fast and controllable image generation with latent consistency models

    Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. PixArt- : Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024 a

    Pith/arXiv arXiv 2024

  57. [61]

    PixArt- : Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In International Conference on Learning Representations (ICLR), 2024 b . arXiv:2310.00426

    Pith/arXiv arXiv 2024

  58. [62]

    Effectively unbiased FID and inception score and where to find them

    Min Jin Chong and David Forsyth. Effectively unbiased FID and inception score and where to find them. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6070--6079, 2020. arXiv:1911.07023

    Pith/arXiv arXiv 2020

  59. [63]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems (NeurIPS), 2013

    2013

  60. [64]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M \"u ller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on M...

    Pith/arXiv arXiv 2024

  61. [65]

    The vendi score: A diversity evaluation metric for machine learning

    Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. Transactions on Machine Learning Research, 2023. arXiv:2210.02410

    Pith/arXiv arXiv 2023

  62. [66]

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

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. arXiv:1706.08500

    Pith/arXiv arXiv 2017

  63. [67]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. arXiv:2207.12598

    Pith/arXiv arXiv 2021

  64. [68]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), 2020. arXiv:2006.11239

    Pith/arXiv arXiv 2020

  65. [69]

    Guiding a diffusion model with a bad version of itself

    Tero Karras, Miika Aittala, Tuomas Kynk \"a \"a nniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. In Advances in Neural Information Processing Systems (NeurIPS), 2024. arXiv:2406.02507

    Pith/arXiv arXiv 2024

  66. [70]

    Imagine flash: Accelerating emu diffusion models with backward distillation

    Jonas Kohler, Albert Pumarola, Edgar Sch \"o nfeld, Artsiom Sanakoyeu, Roshan Sumbaly, Peter Vajda, and Ali Thabet. Imagine flash: Accelerating emu diffusion models with backward distillation. arXiv preprint arXiv:2405.05224, 2024

    Pith/arXiv arXiv 2024

  67. [71]

    SDXL-Lightning : Progressive adversarial diffusion distillation

    Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL-Lightning : Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024

    Pith/arXiv arXiv 2024

  68. [72]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C. Lawrence Zitnick. Microsoft COCO : Common objects in context. In European Conference on Computer Vision (ECCV), 2014. arXiv:1405.0312

    Pith/arXiv arXiv 2014

  69. [73]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In International Conference on Learning Representations (ICLR), 2023. arXiv:2210.02747

    Pith/arXiv arXiv 2023

  70. [74]

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

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In International Conference on Learning Representations (ICLR), 2023. arXiv:2209.03003

    Pith/arXiv arXiv 2023

  71. [75]

    Latent consistency models: Synthesizing high-resolution images with few-step inference

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023 a

    Pith/arXiv arXiv 2023

  72. [76]

    LCM-LoRA : A universal stable-diffusion acceleration module

    Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolin \'a rio Passos, Longbo Huang, Jian Li, and Hang Zhao. LCM-LoRA : A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023 b

    Pith/arXiv arXiv 2023

  73. [77]

    DINOv2 : Learning robust visual features without supervision

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2 : Learning robust visual features without supervision. Transactions on Machine Learning Research (TMLR), 2024. arXiv:2304.07193

    Pith/arXiv arXiv 2024

  74. [78]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), 2023. arXiv:2212.09748

    Pith/arXiv arXiv 2023

  75. [79]

    SDXL : Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M \"u ller, Joe Penna, and Robin Rombach. SDXL : Improving latent diffusion models for high-resolution image synthesis. In International Conference on Learning Representations (ICLR), 2024. arXiv:2307.01952

    Pith/arXiv arXiv 2024

  76. [80]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. arXiv:2103.00020

    Pith/arXiv arXiv 2021

  77. [81]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 2020. arXiv:1910.10683

    Pith/arXiv arXiv 2020

  78. [82]

    Hyper- SD : Trajectory segmented consistency model for efficient image synthesis

    Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper- SD : Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv:2404.13686, 2024

    Pith/arXiv arXiv 2024

  79. [83]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. arXiv:2112.10752

    Pith/arXiv arXiv 2022

  80. [84]

    Align your steps: Optimizing sampling schedules in diffusion models

    Amirmojtaba Sabour, Sanja Fidler, and Karsten Kreis. Align your steps: Optimizing sampling schedules in diffusion models. In International Conference on Machine Learning (ICML), 2024. arXiv:2404.14507

    Pith/arXiv arXiv 2024

Showing first 80 references.