Pith. sign in

REVIEW 5 major objections 4 minor 37 references

EMAG: Self-Rectifying Diffusion Sampling with Exponential Moving Average Guidance

T0 review · 5 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Replacing attention maps with their moving average during diffusion sampling produces harder negatives and lifts human preference over classifier-free guidance.

desk verdict EMAG is a plausible new twist on attention-perturbation guidance—EMA as weak surrogate plus adaptive layer selection—but the headline HPS gain is a tuned 1k-sample point estimate, and the CADS composition claim overstates their own Table 3. read the letter →

arxiv 2512.17303 v2 pith:NPIL373J submitted 2025-12-19 cs.CV

Ankit Yadav , Ta Duc Huy , Lingqiao Liu This is my paper
classification cs.CV
keywords diffusionmodelsclassifier-freeguidanceexponentialmovingaveragehardnegativesattentionhumanpreferencescoretraining-freeflowmatching
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

EMAG is a training-free guidance method for diffusion and flow-matching transformers. At inference, it replaces the self-attention map at a selected layer with that layer's exponential moving average over recent timesteps, producing a slightly degraded 'hard negative' prediction that preserves global structure but drops fine-grained refinements. Contrasting this negative with the original prediction in an auto-guidance update sharpens the sample and improves the Human Preference Score (HPS) by +0.46 over classifier-free guidance, while remaining competitive on FID. A simple per-timestep rule selects the layer whose attention map deviates most from its EMA, and the same mechanism composes with APG and CADS. The paper argues that the hardness and semantic faithfulness of the negative — not the specific perturbation — is what drives the gain.

What carries the argument

The exponential moving average of softmax attention maps, E_t = βE_{t-1} + (1−β)A_t with β = e^{−ln2/H} (default H=50, β=0.988), maintained per layer during reverse sampling. At each timestep, the layer in the middle range (12–15 for DiT-XL/2; 6–8 for SD3) with the largest mean absolute error Δ_t = MAE(E_t, A_t) between the EMA and the current map is selected, and its attention map is replaced by the EMA (hard replacement λ=1, soft blend λ∈[0,1] optional). This perturbed attention yields a weakened score that is contrasted with the original in the auto-guidance update, guiding sampling toward high-quality regions. The Hopfield-network analogy (attention as iterative energy descent) provides

What would settle it

Check whether EMA-replaced attention maps behave as low-pass, semantically faithful degradations: measure attention entropy, spatial-support overlap with the current map, and CLIP similarity of the resulting negative samples to the unperturbed prediction across timesteps. If the negatives are either too easy for the denoiser to repair or semantically drift (e.g., attention ranking flips), the asserted mechanism is not what causes the HPS improvement.

Watch

Extended reading notes

Core claim

The central claim is that controllable hard negatives — subtle, semantically faithful degradations — are the key to effective training-free guidance in diffusion transformers, and that an exponential moving average of the attention map is a direct way to produce them. For each timestep t and selected layer, EMAG replaces the softmax attention map A_t with E_t = βE_{t-1} + (1−β)A_t (β = 0.988, half-life 50), and then applies the auto-guidance update that contrasts the original score with the EMA-perturbed score. Because self-attention in diffusion transformers refines structure from coarse to fine, the EMA lags behind the sharpest recent refinement, nudging the sample slightly off the current

Load-bearing premise

The load-bearing premise is that replacing an attention map with its exponential moving average creates a semantically faithful low-pass degradation that suppresses high-frequency refinements while preserving global structure, rather than a stale or misaligned attention that simply corrupts the negative sample.

Editorial extensions

If this is right

  • Any diffusion or flow-matching transformer can use EMAG at inference time without retraining or fine-tuning, simply by maintaining per-layer attention EMAs and a Δ-based layer-selection rule.
  • The adaptive layer-selection rule depends only on attention statistics, so it should transfer to other transformer backbones and resolutions beyond DiT-XL/2 and SD3-Medium.
  • EMAG composes with complementary guidance methods (CFG, APG, CADS), indicating that multiple failure modes — oversaturation, low diversity, and subtle artifacts — can be addressed simultaneously during sampling.
  • The hard-negative design principle suggests a new axis for guidance research: methods should produce degradations that are hard enough to expose failure modes but faithful enough that the model can correct them, rather than arbitrary corruption.
  • In high-resolution, high-quality regimes, EMAG's gains are more pronounced (e.g., 512×512 DiT and SD3), suggesting that subtle attention-level refinements matter most when the base model already produces strong samples.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: varying the EMA half-life H should monotonically control the difficulty of negatives; finding the optimal H per model and resolution could further improve the HPS-FID trade-off beyond the fixed H=50 used here.
  • Because EMAG only touches attention space, it could generalize to other attention-heavy generative models (e.g., video diffusers or autoregressive image models), where similar coarse-to-fine refinement dynamics are expected.
  • The reported +0.46 HPS gain may partly reflect HPS's sensitivity to sharpness or texture; a paired human-preference study with forced choice would confirm whether the refinements are genuinely preferred, not just score-preferred.
  • The greedy argmax layer-selection rule could be replaced by a differentiable or look-ahead selection that optimizes Δ_t over layers; the paper itself flags this as future work, and it could yield stronger negatives.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes EMAG, a training-free inference-time guidance method for diffusion and flow-matching transformers. EMAG maintains an exponential moving average of self-attention maps during sampling and, at selected layers and timesteps, replaces the current attention map with this EMA to create a degraded “negative” score. The negative is then contrasted with the original score in an auto-guidance/CFG-style update. An adaptive per-timestep layer selection rule chooses the layer with maximal MAE between current and EMA attention. The paper reports an HPS gain of +0.46 over CFG on SD3 text-to-image generation (Table 3), further improvements when combined with APG/CADS/S2, and FID improvements in unconditional DiT/SD3 settings. The mechanism is motivated by a Hopfield-network analogy, which the paper itself describes as an interpretive analogy rather than a proof.

Significance. If the reported gains are reproducible, EMAG is a simple, training-free plug-in with clear practical value, and the composition experiments with APG/CADS are a useful contribution. The paper's strengths include broad baseline comparisons, detailed hyperparameter tables (Table S11), and explicit ablations of the layer-selection mechanism. However, the central quantitative claim is not yet established: HPS is computed on a single 1,000-image subsample with a fixed seed and no confidence intervals, and the headline hyperparameters are tuned on the same benchmark. The semantic-faithfulness mechanism is supported only by qualitative figures and an analogy, not by direct measurement. Additionally, the adaptive layer-selection ablation does not, as reported, show a benefit over a fixed layer. These gaps require additional experiments and clarification before the paper's claims can be fully credited.

major comments (5)
  1. [§3.4.2, Eq. (12) vs. Supplementary Sec. A] The core perturbation is specified inconsistently. The main text states 'we replace ... using a convex blend, see Eq.12 with λ=1 (hard replace) as our default'; the supplementary Hopfield section states 'we keep λ=0' for the same equation. Since Eq. (12) with λ=0 becomes eA_t = A_t, the perturbation is the identity and no EMAG guidance would be applied. All results depend on which value is actually used. Please reconcile these statements, state the exact implementation, and confirm that the reported numbers are produced with the stated default.
  2. [§4.1 / Supplementary Sec. C (HPS evaluation)] The headline +0.46 HPS gain in Table 3 is computed on a random 1,000-image subsample with a fixed seed and no confidence intervals, and the same COCO/HPS benchmark is used for hyperparameter selection (Tables S2–S5, Fig. S3–S4). Guidance scales, EMA decay, temporal windows, and layer ranges are all tuned on this benchmark. The reported gain may therefore be a selection artifact. Please report mean±std over multiple subsamples or a full-set evaluation, and use a held-out prompt set or a validation split for all EMAG hyperparameters.
  3. [§4.3, Table 4] The ablation does not support the adaptive layer-selection contribution. The adaptive row reports FID 28.52 / HPS 29.60, while the fixed-layer L8 row reports FID 28.23 / HPS 29.66 — better on both metrics. The text claims single-layer variants 'improve FID at the expense of HPS,' but L8 dominates adaptive selection. Either the adaptive criterion must be shown to provide a benefit over the best fixed layer, or the contribution should be reframed as a heuristic whose value is not yet demonstrated.
  4. [§3.4.1, Eq. (10)–(11)] The claim that EMA attention produces 'semantically faithful hard negatives' is not directly tested. The adaptive selector chooses the layer with maximal MAE(E_t,A_t), which may equally select layers undergoing global layout reorganization rather than subtle high-frequency refinement. To make the mechanism load-bearing, add a control that perturbs attention with stale-but-matched alternatives, e.g., attention from an earlier fixed step or temporally shuffled A_t with the same L1 deviation. If such controls reproduce the HPS/FID gains, the EMA-specific semantic-faithfulness explanation is unsupported.
  5. [Tables S8/S9] On class-conditional DiT, EMAG's HPS improvements are reported at substantially worse FID than CFG (Table S8: FID 4.16 vs 2.30; Table S9: FID 7.59 vs 3.08). Since stronger guidance scales can trade FID for preference, the claim that EMAG 'improves quality' should be presented as a Pareto analysis or a matched-FID comparison. Without this, the DiT HPS gains may simply reflect a different operating point rather than a genuine improvement.
minor comments (4)
  1. [§3.4.1, Eq. (9)] Eq. (9) updates E_t using X_t, but the preceding text defines the attention map as A_t. Please use A_t consistently and correct 'same shape as X_t' to 'same shape as A_t'.
  2. [Algorithms 1–2] The notation for the updated conditional score is inconsistent: Algorithm 2 defines ȳ_t^c, while Eq. (16) uses ỹ(xt,c)′. Please unify the notation so the two-step update is unambiguous.
  3. [Tables S8/S9] The method is called 'I-ERG' in these tables but 'ERG' elsewhere in the paper. Please standardize the name.
  4. [References] Reference [17] contains a stray '1' at the end of the entry; please clean up the bibliography formatting.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial circularity: the headline HPS gain is a benchmark-tuned operating point, and the supplementary formalization of the EMA replacement is inconsistent (λ=0 makes it identity); the core EMAG equations are not themselves defined in terms of HPS.

  1. fitted input called prediction [Supplementary Sec. B.3 (hyperparameter selection), Sec. 4.1, Fig. 5, Table S11]
    "We further analyze the behaviour of EMAG with respect to PRDC metrics on SD3-Medium in Figures S3 and S4, which we use to select a balanced combination of guidance and EMAG scales ... both studies are evaluated on 1,000 generated samples. For HPS, we randomly subsample 1,000 images (without replacement, using a fixed random seed8) from the pool of generated samples for each method and report the mean HPS over these images."

    The EMAG operating point (w_e = 1.75 for SD3 conditional, plus temporal windows and layer ranges in Table S11) is selected by HPS/FID ablations on 1,000 samples from the same COCO-2014 validation split and the same evaluation setup used for the headline HPS. The reported '+0.46 over CFG' is therefore the HPS at a hyperparameter setting chosen to optimize that same HPS measure, not an out-of-the-box prediction of the EMA hard-negative mechanism. The empirical claim is partly a model-selection artifact on the benchmark that produces the headline number.

  2. other [Supplementary Sec. A, Eq. 12 (contrast with Sec. 3.4.2)]
    "EMAG replaces them with an exponential moving average over past steps, which we denote schematically as eAt = (1−λ)A_t + λ E_t (eq:12) where we keep λ=0. Because eAt is a temporally smoothed version of A_t..."

    With λ=0, Eq. 12 reduces to the identity map eAt = A_t, so the supplementary derivation of EMAG's 'temporally smoothed, low-pass hard negative' describes no perturbation at all. This contradicts the default λ=1 hard replacement stated in Sec. 3.4.2. The justification section's core operation is either empty (identity) or inconsistent with the evaluated method, so the mechanism is not derived from the paper's own equations as written.

full rationale

EMAG's construction is not circular in the strongest sense: no equation defines the claimed HPS gain in terms of the EMA update, and the paper contains no load-bearing self-citation chain or imported uniqueness theorem. The Hopfield discussion is explicitly labeled an interpretive analogy, not a proof. However, the empirical centerpiece is weakened by two concrete issues. First, the EMAG hyperparameters (especially w_e and the temporal windows) are selected using HPS/FID ablations on the same COCO-2014 validation split and the same 1,000-sample HPS evaluation pool that yields the headline '+0.46 over CFG', so the gain is partly a tuned operating point rather than an independent prediction of the hard-negative mechanism. Second, Supplementary Sec. A defines the formal perturbation with λ=0, making Eq. 12 the identity and contradicting the main-text default λ=1; the low-pass/semantic-faithfulness explanation is therefore internally inconsistent in the very passage meant to justify it. These are partial circularity and specification issues, not a full reduction of the derivation to its inputs, hence score 4.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

Central claim rests on five tuned parameters (beta, w_e, windows, layer ranges, lambda) and on an unproven low-pass/hard-negative premise; no new entities are postulated.

free parameters (5)
  • EMA decay beta (half-life H=50) = 0.988 (beta = e^{-ln2/50})
    Hand-set smoothing constant controlling degradation strength; not fit to data but tuned as a hyperparameter.
  • EMAG guidance scale w_e = 1.75 (SD3 cond), 5.125 (SD3 uncond), 1.5 (DiT cond), 7 (DiT uncond)
    Selected via HPS/FID ablations on the same COCO/ImageNet benchmarks used for headline results.
  • EMA window tau_s, tau_e, delta_t = tau_s=26/250, tau_e=5/50, delta_t=5/50
    Start/end/warmup timesteps chosen by hand; affects when guidance is applied.
  • Layer ranges (l_min, l_max) = (6,8) for SD3; (12,15) for DiT
    Chosen following prior work's middle-block guidance [17]; not learned.
  • Blend factor lambda = 1 default (hard replace)
    Default hard replacement; lambda=0 in one supplementary sentence contradicts Sec. 3.4.2.
assumptions (4)
  • domain assumption Modern Hopfield attention dynamics converge monotonically and attention refines from coarse to fine across diffusion timesteps
    Used in Sec. 3.4.3 and Supp. A to motivate EMA as low-pass; paper itself calls it an interpretive analogy, not a theorem for time-dependent noisy attention.
  • domain assumption Replacing A_t with E_t suppresses high-frequency refinements while preserving semantics
    Core premise for 'hard negatives'; assumed, supported only by qualitative figures (Fig. 2-3), not derived.
  • domain assumption Middle layers (SD3 6-8, DiT 12-15) are the right search space for perturbation
    Taken from prior work [1,15,17], not re-derived for transformer backbones.
  • ad hoc to paper Selecting the layer with the largest MAE between E_t and A_t at each timestep is a valid proxy for maximal useful degradation
    Eq. 10-11; introduced by this paper without comparative justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EMAG: Self-Rectifying Diffusion Sampling with Exponential Moving Average Guidance." pith.science (2026). https://pith.science/paper/NPIL373J

@misc{pith2026251217303,
  author       = {Pith},
  title        = {Pith review of: EMAG: Self-Rectifying Diffusion Sampling with Exponential Moving Average Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPIL373J}},
  note         = {Machine review of arXiv:2512.17303}
}
read the original abstract

In diffusion and flow-matching generative models, guidance techniques are widely used to improve sample quality and consistency. Classifier-free guidance (CFG) is the de facto choice in modern systems and achieves this by contrasting conditional and unconditional samples. Recent work explores contrasting negative samples at inference using a weaker model, via strong/weak model pairs, attention-based masking, stochastic block dropping, or perturbations to the self-attention energy landscape. While these strategies refine the generation quality, they still lack reliable control over the granularity or difficulty of the negative samples, and target-layer selection is often fixed. We propose Exponential Moving Average Guidance (EMAG), a training-free mechanism that modifies attention at inference time in diffusion transformers, with a statistics-based, adaptive layer-selection rule. Unlike prior methods, EMAG produces harder, semantically faithful negatives (fine-grained degradations), surfacing difficult failure modes, enabling the denoiser to refine subtle artifacts, boosting the quality and human preference score (HPS) by +0.46 over CFG. We further demonstrate that EMAG naturally composes with advanced guidance techniques, such as APG and CADS, further improving HPS.

Figures

Figures reproduced from arXiv: 2512.17303 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Negative-sample comparison (top: negative samples; bottom: positive samples). Prior guidance (SAG [15], auto guid￾ance [18], ERG [17], S2 -Guidance [4]) often yield obvious “easy” degradations. In contrast, EMAG produces subtle, semantically near-miss negatives that reveal difficult failure modes yet retain global structure, enabling finer refinement. (Samples from DiT) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. CFG(left) vs EMAG hard negatives(right) EMAG samples at varying β; larger β produces stronger degradations. prediction scaled by s; equivalently, Eq. 6 is augmented with the term − s · ϵ ′ 1 (xt, c, m), where m controls which layers are dropped at each timestep, ϵ ′ 1 is the perturbed sub￾network, and s is the S2 scale. In the auto-guidance framework, a weaker model helps steer sampling toward higher-quality regions… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison for (a) Unconditional and (b) Text-conditional settings. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: HPS and FID vs guidance scale (1000 samples). Colors [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 12 linked inside Pith

  1. [1]

    Self-rectifying diffu- sion sampling with perturbed-attention guidance

    Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Ky- ong Hwan Jin, and Seungryong Kim. Self-rectifying diffu- sion sampling with perturbed-attention guidance. InEuro- pean Conference on Computer Vision, pages 1–17. Springer,

  2. [2]

    Consistency-diversity-realism pareto fronts of conditional image generative models.arXiv preprint arXiv:2406.10429, 2024

    Pietro Astolfi, Marlene Careil, Melissa Hall, Oscar Ma ˜nas, Matthew Muckley, Jakob Verbeek, Adriana Romero Sori- ano, and Michal Drozdzal. Consistency-diversity-realism pareto fronts of conditional image generative models.arXiv preprint arXiv:2406.10429, 2024. 2

    arXiv 2024

  3. [3]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023. 1

    arXiv 2023

  4. [4]

    1, 2, 4, 5, 7

    Chubin Chen, Jiashu Zhu, Xiaokun Feng, Nisha Huang, Meiqi Wu, Fangyuan Mao, Jiahong Wu, Xiangxiang Chu, and Xiu Li.s 2-guidance: Stochastic self guidance for training-free enhancement of diffusion models.arXiv preprint arXiv:2508.12880, 2025. 1, 2, 4, 5, 7

    arXiv 2025

  5. [5]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. 6

    2009

  6. [6]

    Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 1, 2, 3, 6, 7

    2021

  7. [7]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning,

  8. [8]

    Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020. 1

    2020

Show all 37 references
  1. [9]

    Evalgim: A library for evaluating generative image models,

    Melissa Hall, Oscar Ma ˜nas, Reyhane Askari, Mark Ibrahim, Candace Ross, Pietro Astolfi, Tariq Berrada Ifriqi, Marton Havasi, Yohann Benchetrit, Karen Ullrich, Carolina Braga, Abhishek Charnalia, Maeve Ryan, Mike Rabbat, Michal Drozdzal, Jakob Verbeek, and Adriana Romero Soria...

  2. [10]

    Clipscore: A reference-free evaluation met- ric for image captioning.arXiv preprint arXiv:2104.08718,

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met- ric for image captioning.arXiv preprint arXiv:2104.08718,

    arXiv

  3. [11]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017

    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 equilib- rium.Advances in neural information processing systems, 30, 2017. 7

    2017

  4. [12]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 1, 2, 3

    2022 arXiv

  5. [13]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1

    2020

  6. [14]

    Smoothed energy guidance: Guiding dif- fusion models with reduced energy curvature of attention

    Susung Hong. Smoothed energy guidance: Guiding dif- fusion models with reduced energy curvature of attention. Advances in Neural Information Processing Systems, 37: 66743–66772, 2024. 1, 2, 4, 7

    2024

  7. [15]

    Improving sample quality of diffusion models us- ing self-attention guidance

    Susung Hong, Gyuseong Lee, Wooseok Jang, and Seungry- ong Kim. Improving sample quality of diffusion models us- ing self-attention guidance. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462– 7471, 2023. 1, 2, 4, 5, 7

    2023

  8. [16]

    Neural networks and physical systems with emergent collective computational abilities.Proceedings of the national academy of sciences, 79(8):2554–2558, 1982

    John J Hopfield. Neural networks and physical systems with emergent collective computational abilities.Proceedings of the national academy of sciences, 79(8):2554–2558, 1982. 5

    1982

  9. [17]

    Entropy rectifying guidance for diffusion and flow models.arXiv preprint arXiv:2504.13987, 2025

    Tariq Berrada Ifriqi, Adriana Romero-Soriano, Michal Drozdzal, Jakob Verbeek, and Karteek Alahari. Entropy rectifying guidance for diffusion and flow models.arXiv preprint arXiv:2504.13987, 2025. 2, 4, 5, 7, 1

    2025

  10. [18]

    Guiding a diffusion model with a bad version of itself.Advances in Neural In- formation Processing Systems, 37:52996–53021, 2024

    Tero Karras, Miika Aittala, Tuomas Kynk ¨a¨anniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself.Advances in Neural In- formation Processing Systems, 37:52996–53021, 2024. 1, 2, 4

    2024

  11. [19]

    Understanding diffu- sion objectives as the elbo with simple data augmentation

    Diederik Kingma and Ruiqi Gao. Understanding diffu- sion objectives as the elbo with simple data augmentation. Advances in Neural Information Processing Systems, 36: 65484–65516, 2023. 1

    2023

  12. [20]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6

    2014

  13. [21]

    Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022. 1, 2, 6

    2022 arXiv

  14. [22]

    Learning-to-cache: Accelerating diffusion trans- former via layer caching.Advances in Neural Information Processing Systems, 37:133282–133304, 2024

    Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion trans- former via layer caching.Advances in Neural Information Processing Systems, 37:133282–133304, 2024. 5

    2024

  15. [23]

    Diffusion models beat gans on image classification.arXiv preprint arXiv:2307.08702, 2023

    Soumik Mukhopadhyay, Matthew Gwilliam, Vatsal Agar- wal, Namitha Padmanabhan, Archana Swaminathan, Srinidhi Hegde, Tianyi Zhou, and Abhinav Shrivastava. Diffusion models beat gans on image classification.arXiv preprint arXiv:2307.08702, 2023. 1

    2023 arXiv

  16. [24]

    Reliable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models. InInternational conference on machine learning, pages 7176–7185. PMLR, 2020. 7

    2020

  17. [25]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021. 1

    2021 arXiv

  18. [26]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 4195–4205,

  19. [27]

    Hopfield networks is all you need.arXiv preprint arXiv:2008.02217, 2020

    Hubert Ramsauer, Bernhard Sch ¨afl, Johannes Lehner, Philipp Seidl, Michael Widrich, Thomas Adler, Lukas Gruber, Markus Holzleitner, Milena Pavlovi ´c, Geir Kjetil Sandve, et al. Hopfield networks is all you need.arXiv preprint arXiv:2008.02217, 2020. 5, 6, 1

    2008 arXiv

  20. [28]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1

    2022

  21. [29]

    Cads: Unleashing the di- versity of diffusion models through condition-annealed sam- pling.arXiv preprint arXiv:2310.17347, 2023

    Seyedmorteza Sadat, Jakob Buhmann, Derek Bradley, Otmar Hilliges, and Romann M Weber. Cads: Unleashing the di- versity of diffusion models through condition-annealed sam- pling.arXiv preprint arXiv:2310.17347, 2023. 1, 2, 6, 7

    2023 arXiv

  22. [30]

    Eliminating oversaturation and artifacts of high guid- ance scales in diffusion models

    Seyedmorteza Sadat, Otmar Hilliges, and Romann M We- ber. Eliminating oversaturation and artifacts of high guid- ance scales in diffusion models. InThe Thirteenth Interna- tional Conference on Learning Representations, 2024. 1, 2, 6, 7

    2024

  23. [31]

    Rethinking the spatial inconsistency in classifier- free diffusion guidance

    Dazhong Shen, Guanglu Song, Zeyue Xue, Fu-Yun Wang, and Yu Liu. Rethinking the spatial inconsistency in classifier- free diffusion guidance. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9370–9379, 2024. 1

    2024

  24. [32]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational confer- ence on machine learning, pages 2256–2265. pmlr, 2015. 1

    2015

  25. [33]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 3, 6

    2010 arXiv

  26. [34]

    Score-based generative modeling through stochastic differential equa- tions.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions.arXiv preprint arXiv:2011.13456, 2020. 1

    2011 arXiv

  27. [35]

    Audit: Audio editing by following instruc- tions with latent diffusion models.Advances in Neural In- formation Processing Systems, 36:71340–71357, 2023

    Yuancheng Wang, Zeqian Ju, Xu Tan, Lei He, Zhizheng Wu, Jiang Bian, et al. Audit: Audio editing by following instruc- tions with latent diffusion models.Advances in Neural In- formation Processing Systems, 36:71340–71357, 2023. 1

    2023

  28. [36]

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

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341,

    arXiv

  29. [2023]

    patterns

    1, 2, 4, 7 10 A. Background As discussed in Section 3.4.3, Eq. 14 gives the update rule for modern Hopfield networks [27]. From Theorems 1 and 2 in [27], for fixed patternsX, the iterations{ζ t}∞ t=0 gener- ated by Eq. 14 monotonically decrease the energyE(ζ t) and converge to...

    2014

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.