Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: Hopfield Perspective

Hyunmin Cho; Kyong Hwan Jin; Woo Kyoung Han

arxiv: 2605.27476 · v1 · pith:FM4CIRUJnew · submitted 2026-05-26 · 💻 cs.LG · cs.AI

Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: Hopfield Perspective

Hyunmin Cho , Woo Kyoung Han , Kyong Hwan Jin This is my paper

Pith reviewed 2026-06-29 19:57 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords diffusion modelsattention mechanismHopfield networksfidelity diversity tradeoffsymmetric decompositionenergy landscapegenerative models

0 comments

The pith

Decomposing the pre-softmax attention matrix into symmetric and skew-symmetric parts links Hopfield stability to the fidelity-diversity trade-off in diffusion models and supplies a circulation knob for control.

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

The paper treats the pre-softmax attention matrix QK^T as an associative memory that records pairwise feature associations. It splits the matrix so the symmetric half defines an energy landscape while the skew-symmetric half drives circulation across that landscape. Stability quantities are then extracted from the energy landscape in the manner of Hopfield networks. These quantities are reported to correlate with how diffusion sampling trades accurate reproduction against output variety. The skew-symmetric term is shown to act as an adjustable parameter that shifts the operating point on the trade-off curve.

Core claim

By viewing QK^T as encoding associations, its symmetric decomposition governs the energy minima that determine stable feature retrieval during sampling; the derived stability indices exhibit direct relations to measured fidelity and diversity scores, while skew-symmetric adjustments serve as a tunable parameter for shifting the operating point on the trade-off curve.

What carries the argument

Symmetric-skew decomposition of the pre-softmax attention matrix, inducing a Hopfield energy landscape whose stability measures quantify retrieved feature robustness.

If this is right

Correlations appear between the stability measures and observed fidelity-diversity metrics across generated samples.
Adjusting the skew-symmetric circulation term provides direct control over the trade-off without retraining the model.
Energy landscape interpretation explains why certain attention patterns lead to mode collapse or excessive diversity in outputs.

Where Pith is reading between the lines

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

Similar decomposition might apply to other attention-based generative models beyond diffusion.
Stability measures could serve as training-time regularizers to target specific fidelity-diversity points.
If the energy minima correspond to data modes, this links attention dynamics to data manifold geometry.

Load-bearing premise

The symmetric component of the attention matrix creates an energy function whose local minima align with the stable features that diffusion sampling retrieves, making the stability numbers causally predictive of output quality.

What would settle it

An experiment that computes the proposed stability measures on attention matrices from many generated samples and finds no statistical correlation with their individual fidelity or diversity scores would falsify the link.

Figures

Figures reproduced from arXiv: 2605.27476 by Hyunmin Cho, Kyong Hwan Jin, Woo Kyoung Han.

**Figure 1.** Figure 1: Skew perturbation and the fidelity–diversity tradeoff. Top: We decompose QK⊤ into symmetric (energy) and skew (circulation) parts. (a) The symmetric part gives stable but low-diversity retrieval. (b) Moderate skew perturbation breaks metastable mixtures while preserving stable states. (c) Excessive perturbation destabilizes even well-formed retrievals, producing artifacts. Bottom: Moderate skew perturbati… view at source ↗

**Figure 2.** Figure 2: Associative memory framework encoding pairwise feature interactions and its decomposition. (a)–(b) We characterize the attention mechanism as an associative memory encoding pairwise feature interactions. (a) Viewing input features X ∈ R L×din as a set of features x (i) ∈ R L , (b) the learned interaction matrix W encodes the association strength between these feature pairs. (c) The resulting attention matr… view at source ↗

**Figure 3.** Figure 3: Visualization of samples generated via decomposed components. Samples generated through the sym. component encapsulate the underlying global structure, whereas those generated via the Skew component manifest fine-grained, irregular details. symmetric component as: EX(ξ) ≜ − 1 2 ξ ⊤Msym(X) ξ. (21) Lower energy (i.e., more negative EX) corresponds to a feature ξ that is more strongly supported by the associ… view at source ↗

**Figure 4.** Figure 4: Qualitative comparison from samples sorted by Alignment Score. For three prompts, we group baseline generations into Stable (top row) and Unstable (bottom row) subsets according to AlignX. Stable samples show coherent, object-centric structures, whereas unstable samples exhibit diverse but less coherent mixtures. White labels indicate the corresponding AlignX values. This expansion explicitly characterizes… view at source ↗

**Figure 5.** Figure 5: provides a schematic summary of these three stability measures [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: Qualitative visualization of the stability spectrum. Baseline samples are sorted by their Alignment Score AlignX(ξ). High-Alignment samples (Stable) exhibit structural coherence and consistent object-centric compositions. In contrast, LowAlignment samples (Unstable) display fragmented structures and incompatible texture mixtures, indicating metastable entrapment. Fidelity–Diversity trade-off via stability… view at source ↗

**Figure 7.** Figure 7: Qualitative results of feature blending. Perturbation on unstable sample (left): perturbation breaks spurious mixture configurations and yields a cleaner, object-centric reconstruction. Perturbation on stable sample (right): perturbation injects variation (texture/background/composition) and may introduce drift, illustrating the operating-point trade-off. Here, α governs the intensity of the circulation pe… view at source ↗

**Figure 9.** Figure 9: Ablation against attention temperature τ scaling. Relative to the SYM.ONLY reference, temperature scaling can introduce unintended structures (e.g., additional leg) due to nonselective strengthening/weakening of interactions across the scene. Instead, our control better preserves strongly supported structure while suppressing weakly supported mixture artifacts. 6.2. Circulation Control and Global Temperin… view at source ↗

**Figure 8.** Figure 8: Effectiveness of adaptive circulation control. For the prompt “A fancy clock ... with red carpet,” moderate circulation improves the baseline structure, whereas excessive static circulation introduces visible distortion. Adaptive control reduces this over-perturbation and preserves a more coherent object structure. where b and h index the sample and attention head, respectively. We then modulate only the… view at source ↗

**Figure 10.** Figure 10: Qualitative results of SDXL + Ours. Left (successful cases): the highlighted concepts are weakly represented or missing in the baseline, and Ours renders them more faithfully (e.g., adding missing objects or correcting on-screen/scene content). Right (failure cases): on prompts the baseline already handles well, Ours can mildly degrade the highlighted aspect (e.g., the sandwich form, the “carrying” pose, … view at source ↗

read the original abstract

We characterize the pre-softmax attention matrix $\mathbf{QK^\top}$ in transformers as an associative memory matrix encoding pairwise associations between input features. By decomposing this matrix into its symmetric and skew-symmetric parts, we interpret the symmetric component as governing the structure of the energy landscape, and the skew-symmetric component as driving circulation on that landscape. Leveraging the energy formulation induced by the symmetric component, we derive Hopfield-style stability measures that quantify the stability of retrieved features. We observe meaningful correlations between Hopfield-style stability measures and the fidelity-diversity trade-offs in generation. Finally, we propose a controllable knob to modulate this trade-off by modifying the circulation of the underlying dynamics. Code is available at our GitHub (https://github.com/hyeon-cho/Attention-Symmetric-Decomposition).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper decomposes pre-softmax attention into symmetric and skew parts to extract Hopfield stability scores and a circulation knob for diffusion fidelity-diversity control, but the mapping from that energy to actual reverse-process attractors stays unshown.

read the letter

The new piece is taking the standard symmetric/skew split of QK^T, reading the symmetric part as a Hopfield energy, pulling stability quantities from it, and then using the skew part as a tunable circulation term to shift the fidelity-diversity balance. They report correlations between those stability numbers and generation metrics, plus a knob that modifies the skew component.

The decomposition itself is clean linear algebra and the Hopfield framing gives a compact way to talk about feature stability inside the attention matrix. If the reported correlations are reproducible and the knob actually moves the metrics without breaking the model, it could be a practical handle for people already running diffusion transformers.

The soft spot is the central assumption that the symmetric component creates an energy landscape whose minima line up with the points stabilized during diffusion sampling. Diffusion follows a learned score function and reverse SDE, not descent on that associative-memory energy. The abstract gives no derivation showing that argmin of the symmetric energy matches fixed points in the denoising chain, nor experiments that isolate the effect of the circulation change from other changes to the score estimate. Without that link, the stability measures remain descriptive rather than explanatory.

The code release helps, but the paper is still early. It is aimed at researchers who work on attention interpretability inside generative models. The combination is fresh enough and the experimental claims are concrete enough that it should go to referees rather than desk reject, though the reviewers will need to press on whether the energy-to-sampling correspondence holds.

Referee Report

2 major / 1 minor

Summary. The paper decomposes the pre-softmax attention matrix QK^T in transformers into symmetric and skew-symmetric parts, interpreting the symmetric component as defining an energy landscape (Hopfield-style) and the skew-symmetric as driving circulation. From the symmetric part it derives stability measures, reports correlations between these measures and fidelity-diversity trade-offs in diffusion-model generation, and proposes a controllable knob obtained by modulating the circulation term.

Significance. If the claimed causal link between the derived Hopfield stability quantities and generation metrics holds and the circulation knob can be applied without retraining or altering the underlying score function, the work would supply a new, interpretable mechanism for trading off fidelity and diversity that is grounded in associative-memory dynamics rather than ad-hoc sampling adjustments.

major comments (2)

[Abstract] Abstract (and the central claim): the mapping from the symmetric part S of QK^T to an energy E whose local minima govern stable features retrieved during diffusion sampling is asserted but not derived or experimentally verified. Diffusion trajectories follow the learned reverse SDE/ODE, not gradient descent on E(S); no alignment between argmin E and stabilized points in the sampling chain is shown, undermining the causal interpretation of the reported correlations.
[Abstract] The proposed circulation-modulation knob is presented as controllable, yet the manuscript provides no ablation confirming that changes to the skew-symmetric component leave the score estimate unchanged while only affecting the claimed energy landscape; this is load-bearing for the claim that the knob balances fidelity and diversity without side effects.

minor comments (1)

The GitHub link is supplied but no statement is made about whether the released code reproduces the exact stability-measure derivations and correlation tables reported in the paper.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's insightful comments on the abstract and central claims. Below we provide point-by-point responses, agreeing to revisions that clarify the interpretive framework and add supporting ablations.

read point-by-point responses

Referee: [Abstract] Abstract (and the central claim): the mapping from the symmetric part S of QK^T to an energy E whose local minima govern stable features retrieved during diffusion sampling is asserted but not derived or experimentally verified. Diffusion trajectories follow the learned reverse SDE/ODE, not gradient descent on E(S); no alignment between argmin E and stabilized points in the sampling chain is shown, undermining the causal interpretation of the reported correlations.

Authors: Our framework draws an analogy to Hopfield networks, where the symmetric component of the weight matrix defines an energy landscape with local minima corresponding to stable patterns. The stability measures are derived from this symmetric part S and shown to correlate with fidelity-diversity trade-offs observed in diffusion generation. We do not claim or derive that the diffusion sampling trajectory performs gradient descent on this energy E; the reverse SDE is followed as learned. The correlations are empirical observations supporting the utility of these measures. We agree that the causal link is not fully established without alignment verification and will revise the abstract to tone down the language from 'govern' to 'analogous to' and add a discussion on the limitations of the analogy. revision: yes
Referee: [Abstract] The proposed circulation-modulation knob is presented as controllable, yet the manuscript provides no ablation confirming that changes to the skew-symmetric component leave the score estimate unchanged while only affecting the claimed energy landscape; this is load-bearing for the claim that the knob balances fidelity and diversity without side effects.

Authors: The knob is designed by modulating the skew-symmetric part while preserving the symmetric part, with the intention that the energy landscape remains the same but circulation changes the dynamics. Since the score function in diffusion models is learned from the full attention, we recognize that an explicit check that the modulated attention does not alter the effective score estimate is missing. We will perform and include an ablation study in the revision that applies the modulation at inference time and verifies that key generation statistics (beyond the target diversity) remain consistent with the unmodulated model, thereby confirming minimal side effects. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation remains self-contained

full rationale

The paper decomposes the pre-softmax QK^T matrix into symmetric and skew-symmetric parts, interprets the symmetric component as inducing an energy landscape, derives Hopfield-style stability measures from that formulation, reports empirical correlations with fidelity-diversity metrics, and proposes a circulation-modifying knob. None of these steps reduce by the paper's own equations to a fitted input renamed as prediction, a self-citation chain, or a definitional tautology; the stability quantities and knob follow directly from the decomposition and external observations rather than being forced by construction. The mapping to diffusion sampling dynamics is an interpretive assumption, not a circular reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claims rest on a matrix decomposition treated as an energy landscape and on an untested mapping from that landscape to diffusion sampling dynamics; no free parameters are numerically fitted in the abstract, but the modulation knob is introduced without stated derivation.

free parameters (1)

circulation_modulation_scale
A controllable knob is proposed to alter circulation; its functional form and any fitting procedure are not specified in the abstract.

axioms (1)

domain assumption The pre-softmax attention matrix QK^T encodes pairwise associations that can be additively decomposed into symmetric and skew-symmetric components with distinct dynamical roles.
This decomposition is the starting point for all subsequent energy and circulation interpretations.

invented entities (1)

Hopfield-style stability measure no independent evidence
purpose: Quantifies how firmly a retrieved feature sits at an energy minimum induced by the symmetric attention component.
Introduced as a derived diagnostic for diffusion generation; no independent falsifiable prediction outside the paper is stated.

pith-pipeline@v0.9.1-grok · 5665 in / 1521 out tokens · 36167 ms · 2026-06-29T19:57:36.608788+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 10 canonical work pages · 3 internal anchors

[1]

URL https://openreview.net/forum? id=hkV9CvCOjH. Amit, D. J., Gutfreund, H., and Sompolinsky, H. Spin- glass models of neural networks.Phys. Rev. A, 32: 1007–1018, Aug 1985. doi: 10.1103/PhysRevA.32

work page doi:10.1103/physreva.32 1985
[2]

1103/PhysRevA.32.1007

URL https://link.aps.org/doi/10. 1103/PhysRevA.32.1007. Bietti, A., Cabannes, V ., Bouchacourt, D., Jegou, H., and Bottou, L. Birth of a transformer: A memory viewpoint. InThirty-seventh Conference on Neural Information Pro- cessing Systems, 2023. URL https://openreview. net/forum?id=3X2EbBLNsk. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., ...

2023
[3]

cc/paper_files/paper/2020/file/ 1457c0d6bfcb4967418bfb8ac142f64a-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 1457c0d6bfcb4967418bfb8ac142f64a-Paper. pdf. Chen, X., Liu, N., Zhu, Y ., Feng, F., and Tang, J. EDT: An efficient diffusion transformer framework inspired by human-like sketching. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems,

2020
[4]

Boundary-Value Problems with Non-Local Initial Condition for Parabolic Equations with Parameter

URL https://openreview.net/forum? id=MihOCXte41. Chengxiang, Z., Dasgupta, C., and Singh, M. P. Retrieval properties of a hopfield model with random asymmetric interactions.Neural Computation, 12(4):865–880, 2000. doi: 10.1162/089976600300015628. Derrida, B., Gardner, E., and Zippelius, A. An exactly solv- able asymmetric neural network model.Europhysics ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1162/089976600300015628 2000
[5]

emnlp-main.595/

URL https://aclanthology.org/2021. emnlp-main.595/. 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. In Guyon, I., Luxburg, U. V ., Bengio, S., Wallach, 10 Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: ...

2021
[6]

cc/paper_files/paper/2017/file/ 8a1d694707eb0fefe65871369074926d-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2017/file/ 8a1d694707eb0fefe65871369074926d-Paper. pdf. Ho, J. and Salimans, T. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. URL https:// openreview.net/forum?id=qw8AKxfYbI. Ho, J., Jain, A., and Abbeel, P. Denoising diffusion ...

2017
[7]

cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper. pdf. Hong, S. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention.Ad- vances in Neural Information Processing Systems, 37: 66743–66772, 2024. Hoover, B., Strobelt, H., Krotov, D., Hoffman, J., Kira, Z., and Cha...

work page doi:10.1073/pnas.79.8 2020
[8]

Hwang, S., Folli, V ., Lanza, E., Parisi, G., Ruocco, G., and Zamponi, F

URL https://www.pnas.org/doi/abs/ 10.1073/pnas.79.8.2554. Hwang, S., Folli, V ., Lanza, E., Parisi, G., Ruocco, G., and Zamponi, F. On the number of limit cycles in asymmetric neural networks.Journal of Statistical Me- chanics: Theory and Experiment, 2019(5):053402, May

work page doi:10.1073/pnas.79.8.2554 2019
[9]

doi: 10.1088/1742-5468/ ab11e3

ISSN 1742-5468. doi: 10.1088/1742-5468/ ab11e3. URL http://dx.doi.org/10.1088/ 1742-5468/ab11e3. Kim, K. and Sim, B. Pladis: Pushing the limits of atten- tion in diffusion models at inference time by leveraging sparsity. InProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 16238–16248, 2025. Krotov, D. and Hopfield, J. J. Dense a...

work page doi:10.1088/1742-5468/ 2025
[10]

FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

URL https://proceedings.neurips. cc/paper_files/paper/2016/file/ eaae339c4d89fc102edd9dbdb6a28915-Paper. pdf. Labs, B. F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y ., Li, C., Lorenz, D., M¨uller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., and Sm...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/0025-5564(74)90031-5 2016
[11]

org/CorpusID:38591603

URL https://api.semanticscholar. org/CorpusID:38591603. Nichol, A. Q. and Dhariwal, P. Improved denoising dif- fusion probabilistic models. In Meila, M. and Zhang, T. (eds.),Proceedings of the 38th International Confer- ence on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pp. 8162–8171. PMLR, 18–24 Jul 2021. URLhttps://proceedi...

2021
[12]

cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740-Paper. pdf. Peebles, W. and Xie, S. Scalable diffusion models with trans- formers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4195–4205, October 2023. Peretto, P. Collective properties of neural networks: A statistic...

work page doi:10.1007/bf00317939 2019
[13]

Podell, D., English, Z., Lacey, K., Blattmann, A., Dock- horn, T., M¨uller, J., Penna, J., and Rombach, R

URL https://openreview.net/forum? id=IWZnhP3YgK. Podell, D., English, Z., Lacey, K., Blattmann, A., Dock- horn, T., M¨uller, J., Penna, J., and Rombach, R. SDXL: Improving latent diffusion models for high-resolution image synthesis. InThe Twelfth International Confer- ence on Learning Representations, 2024. URL https: //openreview.net/forum?id=di52zR8xgf....

2024
[14]

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B

URL https://openreview.net/forum? id=tL89RnzIiCd. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with la- tent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10684–10695, June 2022. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C. W., Wig...

work page doi:10.1103/physreve.52.5261 2022
[15]

cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper. pdf. 12 Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: Hopfield Perspective Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative mod- eling through stochas...

2019
[16]

LLaMA: Open and Efficient Foundation Language Models

URL https://openreview.net/forum? id=PxTIG12RRHS. Stein, G., Cresswell, J. C., Hosseinzadeh, R., Sui, Y ., Ross, B. L., Villecroze, V ., Liu, Z., Caterini, A. L., Taylor, E., and Loaiza-Ganem, G. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. InThirty-seventh Conference on Neural Information Processin...

work page internal anchor Pith review Pith/arXiv arXiv 2023
[17]

bti,hij,bsj->bhts

URL https://proceedings.neurips. cc/paper_files/paper/2017/file/ 3f5ee243547dee91fbd053c1c4a845aa-Paper. pdf. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lam- bert, N., Rasul, K., Davaadorj, M., Nair, D., Paul, S., Berman, W., Xu, Y ., Liu, S., and Wolf, T. Diffusers: State-of-the-art diffusion models. https://github. com/huggingface/diffusers, 20...

work page arXiv 2017

[1] [1]

URL https://openreview.net/forum? id=hkV9CvCOjH. Amit, D. J., Gutfreund, H., and Sompolinsky, H. Spin- glass models of neural networks.Phys. Rev. A, 32: 1007–1018, Aug 1985. doi: 10.1103/PhysRevA.32

work page doi:10.1103/physreva.32 1985

[2] [2]

1103/PhysRevA.32.1007

URL https://link.aps.org/doi/10. 1103/PhysRevA.32.1007. Bietti, A., Cabannes, V ., Bouchacourt, D., Jegou, H., and Bottou, L. Birth of a transformer: A memory viewpoint. InThirty-seventh Conference on Neural Information Pro- cessing Systems, 2023. URL https://openreview. net/forum?id=3X2EbBLNsk. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., ...

2023

[3] [3]

cc/paper_files/paper/2020/file/ 1457c0d6bfcb4967418bfb8ac142f64a-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 1457c0d6bfcb4967418bfb8ac142f64a-Paper. pdf. Chen, X., Liu, N., Zhu, Y ., Feng, F., and Tang, J. EDT: An efficient diffusion transformer framework inspired by human-like sketching. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems,

2020

[4] [4]

Boundary-Value Problems with Non-Local Initial Condition for Parabolic Equations with Parameter

URL https://openreview.net/forum? id=MihOCXte41. Chengxiang, Z., Dasgupta, C., and Singh, M. P. Retrieval properties of a hopfield model with random asymmetric interactions.Neural Computation, 12(4):865–880, 2000. doi: 10.1162/089976600300015628. Derrida, B., Gardner, E., and Zippelius, A. An exactly solv- able asymmetric neural network model.Europhysics ...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1162/089976600300015628 2000

[5] [5]

emnlp-main.595/

URL https://aclanthology.org/2021. emnlp-main.595/. 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. In Guyon, I., Luxburg, U. V ., Bengio, S., Wallach, 10 Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: ...

2021

[6] [6]

cc/paper_files/paper/2017/file/ 8a1d694707eb0fefe65871369074926d-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2017/file/ 8a1d694707eb0fefe65871369074926d-Paper. pdf. Ho, J. and Salimans, T. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. URL https:// openreview.net/forum?id=qw8AKxfYbI. Ho, J., Jain, A., and Abbeel, P. Denoising diffusion ...

2017

[7] [7]

cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper. pdf. Hong, S. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention.Ad- vances in Neural Information Processing Systems, 37: 66743–66772, 2024. Hoover, B., Strobelt, H., Krotov, D., Hoffman, J., Kira, Z., and Cha...

work page doi:10.1073/pnas.79.8 2020

[8] [8]

Hwang, S., Folli, V ., Lanza, E., Parisi, G., Ruocco, G., and Zamponi, F

URL https://www.pnas.org/doi/abs/ 10.1073/pnas.79.8.2554. Hwang, S., Folli, V ., Lanza, E., Parisi, G., Ruocco, G., and Zamponi, F. On the number of limit cycles in asymmetric neural networks.Journal of Statistical Me- chanics: Theory and Experiment, 2019(5):053402, May

work page doi:10.1073/pnas.79.8.2554 2019

[9] [9]

doi: 10.1088/1742-5468/ ab11e3

ISSN 1742-5468. doi: 10.1088/1742-5468/ ab11e3. URL http://dx.doi.org/10.1088/ 1742-5468/ab11e3. Kim, K. and Sim, B. Pladis: Pushing the limits of atten- tion in diffusion models at inference time by leveraging sparsity. InProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 16238–16248, 2025. Krotov, D. and Hopfield, J. J. Dense a...

work page doi:10.1088/1742-5468/ 2025

[10] [10]

FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

URL https://proceedings.neurips. cc/paper_files/paper/2016/file/ eaae339c4d89fc102edd9dbdb6a28915-Paper. pdf. Labs, B. F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y ., Li, C., Lorenz, D., M¨uller, J., Podell, D., Rombach, R., Saini, H., Sauer, A., and Sm...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/0025-5564(74)90031-5 2016

[11] [11]

org/CorpusID:38591603

URL https://api.semanticscholar. org/CorpusID:38591603. Nichol, A. Q. and Dhariwal, P. Improved denoising dif- fusion probabilistic models. In Meila, M. and Zhang, T. (eds.),Proceedings of the 38th International Confer- ence on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pp. 8162–8171. PMLR, 18–24 Jul 2021. URLhttps://proceedi...

2021

[12] [12]

cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ bdbca288fee7f92f2bfa9f7012727740-Paper. pdf. Peebles, W. and Xie, S. Scalable diffusion models with trans- formers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4195–4205, October 2023. Peretto, P. Collective properties of neural networks: A statistic...

work page doi:10.1007/bf00317939 2019

[13] [13]

Podell, D., English, Z., Lacey, K., Blattmann, A., Dock- horn, T., M¨uller, J., Penna, J., and Rombach, R

URL https://openreview.net/forum? id=IWZnhP3YgK. Podell, D., English, Z., Lacey, K., Blattmann, A., Dock- horn, T., M¨uller, J., Penna, J., and Rombach, R. SDXL: Improving latent diffusion models for high-resolution image synthesis. InThe Twelfth International Confer- ence on Learning Representations, 2024. URL https: //openreview.net/forum?id=di52zR8xgf....

2024

[14] [14]

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B

URL https://openreview.net/forum? id=tL89RnzIiCd. Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with la- tent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10684–10695, June 2022. Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C. W., Wig...

work page doi:10.1103/physreve.52.5261 2022

[15] [15]

cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper. pdf. 12 Balancing Fidelity and Diversity in Diffusion Models via Symmetric Attention Decomposition: Hopfield Perspective Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative mod- eling through stochas...

2019

[16] [16]

LLaMA: Open and Efficient Foundation Language Models

URL https://openreview.net/forum? id=PxTIG12RRHS. Stein, G., Cresswell, J. C., Hosseinzadeh, R., Sui, Y ., Ross, B. L., Villecroze, V ., Liu, Z., Caterini, A. L., Taylor, E., and Loaiza-Ganem, G. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. InThirty-seventh Conference on Neural Information Processin...

work page internal anchor Pith review Pith/arXiv arXiv 2023

[17] [17]

bti,hij,bsj->bhts

URL https://proceedings.neurips. cc/paper_files/paper/2017/file/ 3f5ee243547dee91fbd053c1c4a845aa-Paper. pdf. von Platen, P., Patil, S., Lozhkov, A., Cuenca, P., Lam- bert, N., Rasul, K., Davaadorj, M., Nair, D., Paul, S., Berman, W., Xu, Y ., Liu, S., and Wolf, T. Diffusers: State-of-the-art diffusion models. https://github. com/huggingface/diffusers, 20...

work page arXiv 2017