Pith. sign in

REVIEW 2 major objections 5 minor 6 cited by

Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A small network trained only to adjust the starting noise of a frozen diffusion generator can recover much of the quality gain of per-image test-time optimization, via a tractable noise-space objective.

desk verdict A practical, well-executed shortcut for reward-aligned one-step generation, but the theoretical justification leans on a Lipschitz condition the paper never checks. read the letter →

arxiv 2508.09968 v1 pith:ZQXS6O3X submitted 2025-08-13 cs.LG cs.CV

Luca Eyring , Shyamgopal Karthik , Alexey Dosovitskiy , Nataniel Ruiz , Zeynep Akata This is my paper
classification cs.LGcs.CV
keywords noisehypernetworktest-timescalingdiffusionmodelsreward-tilteddistributionnoise-spaceKLdivergenceLoRAtext-to-imagealignmentrewardfine-tuning
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

The paper's claim is that the expensive part of test-time scaling in diffusion models — per-prompt optimization of the starting noise, which takes many forward and backward passes through the generator — can be moved into a one-time post-training stage. A lightweight noise hypernetwork $f_\phi$, built as a LoRA adapter on a frozen, step-distilled generator, maps standard Gaussian noise $z_0$ to a modulated latent $z_0 + f_\phi(z_0)$; passing that through the fixed generator produces outputs close to the reward-tilted distribution $p^\star(x) \propto p_{\mathrm{base}}(x)\exp(r(x)/\alpha)$. The paper derives that the KL divergence keeping the modulated noise faithful to the Gaussian prior reduces, under a small-Lipschitz condition on $f_\phi$, to an $\ell^2$ penalty on the perturbation magnitude, yielding the tractable objective $\mathbb{E}[\tfrac{1}{2}\|f_\phi(z_0)\|^2 - r(g_\theta(z_0 + f_\phi(z_0)))]$. Experiments on SD-Turbo, SANA-Sprint, and FLUX-Schnell show one extra forward pass recovering a substantial share of the quality gains of explicit test-time noise optimization such as ReNO, where direct reward fine-tuning of the same models degrades or collapses. If the paper is right, reward-aligned generation from fast few-step models becomes practical in settings where minutes of per-image optimization is not an option.

What carries the argument

The central object is the noise hypernetwork $f_\phi$: a LoRA adapter on the frozen generator that maps standard noise $z_0$ to a perturbation, so the generator receives $z_0 + f_\phi(z_0)$. The argument is carried by the tilted noise distribution $p^\star_0(z_0) \propto p_0(z_0)\exp(r(g_\theta(z_0))/\alpha)$ — the exact input distribution whose pushforward through $g_\theta$ equals the reward-tilted output — and by the noise-space KL formula $\mathrm{KL}(\rho_\phi^0 \| \rho_0) = \mathbb{E}[\tfrac{1}{2}\|f_\phi\|^2 + \mathrm{Tr}(\log(I + J_{f_\phi}))]$, derived via change of variables and Stein's lemma. Theorem 1 bounds the log-determinant term under a Lipschitz assumption, justifying its re

What would settle it

Measure the empirical Lipschitz constant of the trained hypernetwork — the top singular value of its Jacobian on Gaussian inputs — and compare it with the Theorem 1 regime. If the network operates with Lipschitz constant well above 1 while the GenEval gains are undiminished, the tractable $\ell^2$ objective is not the faithful KL surrogate the theory requires. Alternatively, retrain HyperNoise with the $\ell^2$ regularization removed: if the gains persist without reward-hacking artifacts, the fidelity mechanism is not what is driving the improvements.

Watch

Extended reading notes

Core claim

Steering a frozen distilled generator to the tilted output distribution $p^\star(x) \propto p_{\mathrm{base}}(x)\exp(r(x)/\alpha)$ is equivalent to drawing initial noise from $p^\star_0(z_0) \propto p_0(z_0)\exp(r(g_\theta(z_0))/\alpha)$. The paper learns this tilted noise distribution via a noise-space KL that reduces to the tractable objective $\mathbb{E}[\tfrac{1}{2}\|f_\phi(z_0)\|^2 - r(g_\theta(z_0 + f_\phi(z_0)))]$, with $f_\phi$ a zero-initialized LoRA adapter. The derivation uses change of variables and Stein's lemma; Theorem 1 bounds the leftover log-determinant term when $f_\phi$ is Lipschitz with constant below 1, and the data-processing inequality ties the noise-space KL to outpu

Load-bearing premise

The load-bearing premise is that the average squared magnitude of the noise perturbation equals the true KL divergence between the modulated and original noise distributions; Theorem 1 guarantees this only when the perturbation network is Lipschitz with constant below 1, a condition the training neither enforces nor measures.

Editorial extensions

If this is right

  • Reward alignment becomes a post-training step: one forward pass through the hypernetwork replaces the tens of forward-backward passes that ReNO-style per-image optimization needs, at a fraction of the inference latency.
  • Because the generator is frozen, the same base model can be steered by training a separate hypernetwork per reward function, with no generator fine-tuning and no need to keep multiple generator copies.
  • A hypernetwork trained for one-step generation also improves two-, four-, eight-, and sixteen-step inference of the same model, even though the generator was not trained for that regime.
  • Direct reward fine-tuning of distilled generators degrades GenEval performance and introduces artifacts; the noise-space regularizer does not, so the method preserves base-model fidelity and output diversity while optimizing the reward.

Reading between the lines

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

  • An implication the paper leaves implicit: with the generator frozen, failures of a HyperNoise-aligned model are attributable to the reward model rather than to generator collapse — the redness experiment shows the method cannot exploit out-of-distribution image artifacts, so it could serve as a diagnostic for reward-model blind spots.
  • The objective only needs Gaussian noise, a fixed deterministic generator, and a differentiable reward; nothing in the derivation is text-to-image specific, so the same amortization plausibly transfers to other deterministic generators such as audio, video, or control policies. The paper demonstrates only text-to-image models, making this an extrapolation.
  • A testable extension the paper does not explore: use the hypernetwork's output as a warm start for a short noise-optimization refinement, converting the remaining gap to full test-time optimization into a tunable compute-versus-quality knob.
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

2 major / 5 minor

Summary. The paper proposes HyperNoise, a post-training method for step-distilled text-to-image diffusion models. It trains a lightweight LoRA-based noise hypernetwork f_phi that maps standard Gaussian noise z0 to a modulated latent z0 + f_phi(z0,c); the frozen generator g_theta then produces the image. The training objective is Eq. (11): L = E[ 1/2 ||f_phi(z0)||^2 - r(g_theta(z0 + f_phi(z0,c)), c) ]. The first term is presented as a tractable approximation of KL(p_phi0 || p0), and the second maximizes reward. The authors derive the reward-tilted noise distribution p*_0 ∝ p0 exp(r(g_theta(z0))/alpha) and prove that its pushforward through g_theta equals the data-space tilted distribution p*(x) ∝ p_base(x) exp(r(x)/alpha). They also relate this result to stochastic optimal control for one-step generators. Experiments on SD-Turbo, SANA-Sprint, and FLUX-Schnell report consistent GenEval gains (e.g., SANA-Sprint 0.70→0.75, FLUX-Schnell 0.68→0.72) with near-baseline inference latency, plus improvements on T2I-CompBench and DPG-Bench and generalization across 1-32 NFEs. Comparisons include ReNO, Best-of-N, and prompt optimization.

Significance. The empirical contribution is solid: gains are consistent across three model families and several benchmarks, the method is parameter-efficient, and inference overhead is small. Releasing code and reporting hyperparameters supports reproducibility. The derivation of the tilted noise distribution and its relation to the tilted output distribution is correct, and the connection to stochastic optimal control is a useful sanity check. The main weakness is that the tractable KL surrogate rests on a small-Lipschitz condition that is neither enforced nor verified; the paper itself concedes this in Appendix A.4. If the authors add verification or enforcement, the theoretical claim becomes credible. As written, this is a useful empirical contribution with an incompletely supported theoretical framing.

major comments (2)
  1. [Appendix A.4, Eq. (10)-(11)] The central loss (11) is justified by replacing KL(p_phi0 || p0) with E_{z0~p0}[1/2 ||f_phi(z0)||^2]. Theorem 7 bounds the omitted log-determinant term only when f_phi has Lipschitz constant L < 1, with an error growing like L^2/(1-L). The paper explicitly states in Appendix A.4: 'we do not explicitly enforce L<1 during training.' The L2 magnitude penalty does not control the Jacobian: a residual can have small amplitude with large gradients, and a LoRA residual over transformer blocks can plausibly have singular values above 1. Since Eq. (11) is the training objective, the claimed equivalence between minimizing (11) and minimizing KL(p_phi0 || p*_0) is not established. Please either enforce a Lipschitz/contractivity condition (e.g., spectral normalization or a constraint on the LoRA weights) or empirically verify the approximation regime by measuring the trained f_phi's Jacobian spectra
  2. [Section 3.1 / Appendix A.3, Eq. (5)-(6)] The method restricts modulated noise distributions to the pushforward family (id + f_phi)_# p0. The paper does not show that this family can represent the tilted noise distribution p*_0 in Eq. (5), nor does it bound the approximation error KL(p_phi0 || p*_0) attainable within the family. The LoRA rank ablation in Appendix C.5 shows sensitivity to rank (GenEval 0.75 at ranks 64/128 vs 0.70-0.71 at ranks 8/16), which is indirect evidence of expressivity limitations. A discussion of expressivity, or an empirical estimate of the residual gap between the learned noise distribution and p*_0, would make the theoretical claim more complete.
minor comments (5)
  1. [Theorem numbering] Theorem 1 in the main text is labeled Theorem 7 in Appendix A.4, and Eq. (9) corresponds to Eq. (57). The numbering should be harmonized across the main text and appendix.
  2. [Sec. 3.1, Eq. (12)] The Data Processing Inequality direction appears typeset incorrectly. The correct inequality is KL(p_phi0 || p0) >= KL((g_theta)_# p_phi0 || (g_theta)_# p0), not the reverse.
  3. [Appendix B.2] The text says 'we consider SD-Turbo and SANA-Sprint as our two base models,' although FLUX-Schnell is also used in the experiments. This wording should be updated.
  4. [Table 1] The timing comparison should state explicitly whether the reported times include the HyperNoise forward pass and whether ReNO/Best-of-N times include reward-model evaluations. This would make the 33x-300x speedup claims easier to verify.
  5. [Section 4.2] The cross-reference 'In Tab. 8' appears to point to the appendix table; the corresponding main-text table is Table 2. Please ensure cross-references are correct.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the noise-space objective is derived from the tilted distribution through explicit change-of-variables and Stein-lemma steps, and the independent benchmark evaluations do not reduce to the training rewards.

full rationale

The paper's central derivation is self-contained rather than circular. The tilted noise distribution p*_0 in Eq. (5) is defined and then proven to push forward to the tilted output distribution p* (Theorem 3, Part 1); the KL decomposition in Proposition 4 follows from the definition of p*_0. The tractable objective in Eq. (11) is obtained by an explicit change of variables, Stein's lemma, and a stated log-determinant bound (Appendix A.4, Theorem 7). The L2 proxy for KL(p_phi0 || p0) is an approximation whose validity depends on the Lipschitz constant L < 1, and the paper explicitly acknowledges that this condition is not enforced during training: 'While we do not explicitly enforce L < 1 during training, these practical measures help maintain f_phi in a regime where our approximation remains accurate throughout the optimization process.' That is an unverified assumption and a correctness risk, but it is not a circular step: the L2 term is not defined in terms of the prediction target, nor is any equation forced by construction. The reward ensemble is inherited from the authors' own ReNO paper (self-citation), but it is used only as the training reward signal and as a comparison baseline; the GenEval, DPG-Bench, and T2I-CompBench evaluations use disjoint prompts and are not fitted or defined by the training rewards. No fitted parameter is renamed as a prediction; the noise hypernetwork's output is the trained amortized optimizer, which is standard amortized inference rather than circularity. Overall, the derivation chain does not reduce to its inputs.

Assumptions & free parameters 1 free parameters · 6 assumptions · 1 invented entities

The central practical result depends on four unverified assumptions: global diffeomorphism, L<1, representability of the tilted distribution, and Stein regularity. The only hand-set scalar in the objective is alpha=1; reward ensemble weights are inherited from ReNO without retuning.

free parameters (1)
  • Reward temperature alpha = 1 (fixed, not tuned)
    Controls the strength of the tilt in Equations 5 and 11. The paper sets alpha=1 and does not tune it; the balance between reward and KL is otherwise fixed by the 0.5 L2 coefficient.
assumptions (6)
  • domain assumption The initial noise distribution of the distilled generator is standard Gaussian N(0,I).
    Used throughout the KL derivation and valid by construction for the tested diffusion models.
  • standard math The generator g_theta is measurable and preimages have a well-defined measure structure.
    Standing assumption in Appendix A.1 used to define pushforward densities and the tilted distribution.
  • ad hoc to paper The perturbation map id+f_phi is a global diffeomorphism.
    Assumption 1 in Appendix A.4 is needed for the change-of-variables KL calculation, but LoRA training does not guarantee invertibility.
  • ad hoc to paper The Lipschitz constant L of f_phi is below 1.
    Theorem 1 requires L<1 for the log-determinant term to be dropped. Appendix A.4 admits that L<1 is not explicitly enforced during training.
  • ad hoc to paper The pushforward family (id+f_phi)#p0 can represent the tilted noise distribution p0*.
    No expressivity or convergence proof is given; the method optimizes within this family and assumes it is close enough to the target.
  • standard math Stein's lemma regularity conditions hold for f_phi.
    Needs finite moments and continuous differentiability for the integration-by-parts step in the KL derivation.
invented entities (1)
  • Noise hypernetwork f_phi (LoRA residual on base generator) independent evidence
    purpose: Transforms standard Gaussian noise into a reward-optimized initial noise for the frozen generator.
    This is a trainable network rather than an unobserved physical entity. Its effect is directly measured in GenEval, T2I-CompBench, DPG-Bench, and qualitative outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models." pith.science (2026). https://pith.science/paper/ZQXS6O3X

@misc{pith2026250809968,
  author       = {Pith},
  title        = {Pith review of: Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZQXS6O3X}},
  note         = {Machine review of arXiv:2508.09968}
}
read the original abstract

The new paradigm of test-time scaling has yielded remarkable breakthroughs in Large Language Models (LLMs) (e.g. reasoning models) and in generative vision models, allowing models to allocate additional computation during inference to effectively tackle increasingly complex problems. Despite the improvements of this approach, an important limitation emerges: the substantial increase in computation time makes the process slow and impractical for many applications. Given the success of this paradigm and its growing usage, we seek to preserve its benefits while eschewing the inference overhead. In this work we propose one solution to the critical problem of integrating test-time scaling knowledge into a model during post-training. Specifically, we replace reward guided test-time noise optimization in diffusion models with a Noise Hypernetwork that modulates initial input noise. We propose a theoretically grounded framework for learning this reward-tilted distribution for distilled generators, through a tractable noise-space objective that maintains fidelity to the base model while optimizing for desired characteristics. We show that our approach recovers a substantial portion of the quality gains from explicit test-time optimization at a fraction of the computational cost. Code is available at https://github.com/ExplainableML/HyperNoise

Discussion (0). Sign in to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Don't Settle at the Mode! Mitigating Diversity Collapse in Pretrained Flow Models via Feature Self-Guidance

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    Feature self-guidance disperses internal features of flow models during batch generation and applies manifold regularization to increase output diversity while preserving condition alignment.

  2. FASTER: Value-Guided Sampling for Fast RL

    cs.LG 2026-04 unverdicted novelty 6.0 of 10

    FASTER models multi-candidate denoising as an MDP and trains a value function to filter actions early, delivering the performance of full sampling at lower cost in diffusion RL policies.

  3. Na\"ive PAINE: Lightweight Text-to-Image Generation Improvement with Prompt Evaluation

    cs.CV 2026-03 conditional novelty 6.0 of 10

    A lightweight predictor ranks initial noises by expected human-preference score for a prompt, selecting the best few for diffusion generation and reporting prompt difficulty.

  4. Distribution Matching Distillation Meets Reinforcement Learning

    cs.CV 2025-11 conditional novelty 6.0 of 10

    Combining DMD distillation with RL during training produces few-step text-to-image models that outperform their multi-step teacher on several benchmarks.

  5. Lagrangian Perturbation Diffusion Steering: Latent Reinforcement Learning for Generative Policies

    cs.LG 2026-05 unverdicted novelty 5.0 of 10

    LP-DS improves generative policies for imitation and RL by optimizing latent noise perturbations with a constrained Lagrangian objective, showing up to 25% better returns on manipulation and locomotion tasks.

  6. Can We Predict The Human Preference For Text-to-Image Content Prior To Generation And Is It Even Useful To Do So?

    cs.CV 2026-06 unverdicted novelty 3.0 of 10

    Exploration of pre-generation prediction of human preference metrics (HPM) from noise seeds in diffusion models to improve output quality with negligible added cost.

Reference graph

Works this paper leans on

119 extracted references · 69 canonical work pages · cited by 6 Pith papers

  1. [1]

    A noise is worth diffusion guidance

    Donghoon Ahn, Jiwon Kang, Sanghyun Lee, Jaewon Min, Minjae Kim, Wooseok Jang, Hyoungwon Cho, Sayak Paul, SeonHwa Kim, Eunju Cha, et al. A noise is worth diffusion guidance. ����� �������� ����������������, 2024

    2024

  2. [2]

    Hyperstyle: Stylegan inversion with hypernetworks for real image editing

    Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In ���� , 2022

    2022

  3. [3]

    Building normalizing flows with stochastic interpolants

    Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ���� , 2023

    2023

  4. [4]

    Llms can see and hear without any training, 2025

    Kumar Ashutosh, Yossi Gandelsman, Xinlei Chen, Ishan Misra, and Rohit Girdhar. Llms can see and hear without any training, 2025. URL https://arxiv.org/abs/2501.18096

    arXiv 2025

  5. [5]

    Naesseth

    Grigory Bartosh, Dmitry Vetrov, and Christian A. Naesseth. Neural flow diffusion models: Learnable forward process for improved diffusion modelling, 2025. URL https://arxiv. org/abs/2404.12940

    arXiv 2025

  6. [6]

    D-flow: Differentiating through flows for controlled generation

    Heli Ben-Hamu, Omri Puny, Itai Gat, Brian Karrer, Uriel Singer, and Yaron Lipman. D-flow: Differentiating through flows for controlled generation. In ���� , 2024

    2024

  7. [7]

    Lowering pytorch’s memory consumption for selective differentiation

    Samarth Bhatia and Felix Dangel. Lowering pytorch’s memory consumption for selective differentiation. 2024

    2024

  8. [8]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ���� , 2024

    2024

Show all 119 references
  1. [9]

    Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. In �������� , 2023

    2023

  2. [10]

    Enhancing diffusion models with text-encoder reinforcement learning

    Chaofeng Chen, Annan Wang, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Enhancing diffusion models with text-encoder reinforcement learning. In ���� , 2024

    2024

  3. [11]

    Sana-sprint: One-step diffusion with continuous-time consistency distillation

    Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Enze Xie, and Song Han. Sana-sprint: One-step diffusion with continuous-time consistency distillation. ����� �������� ����������������, 2025

    2025

  4. [12]

    Directly fine-tuning diffusion models on differentiable rewards

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ���� , 2024

    2024

  5. [13]

    Cover and Joy A

    Thomas M. Cover and Joy A. Thomas. �������� �� ����������� ������ ��� ������� ������ ������ �� ������������������ ��� ������ �����������. Wiley-Interscience, July 2006. ISBN 0471241954

    2006

  6. [14]

    Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models

    Fei Deng, Qifei Wang, Wei Wei, Matthias Grundmann, and Tingbo Hou. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In ���� , 2024

    2024

  7. [15]

    Adjoint matching: Fine-tuning flow and diffusion generative models with memoryless stochastic optimal control

    Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky TQ Chen. Adjoint matching: Fine-tuning flow and diffusion generative models with memoryless stochastic optimal control. ����� �������� ����������������, 2024

    2024

  8. [16]

    Hyperdiffusion: Generating implicit neural fields with weight-space diffusion

    Ziya Erkoç, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Hyperdiffusion: Generating implicit neural fields with weight-space diffusion. In ���� , 2023

    2023

  9. [17]

    Scaling rectified flow trans- formers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow trans- formers for high-resolution image synthesis. ����� �������� ����������������, 2024

    2024

  10. [18]

    Reno: Enhancing one-step text-to-image models through reward-based noise optimization

    Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. In �������, 2024. 11

    2024

  11. [19]

    Unbalancedness in neural monge maps improves unpaired domain translation

    Luca Eyring, Dominik Klein, Théo Uscidda, Giovanni Palla, Niki Kilbertus, Zeynep Akata, and Fabian J Theis. Unbalancedness in neural monge maps improves unpaired domain translation. In ��� ������� ������������� ���������� �� �������� ���������������, 2024. URL https://openrevi...

    2024

  12. [20]

    Reinforcement learning for fine-tuning text-to-image diffusion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. �������, 2023

    2023

  13. [21]

    Training-free structured diffusion guidance for compositional text-to-image synthesis

    Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In ���� , 2023

    2023

  14. [22]

    Geneval: An object-focused frame- work for evaluating text-to-image alignment

    Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused frame- work for evaluating text-to-image alignment. In �������, 2023

    2023

  15. [23]

    Priorgrad: Improving conditional denoising diffusion models with data-dependent adaptive prior, 2022

    Sang gil Lee, Heeseung Kim, Chaehun Shin, Xu Tan, Chang Liu, Qi Meng, Tao Qin, Wei Chen, Sungroh Yoon, and Tie-Yan Liu. Priorgrad: Improving conditional denoising diffusion models with data-dependent adaptive prior, 2022. URL https://arxiv.org/abs/2106.06406

    2022 arXiv

  16. [24]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. ����� �������� ����������������, 2025

    2025

  17. [25]

    Initno: Boosting text-to-image diffusion models via initial noise optimization

    Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In ���� , 2024

    2024

  18. [26]

    Hypernetworks

    David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. ����� �������� ����������������, 2016

    2016

  19. [27]

    Hypernet fields: Efficiently training hypernetworks without ground truth by learning weight trajectories

    Eric Hedlin, Munawar Hayat, Fatih Porikli, Kwang Moo Yi, and Shweta Mahajan. Hypernet fields: Efficiently training hypernetworks without ground truth by learning weight trajectories. ����� �������� ����������������, 2024

    2024

  20. [28]

    Clipscore: A reference-free evaluation metric for image captioning, 2022

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022

    2022

  21. [29]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In �������, 2020

    2020

  22. [30]

    Margin- aware preference optimization for aligning diffusion models without reference

    Jiwoo Hong, Sayak Paul, Noah Lee, Kashif Rasul, James Thorne, and Jongheon Jeong. Margin- aware preference optimization for aligning diffusion models without reference. ����� �������� ����������������, 2024

    2024

  23. [31]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ���� , 2022

    2022

  24. [32]

    Ella: Equip diffusion models with llm for enhanced semantic alignment

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. ����� �������� ����������������, 2024

    2024

  25. [33]

    T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In�������, 2023

    2023

  26. [34]

    Openclip, July 2021

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. URL https://doi.org/10.5281/ zenodo.5143773

    2021

  27. [35]

    Hint: hypernetwork instruction tuning for efficient zero-& few-shot generalisation

    Hamish Ivison, Akshita Bhagia, Yizhong Wang, Hannaneh Hajishirzi, and Matthew Peters. Hint: hypernetwork instruction tuning for efficient zero-& few-shot generalisation. ����� �������� ����������������, 2022

    2022

  28. [36]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. ����� �������� ����������������, 2024

    2024

  29. [37]

    Elucidating optimal reward-diversity tradeoffs in text-to-image diffusion models

    Rohit Jena, Ali Taghibakhshi, Sahil Jain, Gerald Shen, Nima Tajbakhsh, and Arash Vahdat. Elucidating optimal reward-diversity tradeoffs in text-to-image diffusion models. ����� �������� ����������������, 2024. 12

    2024

  30. [38]

    Reward fine-tuning two-step diffusion models via learning differentiable latent-space surrogate reward

    Zhiwei Jia, Yuesong Nan, Huixi Zhao, and Gengdai Liu. Reward fine-tuning two-step diffusion models via learning differentiable latent-space surrogate reward. ����� �������� ����������������, 2024

    2024

  31. [39]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In �������, 2022

    2022

  32. [40]

    If at first you don’t succeed, try, try again: Faithful diffusion-based text-to-image generation by selection

    Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. If at first you don’t succeed, try, try again: Faithful diffusion-based text-to-image generation by selection. ����� �������� ����������������, 2023

    2023

  33. [41]

    Scalable ranked preference optimization for text-to-image generation

    Shyamgopal Karthik, Huseyin Coskun, Zeynep Akata, Sergey Tulyakov, Jian Ren, and Anil Kag. Scalable ranked preference optimization for text-to-image generation. ����� �������� ����������������, 2024

    2024

  34. [42]

    Optimizing diffusion noise can serve as universal motion priors

    Korrawe Karunratanakul, Konpat Preechakul, Emre Aksan, Thabo Beeler, Supasorn Suwa- janakorn, and Siyu Tang. Optimizing diffusion noise can serve as universal motion priors. In ���� , 2024

    2024

  35. [43]

    Kingma, Tim Salimans, Ben Poole, and Jonathan Ho

    Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In �������, 2021

    2021

  36. [44]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In �������, 2023

    2023

  37. [45]

    T �" ulu 3: Pushing frontiers in open language model post-training

    Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T �" ulu 3: Pushing frontiers in open language model post-training. ����� �������� ����������������, 2024

    2024

  38. [46]

    Aligning text-to-image models using human feedback

    Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. ����� �������� ����������������, 2023

    2023

  39. [47]

    Blip: Bootstrapping language- image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language- image pre-training for unified vision-language understanding and generation. In ������������� ���������� �� ������� ��������, 2022

    2022

  40. [48]

    Aligning diffusion models by optimizing human utility

    Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. ����� �������� ����������������, 2024

    2024

  41. [49]

    Textcraftor: Your text encoder can be image quality controller

    Yanyu Li, Xian Liu, Anil Kag, Ju Hu, Yerlan Idelbayev, Dhritiman Sagar, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Textcraftor: Your text encoder can be image quality controller. In ���� , 2024

    2024

  42. [50]

    Evaluating text-to-visual generation with image-to-text generation

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. ����� �������� ����������������, 2024

    2024

  43. [51]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. In ���� , 2023

    2023

  44. [52]

    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. ����� �������� ����������������, 2022

    2022

  45. [53]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In����������� �� ��� �������� ������������� ���������� �� �������� ������ ������, 2021

    2021

  46. [54]

    Latent consistency models: Synthesizing high-resolution images with few-step inference.����� �������� ����������������, 2023

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

    2023

  47. [55]

    Inference-time scaling for diffusion models beyond scaling denoising steps

    Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps. ����� �������� ����������������, 2025

    2025

  48. [56]

    Let’s reward step by step: Step-level reward model as the navigators for reasoning

    Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Yang You, and Hongxia Yang. Let’s reward step by step: Step-level reward model as the navigators for reasoning. ����� �������� ����������������, 2023. 13

    2023

  49. [57]

    Improving text-to-image consistency via automatic prompt optimization, 2024

    Oscar Mañas, Pietro Astolfi, Melissa Hall, Candace Ross, Jack Urbanek, Adina Williams, Aishwarya Agrawal, Adriana Romero-Soriano, and Michal Drozdzal. Improving text-to-image consistency via automatic prompt optimization, 2024. URL https://arxiv.org/abs/ 2403.17804

    2024 arXiv

  50. [58]

    Tuning timestep-distilled diffusion model using pairwise sample optimization

    Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. ����� �������� ����������������, 2024

    2024

  51. [59]

    Learning to compress prompts with gist tokens

    Jesse Mu, Xiang Li, and Noah Goodman. Learning to compress prompts with gist tokens. �������� �� ������ ����������� ���������� �������, 36:19327–19352, 2023

    2023

  52. [60]

    s1: Simple test-time scaling

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. ����� �������� ����������������, 2025

    2025

  53. [61]

    Ditto- 2: Distilled diffusion inference-time t-optimization for music generation

    Zachary Novack, Julian McAuley, Taylor Berg-Kirkpatrick, and Nicholas Bryan. Ditto- 2: Distilled diffusion inference-time t-optimization for music generation. ����� �������� ����������������, 2024

    2024

  54. [62]

    Zachary Novack, Julian McAuley, Taylor Berg-Kirkpatrick, and Nicholas J. Bryan. Ditto: Diffusion inference-time t-optimization for music generation, 2024. URL https://arxiv. org/abs/2401.12179

    2024 arXiv

  55. [63]

    Rl for consis- tency models: Faster reward guided text-to-image generation.����� �������� ����������������, 2024

    Owen Oertell, Jonathan D Chang, Yiyi Zhang, Kianté Brantley, and Wen Sun. Rl for consis- tency models: Faster reward guided text-to-image generation.����� �������� ����������������, 2024

    2024

  56. [64]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. ����� �������� ����������������, 2023

    2023

  57. [65]

    Normalizing flows for probabilistic modeling and inference, 2021

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference, 2021. URL https://arxiv.org/abs/1912.02762

    2021 arXiv

  58. [66]

    Benchmark for compositional text-to-image synthesis

    Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In ������� �������� ��� ���������� �����, 2021

    2021

  59. [67]

    Hypertuning: Toward adapting large language models without back-propagation

    Jason Phang, Yi Mao, Pengcheng He, and Weizhu Chen. Hypertuning: Toward adapting large language models without back-propagation. In ���� , pages 27854–27875, 2023

    2023

  60. [68]

    Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023

    2023

  61. [69]

    Aligning text- to-image diffusion models with reward backpropagation

    Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text- to-image diffusion models with reward backpropagation. ����� �������� ����������������, 2023

    2023

  62. [70]

    Video diffusion alignment via reward gradients

    Mihir Prabhudesai, Russell Mendonca, Zheyang Qin, Katerina Fragkiadaki, and Deepak Pathak. Video diffusion alignment via reward gradients. ����� �������� ����������������, 2024

    2024

  63. [71]

    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, et al. Learning transferable visual models from natural language supervision. In ���� , 2021

    2021

  64. [72]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. �������, 2023

    2023

  65. [73]

    Hyper-sd: Trajectory segmented consistency model for efficient image synthesis, 2024

    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, 2024

    2024

  66. [74]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In ���� , 2022

    2022

  67. [75]

    Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control

    Litu Rout, Yujia Chen, Nataniel Ruiz, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control. ����� �������� ����������������, 2024. 14

    2024

  68. [76]

    Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wad- hwa, Michael Rubinstein, and Kfir Aberman. Hyperdreambooth: Hypernetworks for fast personalization of text-to-image models. In ���� , 2024

    2024

  69. [77]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. ����� �������� ����������������, 2023

    2023

  70. [78]

    Scaling llm test-time compute opti- mally can be more effective than scaling model parameters

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling llm test-time compute opti- mally can be more effective than scaling model parameters. ����� �������� ����������������, 2024

    2024

  71. [79]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ���� , 2021

    2021

  72. [80]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ���� , 2021

    2021

  73. [81]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ���� , 2023

    2023

  74. [82]

    Cocono: Attention contrast-and-complete for initial noise optimization in text-to- image synthesis

    Aravindan Sundaram, Ujjayan Pal, Abhimanyu Chauhan, Aishwarya Agarwal, and Srikrishna Karanam. Cocono: Attention contrast-and-complete for initial noise optimization in text-to- image synthesis. ����� �������� ����������������, 2024

    2024

  75. [83]

    Fine-tuning of diffusion models via stochastic control: entropy regularization and beyond, 2024

    Wenpin Tang. Fine-tuning of diffusion models via stochastic control: entropy regularization and beyond, 2024. URL https://arxiv.org/abs/2403.06279

    2024

  76. [84]

    Inference-time alignment of diffusion models with direct noise optimization

    Zhiwei Tang, Jiangweizhi Peng, Jiasheng Tang, Mingyi Hong, Fan Wang, and Tsung-Hui Chang. Inference-time alignment of diffusion models with direct noise optimization. ����� �������� ����������������, 2024

    2024

  77. [85]

    Fine- tuning of continuous-time diffusion models as entropy-regularized control, 2024

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Fine- tuning of continuous-time diffusion models as entropy-regularized control, 2024. URL https://arxiv.org/abs/2402.15194

    2024 arXiv

  78. [86]

    Reward-guided iterative refinement in diffusion models at test-time with applications to protein and dna design, 2025

    Masatoshi Uehara, Xingyu Su, Yulai Zhao, Xiner Li, Aviv Regev, Shuiwang Ji, Sergey Levine, and Tommaso Biancalani. Reward-guided iterative refinement in diffusion models at test-time with applications to protein and dna design, 2025. URL https://arxiv.org/abs/2502. 14944

    2025

  79. [87]

    Inference-time alignment in diffusion models with reward-guided generation: Tutorial and review, 2025

    Masatoshi Uehara, Yulai Zhao, Chenyu Wang, Xiner Li, Aviv Regev, Sergey Levine, and Tommaso Biancalani. Inference-time alignment in diffusion models with reward-guided generation: Tutorial and review, 2025. URL https://arxiv.org/abs/2501.09685

    2025 arXiv

  80. [88]

    Outsourced diffusion sampling: Efficient posterior inference in latent spaces of generative models

    Siddarth Venkatraman, Mohsin Hasan, Minsu Kim, Luca Scimeca, Marcin Sendera, Yoshua Bengio, Glen Berseth, and Nikolay Malkin. Outsourced diffusion sampling: Efficient posterior inference in latent spaces of generative models. ����� �������� ����������������, 2025

    2025

  81. [89]

    Continual learning with hypernetworks

    Johannes V on Oswald, Christian Henning, Benjamin F Grewe, and João Sacramento. Continual learning with hypernetworks. ����� �������� ����������������, 2019

    2019

  82. [90]

    Steering your diffusion policy with latent space reinforcement learning, 2025

    Andrew Wagenmaker, Mitsuhiko Nakamoto, Yunchu Zhang, Seohong Park, Waleed Yagoub, Anusha Nagabandi, Abhishek Gupta, and Sergey Levine. Steering your diffusion policy with latent space reinforcement learning, 2025. URL https://arxiv.org/abs/2506.15799

    2025 arXiv

  83. [91]

    End-to-end diffusion latent optimization improves classifier guidance

    Bram Wallace, Akash Gokul, Stefano Ermon, and Nikhil Naik. End-to-end diffusion latent optimization improves classifier guidance. In ���� , 2023

    2023

  84. [92]

    Diffusion model alignment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In ���� , 2024

    2024

  85. [93]

    Neural network diffusion

    Kai Wang, Zhaopan Xu, Yukun Zhou, Zelin Zang, Trevor Darrell, Zhuang Liu, and Yang You. Neural network diffusion. ����� �������� ����������������, 2024

    2024

  86. [94]

    Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models

    Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. ����� �������� ����������������, 2022. 15

    2022

  87. [95]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

    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. ����� �������� ����������������, 2023

    2023

  88. [96]

    Better aligning text-to- image models with human preference

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Better aligning text-to- image models with human preference. In ���� , 2023

    2023

  89. [97]

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

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In �������, 2023

    2023

  90. [98]

    Long-clip: Unlocking the long-text capability of clip

    Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In �������� ���������� �� �������� ������, pages 310–325. Springer, 2024

    2024

  91. [99]

    Graph hypernetworks for neural architecture search

    Chris Zhang, Mengye Ren, and Raquel Urtasun. Graph hypernetworks for neural architecture search. ����� �������� ����������������, 2018

    2018

  92. [100]

    The unreason- able effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreason- able effectiveness of deep features as a perceptual metric. In ���� , 2018

    2018

  93. [101]

    Learning multi-dimensional human preference for text-to-image generation

    Sixian Zhang, Bohan Wang, Junqiang Wu, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Learning multi-dimensional human preference for text-to-image generation. In ���� , 2024

    2024

  94. [102]

    Large-scale reinforcement learning for diffusion models

    Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. In ���� , 2024

    2024

  95. [103]

    The lessons of developing process reward models in mathematical reasoning

    Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning. ����� �������� ����������������, 2025

    2025

  96. [104]

    Golden noise for diffusion models: A learning framework

    Zikai Zhou, Shitong Shao, Lichen Bai, Zhiqiang Xu, Bo Han, and Zeke Xie. Golden noise for diffusion models: A learning framework. ����� �������� ����������������, 2024. 16 Appendix The Appendix is organized as follows: • Section A provides all of our theoretical derivations. •...

    2024

  97. [105]

    The generator �θ �� d � � d is measurable

  98. [106]

    The reward function � � � d � � is measurable and � � � ∼p� ��r(g� (� � ))/α� � � for our chosen temperature � � �

  99. [107]

    These assumptions are mild and realistic for neural network generators

    For any � � Range��θ�, the preimage set �−1 θ ����� has a well-defined measure structure. These assumptions are mild and realistic for neural network generators. Pushforward Measure and Base Distribution. The base generator density �base��� is the density of the pushforward me...

  100. [108]

    Intractable KL term: Computing � KL��ϕ��base� requires evaluating densities of high- dimensional neural network generators, which involves intractable Jacobian determinants

  101. [109]

    No continuous-time structure: Unlike full diffusion models, distilled generators often lack explicit SDE/ODE structure that would enable techniques from stochastic optimal control 18

  102. [110]

    A.3 The Reward-Tilted Noise Distribution An alternative to modifying the generator�θ is to modify the input noise density�0�� 0�while keeping �θ fixed

    Reward hacking: Without proper regularization, optimization can lead to adversarial exploitation of the reward model, generating unrealistic samples that achieve high reward scores These challenges motivate our alternative approach of modifying the input noise distribution whi...

  103. [111]

    Our direct variational approach and the general stochastic control theory yield the same optimal noise distribution

  104. [112]

    This equivalence arises because for one-step generators, the continuous-time framework naturally collapses to our setting, with their value function �⋆ 0 simplifying to the composed reward � � �θ

  105. [113]

    While both approaches are mathematically equivalent here, our proof provides a more elementary and direct path to the solution, sidestepping the complex machinery of stochastic control. This connection not only validates our result but also situates it as an important special ...

  106. [114]

    �ϕ is continuously differentiable

  107. [115]

    � ϕ is a global diffeomorphism (invertible with continuous derivatives)

  108. [116]

    While Assumption 1 requires � ϕ to be a global diffeomorphism, we provide a practical sufficient condition: Lemma 5 (Lipschitz Condition for Invertibility)

    �ϕ satisfies the regularity conditions for Stein’s lemma: � ���ϕ�� 0��2� � � and � ��� 0���ϕ�� 0���� � for � 0 � � ��� �� Sufficient Condition for Global Diffeomorphism. While Assumption 1 requires � ϕ to be a global diffeomorphism, we provide a practical sufficient condition:...

  109. [117]

    Initialization: Setting �ϕ��� � � ensures ��� � � � initially

  110. [118]

    A photo of a parrot

    Regularization: The term 1 2��ϕ�� 0��2 naturally penalizes large perturbations, helping maintain small eigenvalues of �f� While we do not explicitly enforce � � � during training, these practical measures help maintain �ϕ in a regime where our approximation remains accurate th...

  111. [119]

    reward-hacking

    for evaluating the composi- tional capabilities of text-to-image generation models. We evaluate on the Attribute binding tasks, which includes color, shape, and texture sub-categories, where the model should bind the attributes with the correct objects to generate the complex ...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.