Pith. sign in

REVIEW 3 major objections 5 minor 84 references

Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Inversion-DPO replaces the intractable posterior in diffusion preference alignment with a deterministic DDIM-inversion trajectory, yielding a two-term loss that trains more than twice as fast and improves PickScore, CLIP, and Aesthetic…

desk verdict Real dataset, timely idea, but the central DPO derivation has a sign error that flips the objective; the paper needs major revision. read the letter →

arxiv 2507.11554 v4 pith:HXK2XIZC submitted 2025-07-14 cs.CV cs.AI

Zejian Li , Yize Li , Chenye Meng , Zhongni Liu , Yang Ling , Shengyuan Zhang , Guang Yang , Changyuan Yang
show 2 more authors
Zhiyuan Yang Lingyun Sun
This is my paper
classification cs.CVcs.AI
keywords diffusionmodelsdirectpreferenceoptimizationDDIMinversionhumanalignmenttext-to-imagegenerationcompositionalimagepost-trainingPickScore
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 proposes Inversion-DPO, a post-training method that aligns diffusion image generators with human preferences without training a separate reward model. Its central move is to replace the intractable posterior over noisy trajectories, p_theta(x_{1:T}|x0), which prior Diffusion-DPO approximates with the forward noising process, by the deterministic trajectory recovered through DDIM inversion. This collapses the DPO objective from four KL-divergence terms to two squared-difference terms between the trained noise predictor and the frozen reference predictor. The authors claim this removes the distributional mismatch, gives more than 2x faster convergence on text-to-image generation, and produces state-of-the-art PickScore, CLIP, and Aesthetic scores, including on a newly curated 11,140-image paired dataset for compositional generation.

What carries the argument

The load-bearing object is the DDIM inversion trajectory: a deterministic map from a clean image x0 back to a noise sequence x_{1:T} via the reference model's epsilon predictions. The argument uses the inversion trajectory as a stand-in for the intractable posterior p_theta(x_{1:T}|x0), which lets the DPO log-ratio decompose into a single sum of squared epsilon-prediction differences rather than four KL terms. The claim is that this trajectory is precise enough for the log-probability terms involving p_theta to be treated as constants, so only the reference-model terms remain.

What would settle it

Take a checkpoint trained with Inversion-DPO, recompute DDIM inversion with that checkpoint rather than the frozen reference, and compare the resulting trajectories to the original ones; if they diverge significantly, the assumption that x_{t-1} is common to both distributions is violated and the two-term loss is no longer the DPO log-ratio.

Watch

Extended reading notes

Core claim

The paper's central claim is that preference alignment for diffusion models can be derived from a deterministic inversion trajectory rather than from stochastic forward-process sampling. Given a winning and a losing image, DDIM inversion under the frozen reference model produces paired noise sequences; along those sequences the log-probability ratio in DPO reduces to a sum over timesteps of ||eps_theta(x_t)-eps_theta0(x_t)||^2 differences. The resulting loss is L = -E log $\sigma$($\beta$ sum_t [ ||eps_theta(x^w_t,t)-eps_theta0(x^w_t,t)||^2 - ||eps_theta(x^l_t,t)-eps_theta0(x^l_t,t)||^2 ]). The authors argue this is more precise than Diffusion-DPO's forward-process approximation and that the simplification is what yields faster training and the reported improvements.

Load-bearing premise

The derivation depends on the assumption that the noise trajectory DDIM inversion recovers with the frozen reference model is the same trajectory the model being trained would follow, so that the current model's conditional log-probabilities can be treated as constant; this approximation is only as good as that trajectory match and degrades as training moves the model away from the reference.

Editorial extensions

If this is right

  • Text-to-image models post-trained with Inversion-DPO improve PickScore, CLIP Score, and Aesthetic Score over SDXL, Diffusion-DPO, DDPO, D3PO, Demon, and IterComp in the reported tables.
  • The two-term loss reaches the PickScore of a 2000-step Diffusion-DPO run in about 1000 training steps, and with 80 inversion steps converges more than 4x faster, per the training-curve plots.
  • On compositional generation, Inversion-DPO fine-tuned on the new paired LAION-SG-derived dataset improves FID, SG-IoU, Entity-IoU, and Relation-IoU over SGDiff, SG-Adapter, R3CD, and SDXL-SG.
  • Inversion-DPO can be applied on top of both SDXL and SD1.5, and its derivation removes the need for a separately trained reward model.

Reading between the lines

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

  • If the trajectory substitution holds, the same inversion trick could be applied to other preference objectives (e.g., IPO or KTO) and to flow-matching or consistency models, wherever a deterministic inverse trajectory exists; the paper does not explore this.
  • The two-term matching loss resembles a score-distillation or consistency regularizer; one testable extension is whether it can be combined with prompt-aware weighting of timesteps to avoid over-emphasizing early or late steps.
  • The reported efficiency gain depends partly on drawing full trajectories in advance; a practical extension would be to cache or distill inversion trajectories to avoid two inference passes per sample, a cost the paper's own limitation section acknowledges.
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

3 major / 5 minor

Summary. This paper proposes Inversion-DPO, a post-training alignment method for diffusion models. It replaces the intractable posterior p_theta(x_{1:T}|x_0) in Diffusion-DPO with a deterministic DDIM inversion trajectory computed from the frozen reference model, and derives a DPO-style loss (Eq. 15) that compares squared epsilon-prediction differences between the current and reference models on winning and losing trajectories. The authors claim this removes the distributional mismatch of Diffusion-DPO's forward-process approximation, yields over 2x faster convergence, and achieves state-of-the-art PickScore/CLIP/Aesthetic scores on text-to-image generation and compositional generation, supported by experiments on SDXL and SD1.5 and a newly curated 11,140-image compositional dataset.

Significance. If the derivation were correct, eliminating reward-model training and reducing the DPO loss to two trajectory-matching terms would be a practical contribution, and the reported gains in convergence and metrics would be of interest to the multimedia and diffusion-model community. The paper also contributes a compositional paired dataset and a user study. However, the central derivation is not correct as written: the sign of the key log-ratio term is reversed, and the claim that the frozen-reference DDIM inversion trajectory is also a trajectory of the current model is asserted rather than established. These issues are load-bearing, so the significance claim cannot be credited in the present form.

major comments (3)
  1. [Sec. 4.1, Eqs. (13)-(15); Supp. S5-S6] The sign of the DPO argument is reversed. For Gaussian conditionals p_theta(x_{t-1}|x_t)=N(mu_theta(x_t), sigma_t^2 I), evaluating the log-density ratio at the inversion point x_{t-1}=mu_theta0(x_t) gives log p_theta - log p_theta0 = -||mu_theta - mu_theta0||^2/(2 sigma_t^2) + const, i.e., a term proportional to -||eps_theta - eps_theta0||^2, the opposite of Eq. (14) and (S7). Inserting this into the DPO template yields L = -E log sigma(-beta(d_w - d_l)), not Eq. (15). As written, Eq. (15) drives d_w - d_l to +infinity, which increases the distance from the reference on winning samples and decreases it on losing samples, directly contradicting the alignment direction stated in Sec. 4.1 and Figure 2. This is a load-bearing error, not a typo, because the claimed preference-learning behavior is the opposite of the objective actually written.
  2. [Sec. 4.1 and Supp. S5, assumption on x^w_{t-1}] The derivation assumes that x^w_{t-1} subjects to N(mu_theta, sigma_t^2 I), i.e., that the DDIM-inverted trajectory from the frozen reference model is also a trajectory of the current model. Once theta differs from theta0, Eq. (9) fixes x_{t-1} from eps_theta0, not from eps_theta, so the assumption is exactly what needs to be shown. The claim that log p_theta(x_{t-1}|x_t) is constant in theta depends on this equality, and Supp. S4's visual similarity is not a quantitative validation. Supp. S8 concedes that the assumption is strong and its error impact unclear. As it stands, the 'precise posterior sampling' claim is circular: the posterior p_theta(x_{1:T}|x_0) is effectively defined to be the deterministic inversion path, and then the derivation assumes that path is generated by p_theta. The paper needs either a proof under stated conditions or a controlled experiment measuring how the inversion and sampling trajectories diverge as theta moves away from theta0.
  3. [Eq. (13) and Supp. Eqs. (S5)-(S6)] The displayed chain is internally inconsistent. The first proportionality says -log p_theta0(x_{t-1}|x_t) is proportional to ||mu_theta0(x_t)-x_{t-1}||^2, but x_{t-1} is obtained from the DDIM inversion iteration using eps_theta0, so mu_theta0(x_t)-x_{t-1}=0 in the deterministic setting (or is controlled by sigma_t in the stochastic setting). The second proportionality to ||eps_theta0 - eps_theta||^2 therefore does not follow. The derivation must specify which point x_{t-1} is being evaluated and which density is being compared; currently the equation conflates the inversion trajectory and the current model's mean.
minor comments (5)
  1. [Abstract and Sec. 2] There are several typos, including 'appromixation', 'compostitional image geneation', and 'Out attempt' in Sec. 2.
  2. [Eq. (15)] The expectation is written over x_{1:T} ~ p_theta(x_{1:T}|x_0), but p_theta is intractable and the method actually uses a deterministic inversion; the notation should be a deterministic function of x_0 or a delta distribution, not a p_theta expectation.
  3. [Figs. 3-4 and Sec. 5.4] The '2x faster' claim compares 20-step Inversion-DPO at 1000 steps to the authors' Diffusion-DPO retraining at 2000 steps, but the retrained Diffusion-DPO does not reach the official checkpoint; please clarify the comparison and report variance or confidence intervals.
  4. [Table S2 vs. Table 1] Table S2 reports PickScore 0.228 for Inversion-DPO while Table 1 reports 0.232; please explain whether these are different checkpoints or evaluation settings.
  5. [Supp. S1] The user study has only 10 participants; please state whether the 66.53% and 72.67% figures are aggregate across all 500 trials and report participant-level variance.

Circularity Check

1 steps flagged · score 6.0 of 10

Central derivation assumes the frozen-reference DDIM trajectory is a p_theta trajectory, so the claimed log-ratio is imposed by construction.

  1. self definitional [Sec. 4.1 and Supp. S5; Eq. 9 vs. Eqs. 14/S7]
    "Since x^w_{t-1} is equivalently generated by p_theta (i.e., mu_theta) based on the assumption above, we further assume x^w_{t-1} subjects to N(mu_theta(x^w_t), sigma^2_t I)."

    The trajectory x^w_{t-1} is constructed by DDIM inversion with the frozen reference eps_theta0 (Eq. 9), i.e., x^w_{t-1} = mu_theta0(x^w_t). The derivation then evaluates -log p_theta0 at that point and converts the mean difference into ||eps_theta - eps_theta0||^2 (S5-S7) by assuming the same point is drawn from N(mu_theta, sigma_t^2 I). That assumption is exactly the conclusion needed: that the theta0-inversion trajectory is a p_theta trajectory. The equality holds only if x_{t-1}=mu_theta, true by construction only for the reference, not for the current theta. Hence Eq. 14/S7 is imposed by defining the posterior as the theta0-inversion path, not derived from p_theta.

full rationale

The core derivation is partially circular. Sec. 4.1 announces that Inversion-DPO 'eliminate[s] the distributional mismatch' by approximating p_theta(x_{1:T}|x0) with DDIM inversion, and Eq. 9 constructs the trajectory from the frozen eps_theta0. Supp. S5 then asserts that the inverted x^w_{t-1} is 'equivalently generated by p_theta (i.e., mu_theta)' and 'subjects to N(mu_theta, sigma_t^2 I)', which is the very claim that the theta0-inverted path is a theta path. Without this assertion, evaluating the two Gaussian conditionals at the inversion point gives the opposite sign for the log-ratio, so Eqs. 14/S7 are not a consequence of p_theta but an imposition of the defining assumption. The paper's own S8 concedes that the inversion assumption is 'strong' and valid only for sufficiently many steps, which confirms the premise is assumed rather than derived. Empirical claims are not circular: PickScore/CLIP/Aesthetic/FID are external metrics, preference pairs come from Pick-a-Pic or metric-ranked generated images, and no parameter is fitted to the reported test sets. The LAION-SG self-citation [31] supplies data and baselines for the compositional task, not the DPO derivation, so it is not load-bearing for the circularity score. The sign inconsistency of Eq. 15 (it pushes winning-sample epsilon distance up rather than down) is a correctness issue, not a circularity, and is not scored.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or model entities. Its free parameters are standard hyperparameters plus a dropped weighting coefficient. The key conceptual addition is the assumption that DDIM inversion with the reference model provides the exact posterior trajectory for the current model, which is an ad hoc modeling assumption rather than an independently evidenced fact.

free parameters (3)
  • beta (DPO temperature) = 2000
    Set to 2000 in Sec 5.1. Controls the strength of the preference signal and the implicit KL penalty. No sweep or fitting procedure is reported.
  • DDIM inversion steps = 80 (20, 40, 80 explored)
    Ablated in Sec 5.5. Larger steps improve scores, so 80 is chosen. Not derived from first principles.
  • Loss weighting coefficient (1-alpha_t)/(2 sigma_t^2 alpha_t) = dropped (set to 1)
    In Supp S5 the derived coefficient is removed by adopting the DDPM simple loss. This changes the objective and is a hand choice.
assumptions (4)
  • standard math Bradley-Terry preference model and DPO reparameterization (Eq 1-4)
    Used to derive the preference loss. Standard in the DPO literature.
  • standard math The reverse process is Gaussian with mean as in Eq 8
    Standard DDIM/DDPM parameterization of the denoising transitions.
  • domain assumption DDIM inversion assumption: predicted noises are close in adjacent steps (eps_theta(x_t,t) approx eps_theta(x_{t-1},t-1))
    Sec 3.2 and Eq 9. The inversion trajectory is only a valid approximation of the ODE path if this holds, which requires many steps.
  • ad hoc to paper The DDIM-inverted trajectory from the frozen reference model coincides with the trajectory the current model would sample (x^w_{t-1} ~ N(mu_theta, sigma_t^2 I))
    Sec 4.1 and Supp S6. This is the load-bearing assumption that turns the log-ratio into ||eps_theta - eps_theta0||^2. It is not proven and degrades as theta moves away from theta0.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models." pith.science (2026). https://pith.science/paper/HXK2XIZC

@misc{pith2026250711554,
  author       = {Pith},
  title        = {Pith review of: Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HXK2XIZC}},
  note         = {Machine review of arXiv:2507.11554}
}
read the original abstract

Recent advancements in diffusion models (DMs) have been propelled by alignment methods that post-train models to better conform to human preferences. However, these approaches typically require computation-intensive training of a base model and a reward model, which not only incurs substantial computational overhead but may also compromise model accuracy and training efficiency. To address these limitations, we propose Inversion-DPO, a novel alignment framework that circumvents reward modeling by reformulating Direct Preference Optimization (DPO) with DDIM inversion for DMs. Our method conducts intractable posterior sampling in Diffusion-DPO with the deterministic inversion from winning and losing samples to noise and thus derive a new post-training paradigm. This paradigm eliminates the need for auxiliary reward models or inaccurate appromixation, significantly enhancing both precision and efficiency of training. We apply Inversion-DPO to a basic task of text-to-image generation and a challenging task of compositional image generation. Extensive experiments show substantial performance improvements achieved by Inversion-DPO compared to existing post-training methods and highlight the ability of the trained generative models to generate high-fidelity compositionally coherent images. For the post-training of compostitional image geneation, we curate a paired dataset consisting of 11,140 images with complex structural annotations and comprehensive scores, designed to enhance the compositional capabilities of generative models. Inversion-DPO explores a new avenue for efficient, high-precision alignment in diffusion models, advancing their applicability to complex realistic generation tasks. Our code is available at https://github.com/MIGHTYEZ/Inversion-DPO

Figures

Figures reproduced from arXiv: 2507.11554 by the authors.

Figure 1
Figure 1. Visual comparison of different baselines and Inversion-DPO. Each image is annotated with PickScore [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Computation pipeline of the proposed Inversion-DPO. 1) Given an initial image [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. A demonstration of how PickScore increase with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Relationship between training steps and Aesthetics [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: A comparison of generation details. Existing models often struggle to produce fine-grained details, while Inversion [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

84 extracted references · 46 canonical work pages

  1. [1]

    Oron Ashual and Lior Wolf. 2019. Specifying Object Attributes and Relations in Interactive Scene Generation. 4560–4568

    2019

  2. [2]

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, ...

  3. [3]

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. 2024. Training Diffusion Models with Reinforcement Learning. ArXiv abs/2305.13301 (2024)

    arXiv 2024

  4. [4]

    Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. 2018. Coco-stuff: Thing and stuff classes in context. 1209–1218

    2018

  5. [5]

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. 2023. Attend-and-Excite: Attention-Based Semantic Guidance for Text-to-Image Dif- fusion Models. ACM Trans. Graph. 42, 4, Article 148 (July 2023), 10 pages. doi:10.1145/3592116

    doi:10.1145/3592116 2023

  6. [6]

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhong- dao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. 2024. PixArt- Σ: Weak-to- Strong Training of Diffusion Transformer for 4K Text-to-Image Generation.ArXiv abs/2403.04692 (2024). https://api.semanticscholar.org/CorpusID:268264262

    arXiv 2024

  7. [7]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion Models Beat GANs on Image Synthesis. In Advances in Neural Information Processing Systems , M. Ran- zato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan (Eds.), Vol. 34. 8780–8794

    2021

  8. [8]

    Carles Domingo-Enrich, Michal Drozdzal, Brian Karrer, and Ricky T. Q. Chen

    work page

Show all 84 references
  1. [9]

    Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. 2023. Prompt Tuning Inversion for Text-Driven Image Editing Using Diffusion Models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) . 7396–7406. doi:10.1109/ICCV51070.2023.00683

    2023

  2. [10]

    Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. 2024. ReNO: Enhancing One-step Text-to-Image Models through Reward- based Noise Optimization. In Advances in Neural Information Processing Systems , Vol. 37. Curran Associates, Inc., 125487–125519

    2024

  3. [11]

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. 2023. DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models. In Advances in Neural Information Processing Syste...

    2023

  4. [12]

    Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. 2023. Training- Free Structured Diffusion Guidance for Compositional Text-to-Image Synthesis

    2023

  5. [13]

    Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. 2024. Ranni: Taming Text-to-Image Diffusion for Accurate Instruction Following. 4744– 4753

    2024

  6. [14]

    Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. 2025. ReNoise: Real Image Inversion Through Iterative Noising. In Computer Vision – ECCV 2024 , Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol (Eds...

    2025

  7. [16]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, and Sepp Hochreiter. 2017. GANs Trained by a Two Time-Scale Update Rule Converge to a Nash Equilibrium. ArXiv abs/1706.08500 (2017). https://api.semanticscholar.org/CorpusID:231697514

    2017 arXiv

  8. [17]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems, Vol. 33. 6840–6851

    2020

  9. [18]

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. 2024. ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignment. arXiv:2403.05135 [cs.CV]

    2024 arXiv

  10. [19]

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. 2023. T2I- CompBench: A Comprehensive Benchmark for Open-world Compositional Text- to-image Generation. In Advances in Neural Information Processing Systems , Vol. 36. 78723–78747

    2023

  11. [21]

    Yangqin Jiang, Lianghao Xia, Wei Wei, Da Luo, Kangyi Lin, and Chao Huang

  12. [22]

    Zutao Jiang, Guian Fang, Jianhua Han, Guansong Lu, Hang Xu, Shengcai Liao, Xiaojun Chang, and Xiaodan Liang. 2024. RealignDiff: Boosting Text-to-Image Dif- fusion Model with Coarse-to-fine Semantic Re-alignment. ArXiv abs/2305.19599 (2024)

    2024 arXiv

  13. [23]

    Justin Johnson, Agrim Gupta, and Li Fei-Fei. 2018. Image Generation from Scene Graphs. 1219–1228

    2018

  14. [24]

    Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. 2023. Direct Inversion: Boosting Diffusion-based Editing with 3 Lines of Code. InInternational Conference on Representation Learning

    2023

  15. [25]

    Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. 2022. DiffusionCLIP: Text- Guided Diffusion Models for Robust Image Manipulation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 2416–2425. doi:10. 1109/CVPR52688.2022.00246

    2022

  16. [26]

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. 2023. Pick-a-Pic: An Open Dataset of User Preferences for Text-to- Image Generation. In Advances in Neural Information Processing Systems , Vol. 36. 36652–36663

    2023

  17. [27]

    Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al

  18. [28]

    Jiachen Li, Weixi Feng, Wenhu Chen, and William Yang Wang. 2024. Reward Guided Latent Consistency Distillation. Transactions on Machine Learning Re- search (2024). Featured Certification

    2024

  19. [29]

    Peiming Li, Ziyi Wang, Mengyuan Liu, Hong Liu, and Chen Chen. 2024. ClickDiff: Click to Induce Semantic Contact Map for Controllable Grasp Generation with Diffusion Models. In Proceedings of the 32nd ACM International Conference on Multimedia (MM ’24). Association for Computin...

    2024

  20. [30]

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. 2023. GLIGEN: Open-Set Grounded Text- to-Image Generation. 22511–22521

    2023

  21. [31]

    Zejian Li, Chenye Meng, Yize Li, Ling Yang, Shengyuan Zhang, Jiarui Ma, Jiayi Li, Guang Yang, Changyuan Yang, Zhiyuan Yang, Jinxiong Chang, and Lingyun Sun. 2024. LAION-SG: An Enhanced Large-Scale Dataset for Training Complex Image-Text Models with Structural Annotations. ArXi...

    2024 arXiv

  22. [32]

    Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. 2024. Gen- erative Image Dynamics. 24142–24153

    2024

  23. [33]

    Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Mingxi Cheng, Ji Li, and Liang Zheng. 2025. Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization. ArXiv abs/2406.04314 (2025)

    2025 arXiv

  24. [34]

    Jinxiu Liu and Qi Liu. 2024. R3CD: Scene Graph to Image Generation with Relation-Aware Compositional Contrastive Control Diffusion, Vol. 38. 3657– 3665

    2024

  25. [35]

    Tenenbaum

    Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum

  26. [36]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. arXiv:1711.05101 [cs.LG] https://arxiv.org/abs/1711.05101

    2019 arXiv

  27. [37]

    Yunhong Lu, Qichao Wang, Hengyuan Cao, Xierui Wang, Xiaoyin Xu, and Min Zhang. 2025. InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment. arXiv:2503.18454 [cs.CV] https: //arxiv.org/abs/2503.18454

    2025 arXiv

  28. [38]

    Yunhong Lu, Qichao Wang, Hengyuan Cao, Xiaoyin Xu, and Min Zhang. 2025. Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences. arXiv:2506.02698 [cs.CV] https://arxiv. org/abs/2506.02698

    2025 arXiv

  29. [39]

    Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, and Saining Xie

  30. [40]

    In European Conference on Computer Vision

    Compositional Visual Generation with Composable Diffusion Models. In European Conference on Computer Vision . 423–439

  31. [41]

    Sarthak Mittal, Korbinian Abstreiter, Stefan Bauer, Bernhard Schölkopf, and Arash Mehrjou. 2023. Diffusion based representation learning. In Proceedings of the 40th International Conference on Machine Learning (Honolulu, Hawaii, USA) (ICML’23). JMLR.org, Article 1039, 20 pages

    2023

  32. [42]

    Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schul- man, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, MM ’25, October 27–31, 2025, Dublin, Ireland ...

    2025

  33. [43]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. ArXiv abs/2307.01952 (2023). https://arxiv.org/abs/2307.01952

    2023 arXiv

  34. [44]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings...

    2021

  35. [45]

    arXiv:2501.09732 [cs.CV] https://arxiv.org/abs/2501.09732

    Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps. arXiv:2501.09732 [cs.CV] https://arxiv.org/abs/2501.09732

    arXiv

  36. [46]

    Navonil Majumder, Chia-Yu Hung, Deepanway Ghosal, Wei-Ning Hsu, Rada Mihalcea, and Soujanya Poria. 2024. Tango 2: Aligning Diffusion-based Text-to- Audio Generations through Direct Preference Optimization. In Proceedings of the 32nd ACM International Conference on Multimedia (...

    2024

  37. [47]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. 10674–10685

    2022

  38. [48]

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. 2022. Photore- alistic Text-to-Image Diffusion Models with Deep Lan...

    2022

  39. [49]

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Lud- wig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. 202...

    2022

  40. [50]

    Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive Learning Rates with Sublinear Memory Cost. arXiv:1804.04235 [cs.LG] https://arxiv.org/abs/ 1804.04235

    2018 arXiv

  41. [51]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems 36 (2023), 53728–53741

    2023

  42. [52]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen

  43. [53]

    ArXiv abs/2204.06125 (2022)

    Hierarchical Text-Conditional Image Generation with CLIP Latents. ArXiv abs/2204.06125 (2022). https://api.semanticscholar.org/CorpusID:248097655

    2022 arXiv

  44. [54]

    Ye Tian, Ling Yang, Xinchen Zhang, Yunhai Tong, Mengdi Wang, and Bin Cui

  45. [55]

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Sergey Levine, and Tommaso Biancalani. 2024. Feedback Efficient Online Fine-Tuning of Diffusion Models. In Proceedings of the 41st International Conference on...

    2024

  46. [56]

    Pascal Vincent. 2011. A Connection Between Score Matching and Denoising Autoencoders. Neural Computation 23, 7 (2011), 1661–1674. doi:10.1162/NECO_ a_00142

    2011 doi

  47. [57]

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. 2024. Diffusion Model Alignment Using Direct Preference Optimization. 8228–8238

    2024

  48. [58]

    Guibao Shen, Luozhou Wang, Jiantao Lin, Wenhang Ge, Chaozhe Zhang, Xin Tao, Yuanhui Zhang, Pengfei Wan, Zhong ming Wang, Guangyong Chen, Yijun Li, and Ying cong Chen. 2024. SG-Adapter: Enhancing Text-to-Image Generation with Scene Graph Guidance. ArXiv abs/2405.15321 (2024). h...

    2024 arXiv

  49. [59]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising Diffusion Implicit Models. In International Conference on Learning Representations

    2020

  50. [60]

    Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. 2024. Training-Free Consistent Text-to-Image Generation. ACM Trans. Graph. 43 (2024), 52:1–52:18. https://api.semanticscholar.org/CorpusID: 267412997

    2024

  51. [61]

    Yang Wu, Pengxu Wei, and Liang Lin. 2023. Scene Graph to Image Synthesis via Knowledge Consensus, Vol. 37. 2856–2865

    2023

  52. [62]

    ArXiv abs/2502.12146 (2025)

    Diffusion-Sharpening: Fine-tuning Diffusion Models with Denoising Tra- jectory Sharpening. ArXiv abs/2502.12146 (2025)

    2025 arXiv

  53. [63]

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. 2024. ImageReward: learning and evaluating human preferences for text-to-image generation. In Proceedings of the 37th International Conference on Neural Information Processing Syste...

    2024

  54. [64]

    Jaakkola

    Yilun Xu, Shangyuan Tong, and T. Jaakkola. 2023. Stable Target Field for Reduced Variance Score Estimation in Diffusion Models. In ArXiv

    2023

  55. [65]

    Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. 2024. Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model. 8941–8951

    2024

  56. [66]

    Ruichen Wang, Zekang Chen, Chen Chen, Jian Ma, Haonan Lu, and Xiaodong Lin. 2024. Compositional Text-to-Image Synthesis with Attention Map Control of Diffusion Models, Vol. 38. 5544–5552

    2024

  57. [67]

    Yunnan Wang, Ziqiang Li, Wenyao Zhang, Zequn Zhang, Baao Xie, Xihui Liu, Wenjun Zeng, and Xin Jin. 2024. Scene Graph Disentanglement and Composi- tion for Generalizable Complex Image Generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  58. [68]

    Zixiao Wang, Hongtao Xie, Yuxin Wang, Jianjun Xu, Boqiang Zhang, and Yong- dong Zhang. 2023. Symmetrical Linguistic Feature Distillation with CLIP for Scene Text Recognition. In Proceedings of the 31st ACM International Confer- ence on Multimedia (MM ’23) . Association for Com...

    2023

  59. [69]

    Po-Hung Yeh, Kuang-Huei Lee, and Jun cheng Chen. 2025. Training-Free Diffu- sion Model Alignment with Sampling Demons. In The Thirteenth International Conference on Learning Representations

    2025

  60. [70]

    Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. 2023. BoxDiff: Text-to-Image Synthesis with Training-Free Box-Constrained Diffusion. 7452–7461

    2023

  61. [71]

    Xinchen Zhang, Ling Yang, Guohao Li, Yaqi Cai, Jiake Xie, Yong Tang, Yujiu Yang, Mengdi Wang, and Bin Cui. 2024. IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation. ArXiv abs/2410.07171 (2024)

    2024 arXiv

  62. [72]

    Ziyi Zhang, Li Shen, Sen Zhang, Deheng Ye, Yong Luo, Miaojing Shi, Bo Du, and Dacheng Tao. 2024. Aligning Few-Step Diffusion Models with Dense Reward Difference Learning. ArXiv abs/2411.11727 (2024)

    2024

  63. [73]

    Dewei Zhou, You Li, Fan Ma, Zongxin Yang, and Yi Yang. 2024. Migc++: Advanced multi-instance generation controller for image synthesis. ArXiv abs/2407.02329 (2024)

    2024 arXiv

  64. [74]

    Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang

    Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, G. Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming-Hsuan Yang. 2022. Diffusion-Based Scene Graph to Image Generation with Masked Contrastive Pre-Training. ArXiv abs/2211.11138 (2022). https://api.semanticscholar.org/CorpusID...

    2022 arXiv

  65. [75]

    Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui

  66. [76]

    In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol

    Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235) . 56704– 56721. https://proceedings.mlr.press/v235/yang24ai.html

  67. [77]

    Shentao Yang, Tianqi Chen, and Mingyuan Zhou. 2024. A Dense Reward View on Aligning Text-to-Image Diffusion with Preference. In Proceedings of the 41st International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 55998–56032

    2024

  68. [79]

    Xinchen Zhang, Ling Yang, Yaqi Cai, Zhaochen Yu, Kaini Wang, Jiake Xie, Ye Tian, Minkai Xu, Yong Tang, Yujiu Yang, and Bin Cui. 2024. RealCompo: Balancing Realism and Compositionality Improves Text-to-Image Diffusion Models. ArXiv abs/2402.12908 (2024)

    2024 arXiv

  69. [83]

    Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. 2024. Migc: Multi- instance generation controller for text-to-image synthesis. 6818–6828

    2024

  70. [84]

    Shengzhe Zhou, Zejian Li, Shengyuan Zhang, Lefan Hou, Changyuan Yang, Guang Yang, Zhiyuan Yang, and Lingyun Sun. 2024. Reducing spatial fitting error in distillation of denoising diffusion models. In Proceedings of the Thirty- Eighth AAAI Conference on Artificial Intelligence ...

    2024

  71. [85]

    Yufan Zhou, Bingchen Liu, Yizhe Zhu, Xiao Yang, Changyou Chen, and Jinhui Xu. 2023. Shifted Diffusion for Text-to-image Generation. 10157–10166. Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models MM ’25, October 27–31, 2025, Dublin, Ireland Supplementary M...

    2023

  72. [86]

    log𝜎 𝛽E𝒙 𝑤 1:𝑇∼𝑝𝜃(𝒙 𝑤 1:𝑇|𝒙 𝑤 0) 𝒙𝑙 1:𝑇∼𝑝𝜃(𝒙𝑙 1:𝑇|𝒙𝑙 0) 𝑇∑︁ 𝑡 =1 " ∥𝜖𝜃(𝒙𝑤 𝑡 ,𝑡)− 𝜖𝜃0(𝒙𝑤 𝑡 ,𝑡)∥ 2−∥𝜖𝜃(𝒙𝑙 𝑡,𝑡)− 𝜖𝜃0(𝒙𝑙 𝑡,𝑡)∥ 2 # (S8) Inversion-DPO: Precise and Efficient Post-Training for Diffusion Models MM ’25, October 27–31, 2025, Dublin, Ireland Inversion-DPO∗ Buildingsurro...

    2025

  73. [2017]

    International Journal of Computer Vision 123, 1 (2017), 32–73

    Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision 123, 1 (2017), 32–73

    2017

  74. [2022]

    ArXiv abs/2204.05862 (2022)

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. ArXiv abs/2204.05862 (2022)

    2022 arXiv

  75. [2024]

    In Proceedings of the 32nd ACM International Conference on Multimedia (MM ’24)

    DiffMM: Multi-Modal Diffusion Model for Recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia (MM ’24) . Association for Computing Machinery, 7591–7599. doi:10.1145/3664647.3681498

  76. [2025]

    ArXiv abs/2409.08861 (2025)

    Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models with Memoryless Stochastic Optimal Control. ArXiv abs/2409.08861 (2025)

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.