How Much To Guide: Revisiting Adaptive Guidance in Classifier-Free Guidance Text-to-Vision Diffusion Models

Huixuan Zhang; Junzhe Zhang; Xiaojun Wan

REVIEW 4 major objections 4 minor 1 cited by

How Much To Guide: Revisiting Adaptive Guidance in Classifier-Free Guidance Text-to-Vision Diffusion Models

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that classifier-free guidance only needs to run during the first 30-50% of denoising steps, preserving generation quality while saving 20-30% of inference time across image and video diffusion models.

desk verdict The p=0.5 branch is a genuinely useful speedup, but the 30% speedup claim overreaches and Appendix B has a copy-paste artifact that needs fixing. read the letter →

arxiv 2506.08351 v1 pith:5GHMGBT5 submitted 2025-06-10 cs.CV cs.AIcs.CL

Huixuan Zhang , Junzhe Zhang , Xiaojun Wan This is my paper

classification cs.CVcs.AIcs.CL

keywords classifier-freeguidanceadaptivediffusionmodelstext-to-imagegenerationtext-to-videoinferenceaccelerationsignal-to-noiseratio

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

This paper argues that classifier-free guidance (CFG), the standard way to condition diffusion models on text, does not need to run at every denoising step. The proposed strategy, Step AG, applies CFG only during the first 30-50% of denoising steps and then finishes with a single forward pass per step. Across four image models and two video models, this preserves FID and CLIP/VBench scores while cutting inference time by roughly 20-30%. The paper also proposes SNR AG, a signal-to-noise-ratio threshold version, and argues that the reason late-step guidance is unnecessary is that high SNR makes conditional and unconditional denoising paths converge. The practical payoff is a training-free, model-agnostic acceleration that can be combined with existing step-reduction methods.

What carries the argument

The central object is the signal-to-noise ratio $\lambda_t = \alpha(t)/\sigma(t)$ from the forward process $x_t = \alpha(t)x_0 + \sigma(t)\epsilon$. The proposed strategy selects a cutoff $t_0 = (1-p)T$; for $t \ge t_0$ it uses the guided score $\tilde{\epsilon}_\theta = \epsilon_\theta(x_t,t) + w[\epsilon_\theta(x_t,t,c) - \epsilon_\theta(x_t,t)]$, and for $t < t_0$ it uses a single forward pass, either conditional or unconditional. The argument is that when $\lambda_t$ is large, the data is barely noisy, the set of plausible denoising directions is small, and the conditional and unconditional paths land close together, so the extra forward pass is wasted.

What would settle it

Measure the actual distance between conditional and unconditional score predictions (or final samples) in the late denoising steps of a diffusion model with a very different noise schedule; if they diverge substantially at $p=0.5$ while FID and CLIP Score drop sharply, the convergence premise fails. A simpler experiment: run Step AG with $p=0.5$ on a model whose SNR curve rises much later than those in Figure 3 and check whether CLIP Score collapses.

Watch

Extended reading notes

Core claim

The central claim is that the second model forward required by classifier-free guidance is only needed while the noisy latent is still dominated by noise. Formally, for a schedule with SNR $\lambda_t = \alpha(t)/\sigma(t)$, CFG should be applied for steps with low SNR, i.e., $t \ge t_0$ where $t_0 = (1-p)T$, and a single score (conditional or unconditional) suffices afterwards. Step AG is shown empirically to match full-CFG generation: FID and CLIP Score stay competitive for $p=0.3$ and $p=0.5$ on Stable-Diffusion-3, Stable-Diffusion-XL, Stable-Diffusion-1.5 and PixArt-$\Sigma$-XL, and VBench metrics stay competitive on CogVideoX and ModelScope, with 20-30% average speedup. The paper further shows that the previous similarity-threshold adaptive guidance does not generalize, because cosine similarity between conditional and unconditional scores is high and non-monotonic on these models.

Load-bearing premise

The load-bearing premise is that late-step conditional and unconditional denoising paths converge because the signal-to-noise ratio is high, a heuristic asserted rather than proven, and that a single fixed guidance ratio $p$ transfers across models with different but broadly similar SNR schedules.

Editorial extensions

If this is right

CFG need not run at every step: applying it for the first 30-50% of steps keeps FID and CLIP Score within a few percent of full guidance on tested models.
Inference time drops by 20-30% in practice, close to the theoretical $(1-p)/2$ saving, with the gap accounted for by fixed text-encoding and decoding costs.
The same $p=0.3$ or $0.5$ works across U-Net and DiT architectures, different step counts, and video generators, suggesting the SNR trend generalizes.
Step AG combines with reduced total step counts: fewer steps plus truncated guidance still outperforms simply reducing steps at an equal forward-pass budget.
In the post-guidance phase either conditional or unconditional score works, but the conditional score is safer to avoid CLIP-Score drops at aggressive $p=0.3$.

Reading between the lines

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

If high-SNR convergence is the mechanism, a model-specific $p$ could be set by choosing the SNR threshold where conditional and unconditional predictions begin to differ by less than a set margin, rather than a fixed fraction.
The same reasoning should transfer to other modalities such as audio diffusion, where CFG is also used; a direct test would measure FAD or CLAP scores under Step AG.
Step AG is orthogonal to caching and distillation, so it could be stacked with layer caching or distilled samplers to compound speedups with little extra engineering.
The paper's observation that larger guidance scales can hurt both FID and CLIP suggests that guidance scale and guidance horizon interact; a per-model sweep of $(w,p)$ may be more informative than fixing either.

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.

Desk Editor's Note

read the letter

Here's the short version. Step AG—run CFG for the first p fraction of denoising steps, then switch to a single conditional forward pass—is a genuinely simple and useful trick. At p=0.5 the evidence is decent: across four image and two video models, FID is roughly flat and CLIP drops only 0.6-2.8%. That alone is a 20% inference saving with a small alignment cost, and it's worth knowing.

The p=0.3 branch, which is where the 30% speedup comes from, is much weaker. Their own Table 1 shows conditional-score CLIP drops of 2.1-5.3% across the four image models, and unconditional-score drops of 4.5-23.2%, with PixArt FID worsening by 33%. The paper calls this 'almost no cost' in the abstract and contributions; that is not what the table says. There are no confidence intervals or paired tests anywhere, so we can't tell how much of these differences is noise. The SNR rationale in Section 3.3 is a reasonable post-hoc story—at high SNR the two score paths should be close—but the observed CLIP drops suggest the small differences still matter for text alignment, and the same-p-across-models claim is weakened by their own Figure 3, which shows different SNR curves per model.

The Appendix B ImageNet table is a problem. The SDXL rows in Table 7 are numerically identical to the SDXL rows in Table 1 (the COCO table), which is impossible for a different dataset. This looks like a copy-paste error, but it invalidates the cross-dataset generalization evidence for that model and undermines confidence in the care with which the tables were assembled.

What's genuinely new and useful here: they show the similarity-based adaptive guidance from [3] fails to transfer (their SD1.5 replication is informative), and they validate a fixed-ratio heuristic across a broader set of models than prior work. The recommendation to use the conditional score in the late steps is honest, and the qualitative examples are consistent with the p=0.5 story.

Who is this for? Practitioners with CFG-based text-to-image or text-to-video models who want a training-free inference speedup and can accept a small alignment drop. The paper deserves a serious referee—the idea is simple, practical, and the empirical baseline work is useful—but not in its current form. The authors need to add uncertainty quantification, soften or split the speedup claim (p=0.5 vs. p=0.3), and fix the ImageNet table. If those are addressed, it becomes a solid systems/empirical contribution.

Referee Report

4 major / 4 minor

Summary. The paper proposes Step AG, a training-free adaptive guidance strategy for classifier-free guidance in text-to-vision diffusion models. The method applies CFG only for the first p fraction of denoising steps and uses a single forward pass (conditional or unconditional) for the remaining steps, with p=0.3 and p=0.5 as recommended values. The authors claim this preserves generation quality and image-text alignment while achieving a 20--30% inference speedup. They evaluate on four image diffusion models (SD3, SDXL, SD1.5, PixArt-Σ-XL) using FID and CLIP Score on MSCOCO, and on two video models (CogVideoX, ModelScope) using VBench, and they report an additional ImageNet experiment in Appendix B. They also replicate the prior Similarity AG method and show it fails on Stable-Diffusion-1.5. The central empirical claim is that Step AG is universally applicable across models, step counts, and modalities with minimal performance loss.

Significance. If the central claim holds, Step AG is a practically valuable, architecture-agnostic acceleration: it requires no training, no hyperparameter search beyond p, and no knowledge of the noise scheduler. The paper's breadth—four image and two video models, with default guidance scales and step counts—is a genuine strength, as is the replication showing that the previously proposed Similarity AG fails on a modern model. The method is simple enough for immediate adoption in existing CFG pipelines. However, the headline 'almost no cost' claim is stronger than the data support, particularly for p=0.3, and the Appendix B ImageNet results appear to be duplicated from the MSCOCO experiments, which raises a serious data-integrity concern that must be resolved before the empirical claims can be fully credited.

major comments (4)

[Section 5.1, Table 1, and Abstract]
[Appendix B, Table 7]
[Section 3.3]
[Section 5.3, Tables 8 and 9]

minor comments (4)

[Abstract]
[Section 3.3, Equation (8)]
[Section 4.2]
[Section 5.2, Table 3]

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Step AG is an empirical acceleration strategy validated on external models; its SNR rationale is heuristic and its speedup is arithmetically defined, not fitted.

full rationale

The paper's central strategy, Step AG, is defined by a guidance ratio p (Eq. 10) and is claimed to preserve quality because later denoising steps have high SNR (Section 3.3). This rationale is heuristic rather than derived, but it is not circular: the p values (0.3 and 0.5) are selected before evaluating FID/CLIP, and the quality metrics are computed on external models (SD3, SDXL, SD1.5, PixArt) and datasets (MSCOCO, VBench) against the p=1.0 baseline. The speedup of 20-30% follows arithmetically from using one forward pass instead of two for (1-p)T steps, which is a consequence of the definition, not a fitted prediction. There are no load-bearing self-citations: the cited prior works [3], [18], and [38] are external, and the paper explicitly distinguishes Step AG from them. The SNR explanation in Section 3.3 is an unproven premise that could be wrong, and the Appendix B ImageNet table duplicates SDXL rows from Table 1, which is a data-integrity problem, but neither is a case of the derivation reducing to its own inputs. Consequently, no circular step is identifiable.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim depends on one hyperparameter, p, chosen from empirical SNR curves, and on three domain assumptions about SNR monotonicity, convergence of conditional/unconditional paths, and portability of p across models. No new physical entities are introduced.

free parameters (1)

guidance ratio p = 0.5 and 0.3
Main hyperparameter of Step AG. The values are motivated by SNR curves in Figure 3 (lambda_t near 1 around p=0.3) and are claimed to transfer across models without tuning, which is itself an empirical assumption.

assumptions (4)

standard math Forward diffusion can be written x_t = alpha(t)x0 + sigma(t)epsilon and the denoiser output is a score epsilon_theta(x_t,t,c).
Section 3.1, Eq. 1 and 6. Standard formulation inherited from DDPM/CFG literature.
domain assumption lambda_t = alpha(t)/sigma(t) increases as t decreases and has similar trends across models and step counts.
Section 3.3 and Figure 3. Empirically observed; the shared-p design relies on it, and the paper concedes the same p gives different lambda thresholds per model.
domain assumption At high SNR, conditional and unconditional denoising outcomes are similar enough that using a single forward pass in late steps preserves quality.
Section 3.3. This is the mechanism claimed to justify Step AG, stated as intuition without a formal bound.
domain assumption Cosine similarity between conditional and unconditional scores is not a reliable threshold because it is high and non-monotonic in real models.
Section 3.2, Figure 2. Used to reject Similarity AG; generation-quality evidence is only shown for Stable-Diffusion-1.5 (Table 3).

how reviews work

0 comments

Cite this review

Pith. "Pith review of How Much To Guide: Revisiting Adaptive Guidance in Classifier-Free Guidance Text-to-Vision Diffusion Models." pith.science (2026). https://pith.science/paper/5GHMGBT5

@misc{pith2026250608351,
  author       = {Pith},
  title        = {Pith review of: How Much To Guide: Revisiting Adaptive Guidance in Classifier-Free Guidance Text-to-Vision Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5GHMGBT5}},
  note         = {Machine review of arXiv:2506.08351}
}

read the original abstract

With the rapid development of text-to-vision generation diffusion models, classifier-free guidance has emerged as the most prevalent method for conditioning. However, this approach inherently requires twice as many steps for model forwarding compared to unconditional generation, resulting in significantly higher costs. While previous study has introduced the concept of adaptive guidance, it lacks solid analysis and empirical results, making previous method unable to be applied to general diffusion models. In this work, we present another perspective of applying adaptive guidance and propose Step AG, which is a simple, universally applicable adaptive guidance strategy. Our evaluations focus on both image quality and image-text alignment. whose results indicate that restricting classifier-free guidance to the first several denoising steps is sufficient for generating high-quality, well-conditioned images, achieving an average speedup of 20% to 30%. Such improvement is consistent across different settings such as inference steps, and various models including video generation models, highlighting the superiority of our method.

Figures

Figures reproduced from arXiv: 2506.08351 by the authors.

**Figure 2.** Cosine Similarity 𝛾𝑡 of different models. We calculate cosine similarity under their default inference step settings. a total of 𝑇 steps can then be represented as a set 𝑆 = F𝑇 × F𝑇 −1 × ... × F1, where × represents Cartesian product. The objective of adaptive guidance (AG) is to identify a strategy 𝜁 ∈ 𝑆 that achieves competitive generation performance at a reduced computational cost. For example, the classic class… view at source ↗

**Figure 3.** SNR of different diffusion models. This is calculated under training timestep setting. Inference bears a similar trend. To take a step further, if 𝑝 is needed to be selected for each model specifically, the practical application of this method would still be degraded. However, as can be observed in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Performance of Step AG of different video generation models. SPV(ratio) is compared with spv under 𝑝 = 1.0. similar generation results across different 𝑝 values when 𝑇 is kept constant. This outcome illustrates the effectiveness of Step AG and aligns with our insight:…

**Figure 5.** Figure 5: “ A black Honda motorcycle parked in front of a garage. " generated by Stable-Diffusion-3 under different settings. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: “A room with blue walls and a white sink and door" generated by Stable-Diffusion-1.5 under different settings. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: “This is an open box containing four cucumbers" generated by Stable-Diffusion-XL under different settings. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

**Figure 8.** Figure 8: “A half eaten dessert cake sitting on a cake plate." generated by PixArt- [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

**Figure 9.** Figure 9: “A person is tasting beer" generated by CogVideoX under different settings. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

**Figure 10.** Figure 10: “A motorcycle gliding through a snowy field" generated by ModelScope under different settings. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Rethinking Visual Autoregressive Sampling with Information-Grounding Guidance
cs.CV 2025-09 conditional novelty 5.0 of 10

IGG, an attention-based reweighting of classifier-free guidance, concentrates guidance on important tokens and modestly improves FID/IS in scale-wise autoregressive image generation.

Reference graph

Works this paper leans on

49 extracted references · 14 canonical work pages · cited by 1 Pith paper

[3]

Angela Castillo, Jonas Kohler, Juan C Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. 2023. Adaptive guidance: Training-free acceleration of conditional diffusion models.arXiv preprint arXiv:2312.12487(2023)

arXiv 2023
[1]

Michael S Albergo and Eric Vanden-Eijnden. 2022. Building normalizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571(2022)

arXiv 2022
[2]

Tim Brooks, Aleksander Holynski, and Alexei A Efros. 2023. Instructpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 18392–18402

2023
[4]

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 (2024). arXiv:2403.04692 [cs.CV]

arXiv 2024
[5]

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer How Much To Guide: Revisiting Adaptive Guidance in Classifier-Free Guidance Text-to-Vision Diffusion Models , , vision and pattern recognition. Ieee, 248–255

work page 2009
[6]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis.Advances in neural information processing systems34 (2021), 8780–8794

2021
[7]

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. InForty- first International Conference on Machine Learning

2024
[8]

Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi

Show all 49 references

[9]

Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems30 (2017)

2017
[10]

Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. 2022. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303(2022)

2022 arXiv
[11]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilis- tic Models. InAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33. Curran Asso- ciates, Inc., 6840–6851. https://proce...

2020
[12]

Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598(2022)

2022 arXiv
[13]

Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. 2022. Video Diffusion Models. InAdvances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agar- wal, D. Belgrave, K. Cho, and A. Oh (Eds.), Vol. 35. Curran Associ...

2022
[14]

Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. 2024. VBench: Comprehensive Benchmark Suite for Video Generative Mod...

2024
[15]

Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. 2022. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems35 (2022), 26565–26577

2022
[16]

Diederik Kingma and Ruiqi Gao. 2024. Understanding diffusion objectives as the elbo with simple data augmentation.Advances in Neural Information Processing Systems36 (2024)

2024
[17]

Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2020. Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761(2020)

2020 arXiv
[18]

Tuomas Kynkäänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. 2024. Applying guidance in a limited interval improves sample and distribution quality in diffusion models.arXiv preprint arXiv:2404.07724 (2024)

2024 arXiv
[19]

Lijiang Li, Huixia Li, Xiawu Zheng, Jie Wu, Xuefeng Xiao, Rui Wang, Min Zheng, Xin Pan, Fei Chao, and Rongrong Ji. 2023. Autodiffusion: Training-free optimiza- tion of time steps and architectures for automated diffusion model acceleration. InProceedings of the IEEE/CVF Intern...

2023
[20]

Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. 2024. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds.Advances in Neural Information Processing Systems36 (2024)

2024
[21]

Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. InComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Procee...

2014
[22]

Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le
[23]

Hanxiao Liu, Karen Simonyan, and Yiming Yang. 2018. Darts: Differentiable architecture search.arXiv preprint arXiv:1806.09055(2018)

2018 arXiv
[24]

Xingchao Liu, Chengyue Gong, and Qiang Liu. 2022. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003(2022)

2022 arXiv
[25]

Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu
[26]

Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. 2022. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095(2022)

2022 arXiv
[27]

Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. 2023. VideoFusion: Decom- posed Diffusion Models for High-Quality Video Generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

2023
[28]

Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems35 (2022), 5775–5787

2022
[29]

Xinyin Ma, Gongfan Fang, and Xinchao Wang. 2024. Deepcache: Accelerating diffusion models for free. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15762–15772

2024
[30]

Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. 2023. On distillation of guided diffusion mod- els. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14297–14306

2023
[31]

Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. 2024. Learning- to-Cache: Accelerating Diffusion Transformer via Layer Caching. InAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds...

2024
[32]

William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. InProceedings of the IEEE/CVF International Conference on Computer Vision. 4195–4205

2023
[33]

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 preprint arXiv:2307.01952 (2023)

2023 arXiv
[34]

Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved denoising diffu- sion probabilistic models. InInternational conference on machine learning. PMLR, 8162–8171

2021
[35]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

2021
[36]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10684–10695

2022
[37]

Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. 2021. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning. PMLR, 8599–8608

2021
[38]

Seyedmorteza Sadat, Jakob Buhmann, Derek Bradely, Otmar Hilliges, and Ro- mann M Weber. 2023. CADS: Unleashing the diversity of diffusion models through condition-annealed sampling.arXiv preprint arXiv:2310.17347(2023)

2023 arXiv
[39]

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. Photo- realistic Text-to-Image Diffusion Models with Deep Lan...

2022
[40]

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. InMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings...

2015
[41]

Maximilian Seitzer. 2020. pytorch-fid: FID Score for PyTorch. https://github.com/ mseitzer/pytorch-fid. Version 0.3.0

2020
[42]

Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising Diffusion Implicit Models. InInternational Conference on Learning Representations. https: //openreview.net/forum?id=St1giarCHLP

2021
[43]

Tim Salimans and Jonathan Ho. 2022. Progressive distillation for fast sampling of diffusion models.arXiv preprint arXiv:2202.00512(2022)

2022 arXiv
[44]

Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. 2023. Modelscope text-to-video technical report.arXiv preprint arXiv:2308.06571(2023)

2023 arXiv
[45]

Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al . 2024. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer.arXiv preprint arXiv:2408.06072(2024)

2024 arXiv
[46]

Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. 2021. Score-Based Generative Modeling through Stochas- tic Differential Equations. InInternational Conference on Learning Representations. https://openreview.net/forum?id=PxTIG12RRHS

2021
[49]

A photo of {category}

Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jürgen Schmidhuber. 2024. Cross-attention makes inference cumbersome in text-to-image diffusion models.arXiv e-prints(2024), arXiv–2404. , , Huixuan Zhang, Junzhe Zhang, and Xiaojun Wan∗ A Detailed ...

2024
[2021]

In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing

CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 7514–7528

2021
[2022]

Flow matching for generative modeling.arXiv preprint arXiv:2210.02747 (2022)

2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [3]

Angela Castillo, Jonas Kohler, Juan C Pérez, Juan Pablo Pérez, Albert Pumarola, Bernard Ghanem, Pablo Arbeláez, and Ali Thabet. 2023. Adaptive guidance: Training-free acceleration of conditional diffusion models.arXiv preprint arXiv:2312.12487(2023)

arXiv 2023

[2] [1]

Michael S Albergo and Eric Vanden-Eijnden. 2022. Building normalizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571(2022)

arXiv 2022

[3] [2]

Tim Brooks, Aleksander Holynski, and Alexei A Efros. 2023. Instructpix2pix: Learning to follow image editing instructions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 18392–18402

2023

[4] [4]

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 (2024). arXiv:2403.04692 [cs.CV]

arXiv 2024

[5] [5]

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer How Much To Guide: Revisiting Adaptive Guidance in Classifier-Free Guidance Text-to-Vision Diffusion Models , , vision and pattern recognition. Ieee, 248–255

work page 2009

[6] [6]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis.Advances in neural information processing systems34 (2021), 8780–8794

2021

[7] [7]

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. 2024. Scaling rectified flow transformers for high-resolution image synthesis. InForty- first International Conference on Machine Learning

2024

[8] [8]

Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi