Pith. sign in

REVIEW 3 major objections 6 minor 42 references

TeEFusion: Blending Text Embeddings to Distill Classifier-Free Guidance

T0 review · 3 major / 6 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 TeEFusion distills classifier-free guidance into a single forward pass by fusing conditional and null text embeddings, cutting inference cost up to 6× while matching teacher image quality.

desk verdict A simple, parameter-free CFG-distillation trick that mostly works; the '6x faster' headline is a forward-pass count, not a measured speedup. read the letter →

arxiv 2507.18192 v2 pith:ALRIOHSQ submitted 2025-07-24 cs.CV

Minghao Fu , Guo-Hua Wang , Xiaohao Chen , Qing-Guo Chen , Zhao Xu , Weihua Luo , Kaifu Zhang This is my paper
classification cs.CV
keywords classifier-freeguidancedistillationtextembeddingfusiontext-to-imagegenerationflowmatchingdiffusionmodelinferenceDiTarchitecture
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

TeEFusion claims that classifier-free guidance can be distilled into a text-to-image model with no extra parameters by moving the guidance combination out of the network's output space and into the text embedding space. The paper's recipe is to give the student a single fused embedding $c + w(c - \varnothing)$ instead of running the model twice and computing $(1+w)\epsilon(x,c) - w\epsilon(x,\varnothing)$. Because the student is trained on the teacher's denoised outputs under a complex reflection-based sampler, inference reduces to one simple forward pass per step, and the authors report up to $6\times$ faster generation with comparable or better aesthetic and compositional scores on SD3 and their in-house model. The paper presents this as a minimal, hyperparameter-free alternative to guidance-distillation methods that add extra networks.

What carries the argument

The load-bearing object is the fused text embedding $\hat z_{t,c,\varnothing,w}$, which injects the guidance magnitude into the same joint text-and-timestep embedding that DiT-based text-to-image models already construct. The term $F(c - \varnothing)$ is the projected difference between conditional and null prompt embeddings, and $G(\psi(w))$ plays the role of the coefficient $w$ while bounding the $O(w^2)$ variance growth of a raw product. The training mechanism is direct $\ell^2$ distillation: the student is initialized from the teacher and optimized to reproduce the teacher's denoised output under CFG combined with reflection sampling, so guidance and sampler are distilled simultaneously with no new network components.

What would settle it

Take one prompt, noise sample, and timestep and compare the teacher's true two-pass target $(1+w)\epsilon_{\theta_T}(x_t,c) - w\epsilon_{\theta_T}(x_t,\varnothing)$ with the student's single-pass output $\epsilon_{\theta_S}(x_t, c + w(c - \varnothing))$ on the same seed; if the mean squared error between the two predictions (or the image distance after full sampling) is not small across a suite of ordinary prompts at $w=5$, the central identification fails. A sharper version uses contradictory prompts such as 'cold fire', which the paper itself lists as a failure case.

Watch

Extended reading notes

Core claim

The central discovery is that the two-input structure of classifier-free guidance can be collapsed: instead of computing the guided score as a linear combination of a conditional and an unconditional network output, one can form the linear combination on the text side and then run a single network evaluation on the fused embedding. The paper first verifies that elementwise addition and subtraction of text embeddings reliably adds and removes semantic content, using masked-prompt experiments, and then uses that observation to justify the fusion $\hat c = c + w(c - \varnothing)$. To keep large guidance scales stable, the scalar $w$ is encoded through the same sinusoidal position embedding as the timestep, producing the joint embedding $\hat z_{t,c,\varnothing,w} = G(\psi(t)) + F(c) + G(\psi(w))F(c - \varnothing)$. A student initialized from the teacher and trained with an $\ell^2$ loss to match the teacher's CFG-plus-reflection output then reproduces the teacher's images with plain Euler sampling and one forward pass.

Load-bearing premise

The load-bearing premise is that the network's output is nearly linear in the text embedding, so feeding it the fused embedding $c + w(c - \varnothing)$ behaves like blending the conditional and unconditional outputs with the same weights; the paper supports this empirically on a few prompts and with cosine-similarity scores, but it is not proven and could break for more nonlinear encoders or contradictory prompts.

Editorial extensions

If this is right

  • A single forward pass per denoising step replaces the two or more passes required by CFG and reflection-based samplers, giving guidance-level quality at roughly the cost of unguided generation.
  • Because the student is trained on the teacher's output under a chosen complex sampler, any future test-time-scaling improvement in the teacher's sampling strategy can be folded into the same cheap inference procedure.
  • No architectural change or extra hyperparameter is needed; the distilled model is identical in structure to the undistilled model, so it drops into existing pipelines.
  • Guidance strength remains selectable at inference through $w$, and the distilled model stays sensitive to it, letting users trade prompt adherence against aesthetic quality.
  • In the reported setting, the student avoids the high-$w$ collapse observed in the teacher, which the paper credits for its higher aesthetic scores.

Reading between the lines

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

  • Beyond the paper: if the linearity assumption is what carries the method, applying the same $c + w(c - \varnothing)$ recipe to a text encoder with a different architecture is a direct stress test, and the cosine-similarity evidence suggests the margin may shrink for contradictory prompts.
  • Beyond the paper: the fused-embedding view makes guidance a semantic direction one can measure, so a benchmark could check whether interpolation between two prompt embeddings interpolates the generated images, quantifying when linearity holds.
  • Beyond the paper: combining this fusion with step-distillation or one-step consistency training could compound the savings, since each attacks a different part of the inference budget.
  • Beyond the paper: the $\ell^2$ objective ties the student to the teacher's output distribution, so on prompts far outside the training distribution the student may drift; testing on unseen styles would define the operational envelope.
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 / 6 minor

Summary. The paper proposes TeEFusion, a distillation method that eliminates the two-forward-pass overhead of classifier-free guidance (CFG) in text-to-image diffusion models. The key idea is to fuse the conditional and unconditional text embeddings via the linear operation c + w(c - ∅), with the guidance scale w encoded through a sinusoidal embedding and multiplied into the embedding difference using existing MLPs from the pretrained model. The student model is initialized from the teacher and trained to match the teacher's output produced with a complex sampling strategy (e.g., W2SD+CFG), while the student itself uses simple Euler sampling without CFG. Experiments on SD3 and an In-house T2I model report consistent improvements over the DistillCFG baseline on HPS, DPG-Bench, and prompt-following metrics, with a claimed inference speedup of up to 6x.

Significance. If the results are reproducible, TeEFusion is a simple and practical contribution: it requires no architectural changes, adds no trainable parameters, and is compatible with diverse teacher samplers. The paper's strengths include evaluation on large-scale models (SD3 2B, In-house 1B), standard benchmarks such as HPS and DPG-Bench, and public code. The method's core hypothesis—that linear operations in text embedding space approximate CFG in output space—is clearly stated and empirically tested, with honest reporting of failure cases in Appendix A.2. However, the absolute performance gains over the baseline are modest and the speedup claim is based on a cost model rather than measured latency, so the practical advantage over simpler baselines remains to be fully substantiated.

major comments (3)
  1. [Sec. 6.2 and Table 1] The 'up to 6× faster' claim is not backed by wall-clock timing. The Cost column in Table 1 lists forward-pass multipliers (2×, 6×, 1×), and the caption states 'Cost' refers to inference time 'measured and compared separately for each model,' but the numbers are clearly derived from the forward-pass counts in Figure 1, not from latency measurements. The teacher's W2SD+CFG uses weak and strong models (Sec. 6.1, from CHATS); if the weak model is cheaper than the strong model, the actual teacher cost is less than six strong-model passes, and the true speedup is correspondingly lower. The authors should report parameter counts, per-step FLOPs, and measured end-to-end inference times (including text encoding and sampler overhead) for both the teacher and student pipelines to substantiate the headline efficiency claim.
  2. [Tables 1–4] All quantitative results are from single runs without standard deviations, confidence intervals, or significance tests. For example, Table 1 shows TeEFusion outperforming DistillCFG by 1.23 HPS points on the SD3 Anime group and 0.80 points on the Photo group, while Table 2 shows TeEFusion below the teacher on DPG-Bench Global (81.76 vs. 82.98) and Overall (84.56 vs. 86.56). Without variance estimates, the claim that TeEFusion 'significantly outperforms' the baseline and 'even the teacher model' is not statistically supported. The authors should add multiple seeds (or at least report the number of generated images per prompt and measure run-to-run variation) and, where possible, run paired significance tests on the evaluation metrics.
  3. [Sec. 4 and Eq. 5] The central assumption that linear combinations of text embeddings reproduce CFG in output space is validated only through qualitative examples (Fig. 2) and cosine similarity (Appendix A.1). A direct quantitative validation—for instance, comparing ϵθ(x_t, c + w(c − ∅)) against (1 + w)ϵθ(x_t, c) − wϵθ(x_t, ∅) over a diverse set of prompts and timesteps, and reporting the L2 distance or a distributional metric—would strengthen the paper. Appendix A.2 lists failure cases such as 'cold fire' and 'not a cat' but does not characterize their frequency across a benchmark or identify the conditions under which the linear-fusion approximation is reliable. These additions would either confirm the method's applicability or give practitioners clear boundaries.
minor comments (6)
  1. [Sec. 5, Eq. 7] The notation F(c − ∅) is ambiguous because c and ∅ denote text embeddings, but the paper states the difference is computed 'after projection into the text embedding space.' Clarify whether F is applied to the difference vector or the difference of the projected embeddings, and make the equation's variable definitions explicit.
  2. [Sec. 6.1] The claim that TeEFusion introduces 'no extra hyperparameters' should be qualified: the guidance scale range [wmin, wmax] = [2, 14] in Algorithm 1 is a hyperparameter, as are the learning rate, batch size, and training steps. The absence of newly introduced architecture hyperparameters is a strength, but the phrasing is too broad.
  3. [Fig. 5a] The y-axis labeled 'Score' mixes HPS and CLIP, which have different scales and interpretations. Plot them on separate axes or normalize, and state the prompt set used for the guidance-scale sweep.
  4. [Sec. 6.2] The text says TeEFusion 'even surpasses the teacher model' based on Table 1, but the DPG-Bench results in Table 2 show the student below the teacher on Global and Overall for SD3, and below the teacher on several columns for the In-house model. Qualify the claim to specify the metric and setting, or discuss the discrepancies.
  5. [Sec. 7 and Appendix A.2] The failure cases (e.g., 'not a cat', 'cold fire') are acknowledged but not analyzed in depth. A short discussion of when such semantic inconsistencies are more likely (negations, contradictory noun-adjective combinations, etc.) would help users understand the method's limits and guide future work.
  6. [References] Reference [20] is cited as 'DistillCFG' but the title is 'On distillation of guided diffusion models.' Ensure the citation style matches the terminology used in the text, and verify that the baseline implementation in the code corresponds to the description in Sec. 6.1.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: the student is evaluated on external benchmarks, and the fusion rule is a fixed input transformation, not a fitted predictor.

full rationale

The derivation chain is self-contained. TeEFusion starts from the CFG formula in Eq. 2, \tilde\epsilon = (1+w)\epsilon(c) - w\epsilon(\emptyset), and proposes the ansatz \hat\epsilon = \epsilon(x_t, c + w(c - \emptyset)) in Eq. 5. This is explicitly presented as an empirical hypothesis, tested in Sec. 4 via masked-prompt addition and cosine-similarity measurements; it is not derived from the definitions of CFG. The student is initialized from the teacher weights and trained with the distillation loss in Algorithm 1, \|\epsilon_{\theta_S}(x_t, \hat z) - \tilde\epsilon_{\theta_T}(x_t, w, c)\|^2, which is standard imitation of a teacher's output. Evaluation uses external benchmarks (HPS, DPG-Bench, CLIP, PickScore, ImageReward), so no reported score is forced by the construction. The only self-referential element is citation [6] (CHATS, by overlapping authors), which is used to build the weak and strong teacher models for W2SD sampling in Sec. 6.1. That citation supplies an input teacher model rather than a premise that determines the student's benchmark outcomes, so it is not load-bearing for the central claim. The '6x faster' statement is a forward-pass count (six teacher passes vs one student pass per sampling step) and may overstate wall-clock speedup if the weak model is cheaper, but this is a correctness concern, not circularity. No equation or fitted quantity in the paper reduces to its own input by construction.

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

The method is intentionally simple and introduces no new parameters, so the free-parameter list is empty. The central assumptions are about the linearity of text embedding space and the transferability of pretrained MLPs to encode guidance, both of which are empirically tested in the paper. No new physical or conceptual entities are posited.

assumptions (3)
  • domain assumption The text embedding space is approximately linear for semantic addition and subtraction, so the linear combination (1+w)c - w*empty in embedding space approximates the CFG combination in output space.
    Stated in Section 4 and Eq. 5. The paper tests this with a few qualitative examples and cosine similarities in Appendix A.1, but it is not proven and could fail for models with strongly nonlinear text encoders.
  • ad hoc to paper The MLP F and the time embedding MLP G from the pretrained model can be repurposed to encode the guidance scale w as G(psi(w)) without disrupting the joint embedding.
    Eq. 7 injects w through G(psi(w)) and F(c - empty), reusing pretrained modules that were never trained for this purpose. The paper argues this works empirically, and the ablation in Table 3 shows it helps, but the mechanism is not theoretically motivated.
  • domain assumption The teacher model with W2SD+CFG sampling (constructed using CHATS, a prior work by the same authors) is a valid and reliable distillation target.
    The teacher's behavior is the training target. If CHATS is not publicly available or is not robust, the exact teacher is hard to reproduce, which weakens reproducibility but does not invalidate the method itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TeEFusion: Blending Text Embeddings to Distill Classifier-Free Guidance." pith.science (2026). https://pith.science/paper/ALRIOHSQ

@misc{pith2026250718192,
  author       = {Pith},
  title        = {Pith review of: TeEFusion: Blending Text Embeddings to Distill Classifier-Free Guidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALRIOHSQ}},
  note         = {Machine review of arXiv:2507.18192}
}
abstract

Recent advances in text-to-image synthesis largely benefit from sophisticated sampling strategies and classifier-free guidance (CFG) to ensure high-quality generation. However, CFG's reliance on two forward passes, especially when combined with intricate sampling algorithms, results in prohibitively high inference costs. To address this, we introduce TeEFusion (Text Embeddings Fusion), a novel and efficient distillation method that directly incorporates the guidance magnitude into the text embeddings and distills the teacher model's complex sampling strategy. By simply fusing conditional and unconditional text embeddings using linear operations, TeEFusion reconstructs the desired guidance without adding extra parameters, simultaneously enabling the student model to learn from the teacher's output produced via its sophisticated sampling approach. Extensive experiments on state-of-the-art models such as SD3 demonstrate that our method allows the student to closely mimic the teacher's performance with a far simpler and more efficient sampling strategy. Consequently, the student model achieves inference speeds up to 6$\times$ faster than the teacher model, while maintaining image quality at levels comparable to those obtained through the teacher's complex sampling approach. The code is publicly available at https://github.com/AIDC-AI/TeEFusion.

Figures

Figures reproduced from arXiv: 2507.18192 by the authors.

Figure 1
Figure 1. Comparison of different sampling strategy on two [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Qualitative results by different prompting strategies. The leftmost three columns display images generated from different prompt [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of TeEFusion. Our approach shifts the fusion of the DiT [ [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison among various methods across different guidance scales [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Ablation study on varying (a) guidance scale [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Illustration of embedded w, with values randomly sam￾pled from the range [2, 14], visualized using t-SNE [31]. Switching Teacher Sampling. We also investigate the effectiveness of TeEFusion under different sampling strate￾gies used by teacher models (cf [PITH_FULL_IMA…
Figure 7
Figure 7. Figure 7: Generation examples of failure cases. Prompt: [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 32 canonical work pages

  1. [1]

    Zigzag diffusion sampling: Diffusion models can self-improve via self-reflection

    Lichen Bai, Shitong Shao, Zikai Zhou, Zipeng Qi, Zhiqiang Xu, Haoyi Xiong, and Zeke Xie. Zigzag diffusion sampling: Diffusion models can self-improve via self-reflection. In In- ternational Conference on Learning Representations, pages 1–14, 2025. 1, 3, 4

    work page 2025

  2. [2]

    Weak-to- strong diffusion with reflection

    Lichen Bai, Masashi Sugiyama, and Zeke Xie. Weak-to- strong diffusion with reflection. arXiv:2502.00473, 2025. 1, 3, 4, 6

    arXiv 2025

  3. [3]

    Universal guidance for diffusion models

    Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geip- ing, and Tom Goldstein. Universal guidance for diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 843–852, 2023. 3

    work page 2023

  4. [4]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. In- structPix2Pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392–18402, 2023. 1

    work page 2023

  5. [5]

    Scal- ing rectified flow Transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, et al. Scal- ing rectified flow Transformers for high-resolution image synthesis. In International Conference on Machine Learn- ing, pages 12606–12633, 2024. 1, 2, 3, 5, 6

    work page 2024

  6. [6]

    CHATS: Combin- ing human-aligned optimization and test-time sampling for text-to-image generation

    Minghao Fu, Guo-Hua Wang, Liangfu Cao, Qing-Guo Chen, Zhao Xu, Weihua Luo, and Kaifu Zhang. CHATS: Combin- ing human-aligned optimization and test-time sampling for text-to-image generation. arXiv:2502.12579, 2025. 6

    arXiv 2025

  7. [7]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 1, 3, 4, 6

    arXiv 2022

  8. [8]

    EllA: Equip diffusion models with LLM for en- hanced semantic alignment

    Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. EllA: Equip diffusion models with LLM for en- hanced semantic alignment. arXiv:2403.05135, 2024. 2, 3, 7

    arXiv 2024

Show all 42 references
  1. [9]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Samuli Laine, and Timo Aila. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Sys- tems, pages 26565 – 26577, 2022. 3

    2022

  2. [10]

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

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Ma- tiana, Joe Penna, and Omer Levy. Pick-a-Pic: An open dataset of user preferences for text-to-image generation. In Advances in Neural Information Processing Systems , pages 36652–36663, 2023. 7

    2023

  3. [11]

    FLUX.1-dev

    Black Forest Labs. FLUX.1-dev. https : / / huggingface . co / black - forest - labs / FLUX . 1-dev, 2024. 1, 2, 5, 6

    2024

  4. [12]

    One diffusion step to real-world super-resolution via flow trajectory distillation

    Jianze Li, Jiezhang Cao, Yong Guo, Wenbo Li, and Yulun Zhang. One diffusion step to real-world super-resolution via flow trajectory distillation. arXiv:2502.01993, 2025. 8

    2025 arXiv

  5. [13]

    SnapFusion: Text- to-image diffusion model on mobile devices within two sec- onds

    Yanyu Li, Huan Wang, Qing Jin, et al. SnapFusion: Text- to-image diffusion model on mobile devices within two sec- onds. In Advances in Neural Information Processing Sys- tems, pages 20662–20678, 2023. 2

    2023

  6. [14]

    Flow matching for generative mod- eling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling. In International Conference on Learning Representa- tions, pages 1–13, 2023. 1

    2023

  7. [15]

    Pseudo numerical methods for diffusion models on manifolds

    Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. InIn- ternational Conference on Learning Representations, pages 1–11, 2022. 3

    2022

  8. [16]

    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. In International Conference on Learning Rep- resentations, pages 1–15, 2023. 1, 6

    2023

  9. [17]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. In International Conference on Learning Representations, pages 1–18, 2019. 6

    2019

  10. [18]

    RePaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. RePaint: Inpainting using denoising diffusion probabilistic models. InIEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 11451–11461, 2022. 3

    2022

  11. [19]

    Efficient dif- fusion models: A comprehensive survey from principles to practices

    Zhiyuan Ma, Yuzhu Zhang, Guoli Jia, et al. Efficient dif- fusion models: A comprehensive survey from principles to practices. arXiv:2410.11795, 2024. 1, 2, 3

    2024 arXiv

  12. [20]

    On distillation of guided diffusion models

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14297–14306, 2023. 2, 3, 5, 6

    2023

  13. [21]

    DALL·E 2, 2022

    OpenAI. DALL·E 2, 2022. https://openai.com/ dall-e-2/. 1

    2022

  14. [22]

    Scalable diffusion models with Transformers

    William Peebles and Saining Xie. Scalable diffusion models with Transformers. In IEEE/CVF International Conference on Computer Vision, pages 4172–4182, 2023. 1, 2, 3, 6

    2023

  15. [23]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learn- ing transferable visual models from natural language super- vision. In International Conference on Machine Learning , pages 8748–8763, 2021. 2, 3, 7

    2021

  16. [24]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Confer- ence on Learning Representations, pages 1–21, 2022. 2, 3, 4

    2022

  17. [25]

    Fast high- resolution image synthesis with latent adversarial diffusion distillation

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high- resolution image synthesis with latent adversarial diffusion distillation. In ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in As...

    2024

  18. [26]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87–103, 2024. 3, 5

    2024

  19. [27]

    LAION-5B: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In Advances in neural information processing systems, pages 25278–25294, 2022. 6

    2022

  20. [28]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In International Conference on Learning Representations, pages 1–12, 2022. 3

    2022

  21. [29]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, pages 32211 – 32252, 2023. 3

    2023

  22. [30]

    Diffusion models without classifier-free guidance

    Zhicong Tang, Jianmin Bao, Dong Chen, and Baining Guo. Diffusion models without classifier-free guidance. arXiv:2502.12154, 2025. 3 9

    2025 arXiv

  23. [31]

    Visualizing data using t-SNE

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9 (86):2579–2605, 2008. 8

    2008

  24. [32]

    Atten- tion is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, et al. Atten- tion is all you need. In Advances in Neural Information Pro- cessing Systems, page 6000–6010, 2017. 5

    2017

  25. [33]

    Tune-A- Video: One-shot tuning of image diffusion models for text- to-video generation

    Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, et al. Tune-A- Video: One-shot tuning of image diffusion models for text- to-video generation. In IEEE/CVF International Conference on Computer Vision, pages 7589–7599, 2023. 3

    2023

  26. [34]

    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. arXiv:2306.09341, 2023. 1, 2, 3, 6, 7

    2023 arXiv

  27. [35]

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

    Jiazheng Xu, Xiao Liu, Yuchen Wu, et al. ImageRe- ward: Learning and evaluating human preferences for text- to-image generation. In Advances in Neural Information Processing Systems, pages 15903–15935, 2024. 7

    2024

  28. [36]

    One-step dif- fusion models with f-Divergence distribution matching

    Yilun Xu, Weili Nie, and Arash Vahdat. One-step dif- fusion models with f-Divergence distribution matching. arXiv:2502.15681, 2025. 2

    2025 arXiv

  29. [37]

    Freeman, and Taesung Park

    Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613–6623, 2024

    2024

  30. [38]

    Diffusion models are innate one-step generators

    Bowen Zheng and Tianming Yang. Diffusion models are innate one-step generators. arXiv:2405.20750, 2024. 2

    2024 arXiv

  31. [39]

    Fast sampling of dif- fusion models via operator learning

    Hongkai Zheng, Weilie Nie, Arash Vahdat, Kamyar Aziz- zadenesheli, and Anima Anandkumar. Fast sampling of dif- fusion models via operator learning. In International Con- ference on Machine Learning , pages 42390 – 42402, 2023. 3

    2023

  32. [40]

    DPM- Solver-v3: Improved diffusion ode solver with empirical model statistics

    Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. DPM- Solver-v3: Improved diffusion ode solver with empirical model statistics. In Advances in Neural Information Pro- cessing Systems, pages 55502 – 55542, 2024. 3

    2024

  33. [41]

    Score identity Distillation: Exponentially fast distillation of pretrained diffusion mod- els for one-step generation

    Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity Distillation: Exponentially fast distillation of pretrained diffusion mod- els for one-step generation. In International Conference on Machine Learning, pages 62307–62331, 2024. 2

    2024

  34. [42]

    cold fire

    Zhenyu Zhou, Defang Chen, Can Wang, Chun Chen, and Siwei Lyu. DICE: Distilling classifier-free guidance into text embeddings. arXiv preprint arXiv:2502.03726, 2025. 3 10 Figure 7. Generation examples of failure cases. Prompt: 1) not a cat. 2) liquid glass. 3) cold fire. A. Mor...

    2025

Pith tools

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