Pith. sign in

REVIEW 3 major objections 6 minor 46 references

Can We Achieve Efficient Diffusion without Self-Attention? Distilling Self-Attention into Convolutions

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that self-attention in diffusion models is predominantly local, and that a distilled convolutional block can replace it with comparable fidelity and drastically lower cost.

desk verdict Worth engaging: the locality analysis and DeltaConvBlock are solid, but the headline 6929x/3.4x gains are module-level only, so the efficiency claim overstates end-to-end speedups. read the letter →

arxiv 2504.21292 v1 pith:UQXSVEZE submitted 2025-04-30 cs.CV

ZiYi Dong , Chengxing Zhou , Weijian Deng , Pengxu Wei , Xiangyang Ji , Liang Lin This is my paper · ORCID
classification cs.CV
keywords self-attentiondiffusionmodelsconvolutionalnetworksknowledgedistillationlocalattentionpatternstext-to-imagegenerationefficientinferencelinearcomplexity
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 tries to establish that self-attention in pretrained text-to-image diffusion models is overwhelmingly local, not global, and that it can therefore be replaced by convolutional computation without losing generation quality. Layer-wise and spectral analyses of SD1.5, SDXL, and PixArt show attention mass concentrated near each query and decaying roughly quadratically with distance, with effective receptive fields mostly below 15 to 20 pixels. The proposed DeltaConvFusion replaces every self-attention block with a pyramid-convolution block plus an average-pooling bias branch, trained by distilling the frozen teacher's features and noise predictions. The paper reports that this substitution matches or improves the teacher's image quality while cutting reported self-attention cost by up to 6929 times and beating LinFusion in efficiency, a result that, if right, removes the main scaling bottleneck for high-resolution generation.

What carries the argument

The load-bearing object is the DeltaConvBlock, a convolutional replacement for a self-attention block. It has two branches matched to the two observed components of attention maps: a pyramid-convolution branch, where the input is downsampled at several scales, passed through depthwise convolutions with a scaled gating nonlinearity, and upsampled back, so nearby pixels accumulate a hierarchy of receptive fields; and an average-pooling branch that produces a spatially invariant channel bias. The two branches are recombined with pointwise convolutions, and the whole module is trained with feature-level and output-level distillation while the pretrained model's other weights are frozen. Because the pyramid convolutions run at downsampled scales, the module's complexity is linear in the number of spatial positions rather than quadratic, which is the property that removes the self-attention bottleneck.

What would settle it

Measure total inference time and FLOPs for the complete denoising network, including frozen cross-attention, FFN, and ResNet blocks, at 512 by 512 and 1024 by 1024 pixels before and after the replacement on identical hardware. If the full-model speedup is close to 1x, the central efficiency claim fails; if it is a large multiple, the claim holds.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that global interactions are not the source of what self-attention contributes to current diffusion models. The attention map of a query pixel is concentrated in a small neighborhood; high-frequency structure decays quadratically with distance, and the remaining signal is a broad, spatially invariant bias. This pattern appears in both U-Net and DiT backbones, and replacing all self-attention layers with localized neighborhood attention with kernel size 13 already preserves coherent generation. DeltaConvFusion makes that observation operational: it distills the attention maps into DeltaConvBlocks, so the convolutional student reproduces the teacher's spatial behavior while the rest of the network stays frozen. On SD1.5, SDXL, and PixArt, the paper reports comparable or better DINOv2-based quality and lower FLOPs and latency than the attention-based originals.

Load-bearing premise

The headline efficiency gains assume the FLOPs and latency figures in Tables 1 and 2 measure the same scope for every method; if they count only the replaced self-attention module, the reported thousands-fold reduction does not produce an equivalent speedup for the whole model.

Editorial extensions

If this is right

  • If the locality finding holds, high-resolution generation can replace quadratic self-attention with linear-complexity local operators, removing the main scaling bottleneck at 4K and beyond.
  • Freezing the rest of the network and distilling only the replaced blocks is enough to keep the paper's image-quality, fidelity, and text-alignment metrics on par with or better than the teacher.
  • The same recipe applies to both U-Net and DiT backbones, indicating that the localized behavior is not specific to one architecture family.
  • A model trained only at 512 by 512 pixels can generate coherent 1024 by 1024 images after the replacement, whereas the self-attention baseline fragments.
  • The method is orthogonal to attention-sharing acceleration, so its efficiency gains can stack with caching and timestep-sharing approaches.

Reading between the lines

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

  • A natural extension the authors do not pursue is training a diffusion model from scratch with DeltaConvBlocks instead of distilling an attention teacher; the locality evidence suggests such a backbone could work, but the paper only demonstrates the distillation route.
  • The reported efficiency gain should be read as module-level: frozen cross-attention, FFN, and ResNet blocks still dominate the full network, so an end-to-end latency comparison would likely show a much smaller speedup than the headline 6929 times.
  • The localization result may not transfer to tasks requiring genuine long-range dependency, such as video or multi-object scenes with repeated elements; a test would be replacing self-attention in those models and checking whether quality degrades.
  • A quick validation of the paper's core observation would be to run the same attention-map analysis on a non-diffusion generative model; if those maps are not local, the locality is a feature of this training regime rather than of attention in general.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 studies whether self-attention in pre-trained diffusion models is truly needed for global reasoning, and proposes a convolutional replacement module (DeltaConvBlock) with pyramid convolution branches and an average-pooling bias branch. The authors analyze attention maps in SD1.5, SDXL, and PixArt, arguing that attention is predominantly local and can be decomposed into a quadratic-decay high-frequency component plus a spatially invariant low-frequency component. They then replace self-attention blocks with DeltaConvBlocks, freeze all other parameters, and train only the new blocks using feature-level and output-level distillation. Experiments on SD1.5, SDXL, and PixArt report generative quality comparable to or better than baselines on DINOv2-based metrics, and large FLOP/latency reductions. The central claims are that self-attention can be replaced by convolutions without quality loss, and that this yields a 6929x computational reduction and 3.4x latency improvements over prior efficient attention alternatives.

Significance. If the central claims hold, the paper would make a useful contribution to efficient diffusion model inference: it provides evidence against the necessity of global attention at every block, introduces a simple convolution-based replacement module, and demonstrates that distillation from a frozen teacher can transfer the relevant spatial behavior. The paper ships concrete architecture details, a training recipe, and a series of qualitative and quantitative comparisons across U-Net and DiT backbones. However, the significance is tempered by two issues: the headline efficiency numbers are scoped to the replaced module rather than to end-to-end generation, and the ERF matching evidence is partly guaranteed by the feature-distillation objective. The quality evidence is suggestive but based on nonstandard metrics and lacks statistical uncertainty quantification.

major comments (3)
  1. [Abstract; Sec. 5.2, Tables 1 and 2] The headline efficiency claims are not supported as stated because Tables 1 and 2 appear to measure only the self-attention/DeltaConvBlock module, not the full denoising network. For example, the 6929x reduction is exactly the ratio of the SD1.5 16K self-attention FLOPs (11,010,934 G) to the DeltaConvBlock FLOPs (1,589 G), and the 0.78 ms latency for 'Self-Attention' at 512x512 is far too small to be a whole-model forward pass. The abstract's 'reducing computational cost by 6929x' and 'surpassing LinFusion by 5.42x in efficiency' therefore refer to module-level savings, which do not translate to end-to-end generation speedups of the same magnitude because the frozen ResNet, FFN, cross-attention, text-encoder, and VAE components dominate the actual generation cost. The authors should either explicitly re-scope all such claims to the self-attention module or report full-model FLOPs and end-to-end per-image latency for a complete denoising run.
  2. [Sec. 5.5, Fig. 9; Sec. 4.3, Eq. (5)] The effective receptive field matching in Fig. 9 is largely a consequence of the feature-level distillation loss Lf in Eq. (5), which directly minimizes the difference between DeltaConvBlock outputs and self-attention outputs at every layer. Thus the observation that 'DeltaConvFusion exhibits an ERF pattern closely matching that of self-attention-based models' is not an independent validation of the architectural decomposition; it demonstrates that the convolutional block can imitate the teacher when explicitly trained to do so. To support the claim that the pyramid-plus-pooling design intrinsically captures the two attention components, the authors should compare the ERF of an undestilled or randomly initialized DeltaConvBlock against the teacher, or at minimum explicitly acknowledge that the match is by construction due to distillation.
  3. [Table 3, Sec. 5.4, Sec. 5.1] The quantitative quality comparison is based on point estimates of newly introduced metrics (DS and FDD from DINOv2) and a custom evaluation set of 10,000 LAION images annotated by InternVL2-8B, with no error bars or statistical significance tests. Since some differences in Table 3 are small (e.g., PixArt 1024 CLIP scores of 30.60 vs 30.55), it is difficult to assess whether the claimed parity or improvement is meaningful. The paper should report confidence intervals or variance estimates, and ideally also evaluate on a standard benchmark (e.g., MS-COCO or a widely used FID/CLIP setup) to allow external comparison, while noting why a custom set was preferred.
minor comments (6)
  1. [Sec. 4.1, Eq. (3) and gate definition] The definition of the scaled simple gate rho(f) is garbled in the text; please rewrite it as a clear mathematical expression, including how the feature is split into two channel halves and how the scaling by sqrt(C') is applied.
  2. [Table 1] The header of Table 1 contains a typo: '512 512 1024 1204' should be '512x512 1024x1024 2K 4K 8K 16K' (or similar), and the table should clarify in the caption that the FLOPs are for the self-attention/DeltaConvBlock component only, not the whole model.
  3. [Sec. 5.1] The phrase '4K curated real images from LAION' is ambiguous: it likely means 4,000 images, but '4K' usually denotes 3840x2160 resolution; please specify the exact number of images.
  4. [Sec. 5.1, Implementation details] The cross-reference 'In Equ.13' does not match the displayed equation numbering; the Min-SNR weighting appears as Eq. (6), so the reference should be corrected.
  5. [Author affiliations] The affiliations contain typos: 'Unviersity' should be 'University'; please proofread the author block.
  6. [Sec. 3.2, II. Frequency Analysis] The text says 'apply Discrete Fourier Transform (DFT) and then utilize a high-pass Butterworth filter'; it would be clearer to specify whether the DFT is two-dimensional and how the filter cutoff frequency was chosen, since this affects reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

ERF-matching validation is by construction; central quality and efficiency claims are not circular.

  1. fitted input called prediction [Sec. 4.3 Eq. (5) and Sec. 5.5, Fig. 9]
    "For feature-level alignment, we minimize the discrepancy Lf between the outputs of the ΔConvBlock and the original self-attention module across all the layers: Lf = Σ ||Δlθ(zlt) − zlt,out||^2 ... In Fig. 9, ΔConvFusion exhibits an ERF pattern closely matching that of self-attention-based models, indicating that ΔConvBlock effectively captures both high- and low-frequency characteristics of attention maps. This strong similarity further validates ΔConvBlock as a computationally efficient yet robust alternative that preserves essential feature extraction capabilities."

    The DeltaConvBlock is explicitly trained to reproduce the teacher self-attention output at every layer through the feature-level distillation loss Lf. Therefore the ERF similarity reported in Fig. 9 is an expected consequence of the training objective, not an independent confirmation of the architectural claim. Presenting this learned similarity as 'further validates' the design is a mild circular validation. The central generation-quality results (DS, FDD, CLIP in Table 3) and the FLOPs/latency arithmetic are external to this loop and do not reduce to the training objective.

full rationale

The paper's main claim is that self-attention in diffusion models is predominantly local and can be replaced by a distilled convolutional block. The locality observation is an empirical analysis of attention maps (Sec. 3.2), and the replacement is tested with external metrics (DINOv2 Score, FDD, CLIP) against baseline models (Table 3) and with end-to-end generated images; these are not derived from the training loss by construction. The efficiency comparison, while possibly scoped to the replaced self-attention module rather than the full denoising network, is an arithmetic/comparability concern rather than circularity. No load-bearing self-citation chain or imported uniqueness theorem is present. The only circular element is the ERF-matching presentation in Sec. 5.5, which is a by-product of the distillation objective; since it is not the central claim, the overall circularity is minor.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central result is an empirical architecture and efficiency claim rather than a mathematical derivation. It leans on several chosen hyperparameters (kernel sizes, pyramid stages, loss weights), one ad hoc decomposition of attention into high- and low-frequency components, and strong domain assumptions that measured locality and distillation transfer hold at test time and at higher resolutions. There are no invented physical entities.

free parameters (5)
  • Kernel size K = K=13 for SDXL and PixArt, K=9 for SD1.5, K=25 in an ablation
    Chosen from the effective receptive field analysis (Sec. 3.2, 5.1) rather than derived; the central efficiency-quality tradeoff depends on it.
  • Number of pyramid stages n = n=2 for all models
    Selection is reported without analysis of n; different n changes receptive field coverage and FLOPs.
  • Feature distillation weight beta = beta=0.001
    Set to match the loss scales (Sec. 5.1); the balance between feature-level and output-level distillation affects final quality.
  • Min-SNR weight gamma = gamma=5
    Adopted from prior work [11]; still a chosen constant in the output distillation loss (Eq. 6).
  • ASM and ERF threshold = 0.8 (80 percent)
    The effective receptive field is defined as the smallest kernel containing 80 percent of high-pass attention score mass (Eq. 2); this threshold sets the kernel-size design target.
assumptions (5)
  • domain assumption The measured attention locality on a limited set of prompts and timesteps generalizes to the full generation distribution, including multi-step sampling, classifier-free guidance, and out-of-training resolutions.
    Sec. 3.2 aggregates attention maps over T timesteps for prompts from Text2Image-Multi-Prompt; if these are not representative, the locality claim and the whole conv replacement rest on shaky ground.
  • domain assumption A feature-level and output-level distillation loss is sufficient to transfer behavior from the frozen teacher to the conv student.
    The method assumes that matching intermediate outputs and final epsilon-prediction on training images transfers to generated samples; no theory or strong out-of-distribution test supports this.
  • domain assumption The DINOv2-based DS and FDD metrics, plus CLIP score on a high-aesthetic LAION subset, are valid proxies for generative fidelity.
    The paper deliberately abandons standard benchmarks such as MS-COCO and inception-based metrics (Sec. 5.1), so the entire quality comparison relies on this evaluation choice.
  • standard math Softmax attention scores can be interpreted as a 2D spatial signal to which DFT and Butterworth high-pass filtering can be applied for ASM analysis.
    Used in Sec. 3.2 to derive the quadratic-decay and 80 percent ERF claims; the operations are standard, but applying them to attention probability distributions is an unvalidated modeling choice.
  • ad hoc to paper The two-component decomposition of self-attention, high-frequency quadratic-decay plus low-frequency spatially invariant bias, is complete enough to be replaced by pyramid convolutions plus average pooling.
    This decomposition motivates the DeltaConvBlock design (Sec. 4) but is not proven to be necessary or sufficient; the architecture could match observed maps for other reasons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can We Achieve Efficient Diffusion without Self-Attention? Distilling Self-Attention into Convolutions." pith.science (2026). https://pith.science/paper/UQXSVEZE

@misc{pith2026250421292,
  author       = {Pith},
  title        = {Pith review of: Can We Achieve Efficient Diffusion without Self-Attention? Distilling Self-Attention into Convolutions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQXSVEZE}},
  note         = {Machine review of arXiv:2504.21292}
}
abstract

Contemporary diffusion models built upon U-Net or Diffusion Transformer (DiT) architectures have revolutionized image generation through transformer-based attention mechanisms. The prevailing paradigm has commonly employed self-attention with quadratic computational complexity to handle global spatial relationships in complex images, thereby synthesizing high-fidelity images with coherent visual semantics.Contrary to conventional wisdom, our systematic layer-wise analysis reveals an interesting discrepancy: self-attention in pre-trained diffusion models predominantly exhibits localized attention patterns, closely resembling convolutional inductive biases. This suggests that global interactions in self-attention may be less critical than commonly assumed.Driven by this, we propose \(\Delta\)ConvFusion to replace conventional self-attention modules with Pyramid Convolution Blocks (\(\Delta\)ConvBlocks).By distilling attention patterns into localized convolutional operations while keeping other components frozen, \(\Delta\)ConvFusion achieves performance comparable to transformer-based counterparts while reducing computational cost by 6929$\times$ and surpassing LinFusion by 5.42$\times$ in efficiency--all without compromising generative fidelity.

Figures

Figures reproduced from arXiv: 2504.21292 by the authors.

Figure 1
Figure 1. Computational cost comparison between self-attention [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Attention visualization across different layers for FLUX, PixArt, and SD1.5. Each column shows attention maps at different [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overall diffusion architectures of U-Net and our proposed [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: ASM and its gradient across kernels with varying sizes. From (a), both PixArt and SD1.5 exhibit a consistent quadratic relationship between kernel size K and ASM. This indicates that attention primarily captures broad low-frequency information, as the accumulated atten…
Figure 5
Figure 5. Figure 5: Effective Receptive Field (ERF) of self-attention across [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Replacing all the self-attention layers in Stable Diffusion [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Overview of ∆ConvFusion. Left: Structure of ∆ConvBlock, consisting of two key components: (1) Pyramid convolution, which captures high-frequency, distance-dependent features through multi-scale convolutions. (2) Average pooling, which models the low￾frequency, spatiall…
Figure 8
Figure 8. Figure 8: Comparison of cross-resolution generation between [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Effective Receptive Field (ERF) visualization of [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    Butterworth

    S. Butterworth. On the Theory of Filter Amplifiers. Experi- mental Wireless & the Wireless Engineer, 7:536–541, 1930. 4

    work page 1930

  2. [2]

    Flash diffusion: Accelerating any conditional diffusion model for few steps image generation

    Clement Chadebec, Onur Tasar, Eyal Benaroche, and Ben- jamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 3

    arXiv 2024

  3. [3]

    Pixart-Σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-Σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91, 2024. 1, 3, 4, 8

    work page 2024

  4. [4]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Confer- ence on Computer Vision, pages 17–33, 2022. 6

    work page 2022

  5. [5]

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites

    Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 7

    arXiv 2024

  6. [6]

    Rifegan: Rich feature generation for text-to-image synthesis from prior knowledge

    Jun Cheng, Fuxiang Wu, Yanling Tian, Lei Wang, and Dapeng Tao. Rifegan: Rich feature generation for text-to-image synthesis from prior knowledge. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10908– 10917, 2020. 3

    work page 2020

  7. [7]

    Kaplan, and Enrico Shippole

    Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z. Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthe- sis with hourglass diffusion transformers. In International Conference on Machine Learning, 2024. 3

    work page 2024

  8. [8]

    Transformers are ssms: Generalized models and efficient algorithms through structured state space duality

    Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. In International Conference on Machine Learning,

    work page

Show all 46 references
  1. [9]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bj¨orn Ommer. Taming transformers for high-resolution image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition , pages 12873–12883, 2021. 3

    2021

  2. [10]

    Mamba: Linear-time sequence model- ing with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence model- ing with selective state spaces. CoRR, abs/2312.00752, 2023. 1

    2023 arXiv

  3. [11]

    Efficient diffu- sion training via min-snr weighting strategy

    Tiankai Hang, Shuyang Gu, Chen Li, Jianmin Bao, Dong Chen, Han Hu, Xin Geng, and Baining Guo. Efficient diffu- sion training via min-snr weighting strategy. In IEEE/CVF International Conference on Computer Vision, pages 7407– 7417, 2023. 2, 6, 7

    2023

  4. [12]

    Neighborhood attention transformer

    Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6185–6194, 2023. 3, 5

    2023

  5. [13]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 3

    2021

  6. [14]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. In Advances in Neural Information Processing Systems, 2020. 1, 3

    2020

  7. [15]

    Alias-free generative adversarial networks

    Tero Karras, Miika Aittala, Samuli Laine, Erik H ¨ark¨onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. In Advances in Neural Infor- mation Processing Systems, pages 852–863, 2021. 3

    2021

  8. [16]

    Elucidating the design space of diffusion-based generative models

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

    2022

  9. [17]

    Analyzing and improving the training dynamics of diffusion models

    Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 24174–24184, 2024. 1

    2024

  10. [18]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 1, 3

    2024

  11. [19]

    Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip H. S. Torr. Controllable text-to-image generation. In Advances in Neural Information Processing Systems, 2019. 3

    2019

  12. [20]

    Faster diffu- sion via temporal attention decomposition

    Haozhe Liu, Wentian Zhang, Jinheng Xie, Francesco Fac- cio, Mengmeng Xu, Tao Xiang, Mike Zheng Shou, Juan- Manuel Perez-Rua, and J ¨urgen Schmidhuber. Faster diffu- sion via temporal attention decomposition. arXiv preprint arXiv:2404.02747, 2024. 3

    2024 arXiv

  13. [21]

    Linfusion: 1 gpu, 1 minute, 16k image

    Songhua Liu, Weihao Yu, Zhenxiong Tan, and Xinchao Wang. Linfusion: 1 gpu, 1 minute, 16k image. CoRR, abs/2409.02097, 2024. 1, 2, 3, 7, 8

    2024 arXiv

  14. [22]

    More control for free! image synthesis with semantic diffusion guidance

    Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Ar- man Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. CoRR, abs/2112.05744,

    arXiv

  15. [23]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, , pages 9992–10002, 2021. 1

    2021

  16. [24]

    Token caching for diffusion transformer acceleration

    Jinming Lou, Wenyang Luo, Yufan Liu, Bing Li, Xinmiao Ding, Weiming Hu, Jiajiong Cao, Yuming Li, and Chenguang Ma. Token caching for diffusion transformer acceleration. arXiv preprint arXiv:2409.18523, 2024. 3

    2024

  17. [25]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongx- uan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. CoRR, abs/2211.01095, 2022. 7

    2022 arXiv

  18. [26]

    Understanding the effective receptive field in deep convolu- tional neural networks

    Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolu- tional neural networks. In Advances in Neural Information Processing Systems, 2016. 5

    2016

  19. [27]

    Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans

    Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P. 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. 3

    2023

  20. [28]

    GLIDE: towards photorealistic image gen- eration and editing with text-guided diffusion models

    Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image gen- eration and editing with text-guided diffusion models. In International Conference on Machine Learning, p...

    2022

  21. [29]

    Maxime Oquab, Timoth´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael R...

    2024

  22. [30]

    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. IEEE, 2023. 1, 3

    2023

  23. [31]

    SDXL: improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M¨uller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952, 2023. 1, 3, 8

    2023 arXiv

  24. [32]

    Proakis and Dimitris G

    John G. Proakis and Dimitris G. Manolakis. Digital signal processing (3rd ed.): principles, algorithms, and applications. USA, 1996. 4

    1996

  25. [33]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10674–10685, 2022. 1, 3, 4, 8

    2022

  26. [34]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Inter- vention, pages 234–241. Springer, 2015. 1

    2015

  27. [35]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion mo...

    2022 arXiv

  28. [36]

    Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis

    Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International Confer- ence on Machine Learning, pages 30105–30118, 2023. 3

    2023

  29. [37]

    Laion-5b: A large-scale dataset for training ai models

    Christoph Schuhmann, Richard Vencu, Romain Beaumont, et al. Laion-5b: A large-scale dataset for training ai models. https://laion.ai/, 2022. Accessed: 2023-10-01. 7

    2022

  30. [38]

    Fora: Fast-forward caching in diffusion transformer acceleration

    Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration. arXiv preprint arXiv:2407.01425,

    arXiv

  31. [39]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 3

    2021

  32. [40]

    Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations ,

  33. [41]

    Expos- ing flaws of generative model evaluation metrics and their unfair treatment of diffusion models

    George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Expos- ing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. Advances...

    2023

  34. [42]

    Rethinking the incep- tion architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the incep- tion architecture for computer vision. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2818–2826,

  35. [43]

    text2image-multi-prompt: A multi-prompt dataset for text-to-image generation

    Piotr Szyma ´nski. text2image-multi-prompt: A multi-prompt dataset for text-to-image generation. https : / / huggingface . co / datasets / pszemraj / text2image - multi - prompt , 2023. Accessed: 2023-10-01. 7

    2023

  36. [44]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8

    2017

  37. [45]

    midjourney-v5-202304-clean

    Wanng. midjourney-v5-202304-clean. https : / / huggingface . co / datasets / wanng / midjourney-v5-202304-clean , 2023. 7

    2023

  38. [46]

    Ditfastattn: Attention compression for diffusion transformer models

    Zhihang Yuan, Hanling Zhang, Lu Pu, Xuefei Ning, Lin- feng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. Ditfastattn: Attention compression for diffusion transformer models. In Advances in Neural Information Pro- cessing Systems, 2024. 1, 3, 7, 8

    2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.