Pith. sign in

REVIEW 3 major objections 6 minor 63 references

Energy-Guided Flow Matching

T0 review · 3 major / 6 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 This paper claims that replacing the fixed clean endpoint in flow matching with a heat-filtered moving endpoint, scheduled by each image's spectral energy, improves pixel-space generation quality and training efficiency.

desk verdict A clean derivation and promising FID results for a moving-endpoint flow-matching path, but the headline comparison is muddied by a CFG-tuning asymmetry and missing error bars. read the letter →

arxiv 2608.05811 v1 pith:YIXQFVU7 submitted 2026-08-06 cs.CV

Haoyang Tong (1 , 2) , Yu He (2) , Fang Li (2) , Lichen Ma (2 , 3) , Jingling Fu (2) , Dong Chen (2)
show 6 more authors
Zhen Chen (2) Junshi Huang (2) Jie Cao (1) ((1) MAIS & NLPR CASIA (2) JD.com (3) Xi'an Jiaotong University)
This is my paper · ORCID
classification cs.CV
keywords energy-guidedflowmatchingpixel-spacegenerationmovingendpointheat-kernelfilteringspectralenergyschedulingcoarse-to-finediffusiontransformers
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

Energy-Guided Flow Matching (EG-FM) claims that pixel-space generative models learn faster and better when the flow-matching trajectory is explicitly organized from low to high spatial frequencies. Standard flow matching interpolates noise toward a fixed clean image, so the order in which global structure and fine detail appear is left implicit. EG-FM replaces that fixed endpoint with a heat-kernel-filtered moving endpoint whose high-frequency content is released per image according to a global energy clock, and it derives the exact velocity target including an endpoint-motion term. On ImageNet class-conditional generation at 256×256, the paper reports FID 1.55 at 200 epochs and 1.45 at 600 epochs on PixelDiT-XL, beating the baseline's 1.61 at 320 epochs and 1.54 at 800 epochs, with similar gains on other backbones and on text-to-image generation. The method changes only the path and training target, so it adds negligible compute and requires no backbone or data changes.

What carries the argument

The load-bearing object is the moving spectral endpoint $y_t(x)$, a heat-kernel low-pass filtered version of the clean image $x$ whose cutoff sharpens as path time $t$ grows. Its behavior is controlled by the sample-adaptive heat time $h(x,t)$, solved from the energy-matching equation $G_x(h)/\tilde G_x = q(t)$ by bisection, and by the derived velocity target $v_t=y_t(x)-\epsilon+t\,\partial_t y_t(x)$, where the endpoint-motion term $t\,\partial_t y_t(x)$ is computed by implicit differentiation of the same equation. This machinery converts a coarse-to-fine prior into explicit training supervision while preserving the boundary distributions $z_0=\epsilon$, $z_1=x$ of ordinary flow matching.

What would settle it

Train EG-FM and standard flow matching on a dataset where global structure is identical across images but high-frequency texture energy varies widely; if the energy-matched schedule does not improve FID over the fixed-endpoint baseline, the coarse-to-fine energy ordering is not carrying the benefit. Alternatively, measure perceptual similarity, rather than L2 energy, between $y_t(x)$ and $x$ across images at fixed $t$; large variance across images would show the release clock does not align perceptual progress.

Watch

Extended reading notes

Core claim

The central claim is that a flow-matching trajectory with a sample-adaptive moving endpoint provides a better inductive bias for pixel-space generation than the standard fixed-endpoint interpolant. For each image, EG-FM constructs a family of endpoints $y_t(x)=\mathcal{F}^{-1}(R(h(x,t),\rho)\hat{x}(\rho))$, where the heat-kernel response $R(h,\rho)=\exp(-a h \rho^2)$ attenuates high radial frequencies according to a heat time $h(x,t)$. The heat time is chosen image-by-image so that the recovered spectral energy $G_x(h)$ divided by the total missing energy $\tilde G_x$ equals a single global release clock $q(t)$; thus the same path time means the same fraction of achievable frequency recovery for every image. The regression target becomes the exact path derivative $v_t = y_t(x) - \epsilon + t\,\partial_t y_t(x)$, with $\partial_t y_t(x)$ obtained by implicit differentiation of the energy constraint. The paper reports that this reparameterization consistently lowers FID and shortens training across PixelDiT, DeCo, and HyperDiT backbones, and improves GenEval and DPG-Bench scores in text-to-image transfer.

Load-bearing premise

The method assumes that the image's spectral energy, meaning the amount of signal at each spatial frequency, is the right ordering for what should be generated first, and that matching this ordering across images makes the same training time correspond to comparable perceptual progress for every image.

Editorial extensions

If this is right

  • On ImageNet 256×256, PixelDiT-XL with EG-FM reaches FID 1.55 at 200 epochs and 1.45 at 600 epochs, compared with 1.61 at 320 epochs and 1.54 at 800 epochs for the standard baseline.
  • On ImageNet 512×512, continuing from a 256×256 checkpoint for only 40 epochs gives FID 1.58 with HyperDiT-H and 1.68 with PixelDiT-XL, outperforming baselines trained for hundreds more epochs.
  • For text-to-image generation at 512×512, EG-FM-T2I scores 0.85 on GenEval and 83.9 on DPG-Bench, improving over the PixelDiT-T2I baseline while using the same backbone.
  • The added cost is negligible: per-sample FLOPs increase by about 0.01%, per-step wall time by 0.41–4.81%, and inference uses exactly the same solver, time grid, and number of function evaluations as standard flow matching.
  • The path prior transfers across three architecturally different backbones and three model sizes, and also improves an x-prediction baseline, though the x-prediction gain is smaller.

Reading between the lines

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

  • One testable consequence not pursued in the paper is that the release clock $q(t)$ could serve as a perceptual-progress clock at inference: if EG-FM makes path time correspond to comparable recovered energy across images, sampling schedules or per-image step allocation could be tuned from the same energy ratio.
  • Whether the energy-guided ordering helps likely depends on radial L2 spectral energy being a reasonable proxy for what a viewer perceives as global structure versus detail; a distribution dominated by high-frequency texture would test this directly.
  • The same heat-kernel and energy-clock construction could be applied to temporal frequency in video or audio generation, but the paper does not evaluate those settings.
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 Energy-Guided Flow Matching (EG-FM), a modification of the flow-matching training path for pixel-space image generation. Instead of interpolating noise toward a fixed clean image x, EG-FM uses a moving spectral endpoint y_t(x) obtained by heat-kernel low-pass filtering, with a sample-adaptive heat time h(x,t). The heat time is determined by matching the recovered spectral-energy fraction G_x(h)/\tilde G_x to a global release clock q(t) (Eq. 9), so that the same path time corresponds to the same fraction of energy recovery across images. The training target is the exact path derivative (Eq. 12), composed of the usual transport term and an endpoint-motion term. Experiments on ImageNet 256x256 and 512x512 report lower FID with fewer epochs on DeCo, HyperDiT, and PixelDiT, and improved GenEval/DPG scores on text-to-image generation; ablations study initial endpoint strength, schedule granularity, release clock, model size, and CFG scale.

Significance. The mathematical derivation is sound: the moving endpoint yields a valid conditional probability path, G_x(h) is monotonic so the bisection root is unique, and the velocity target in Eq. (12) is the exact path derivative. The method is simple, backbone-agnostic, and adds negligible training overhead and no inference overhead. The reported convergence improvements are striking if they hold under a fair evaluation protocol. However, the current evidence is weakened by evaluation-protocol issues, specifically asymmetric CFG tuning and the absence of repeated runs or error bars, so the central quantitative claim is not yet fully established.

major comments (3)
  1. [§4.4, Fig. 7; §4.1] The CFG scale is tuned for EG-FM (2.55), but no CFG sweep is reported for the Standard-FM baselines, and the baseline CFG values are not stated. Since Figure 7 shows that FID varies from 11.60 without guidance to 1.45 at the tuned CFG, comparing EG-FM at its tuned CFG against baselines at their original CFG is not a controlled comparison. Please report FID versus CFG for both EG-FM and each standard baseline, or fix the same CFG for both, and state the CFG values explicitly in the evaluation protocol.
  2. [§4.2, Tables 1 and 2] All FID numbers appear to be from single runs with no error bars or seed information. The reported differences between EG-FM and the corresponding baseline at comparable epochs are small (e.g., 1.45 vs 1.54 at 600/800 epochs, 1.51 vs 1.56 at 220/600 epochs, 1.63 vs 1.69 at 440/600 epochs), so seed-to-seed variation could change the ranking. Please provide repeated-run statistics (at least three seeds) or error bars for the main comparisons, or otherwise justify that the differences exceed run-to-run variability.
  3. [§4.2, Table 1] The claim that EG-FM 'consistently improves' quality is FID-specific: on DeCo-XL/16, sFID worsens (4.78 vs 4.59) and IS drops (300.1 vs 304.0); on HyperDiT-H, IS drops markedly (293.4 vs 306.5) and precision drops (0.78 vs 0.80). Please clarify whether EG-FM trades off IS/sFID for FID, and discuss this trade-off in the limitations or adjust the wording of the 'consistently improves generation quality' claim in the conclusion.
minor comments (6)
  1. [Abstract] Add a space in 'Flow Matching(EG-FM)'.
  2. [§3.1] In the sentence 'the high-frequency single of x is progressively released', 'single' should be 'signal'.
  3. [§3.2] 'guaranties' should be 'guarantees'.
  4. [§4.4] The caption and axis annotations in Figure 7 are difficult to read; please make the CFG values and the baseline CFG explicit in the figure or caption.
  5. [§C.4] The efficiency paragraph contains a duplicated, garbled sentence: 'Apart from PixelDiT-B/16, whose measured overhead Across all matched runs, the per-step wall-clock overhead ranges from 0.41% to 4.81%.' Please rewrite this passage.
  6. [§5] 'a effective design' should be 'an effective design'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: EG-FM's path construction, heat-time inversion, and exact velocity target are derived internally and benchmarked against external baselines; hyperparameter and CFG choices are evaluation concerns, not circular reductions.

full rationale

The paper's central derivation is self-contained. Section 3.1 defines the moving endpoint y_t(x) via a heat-kernel response (Eqs. 3-6). Section 3.2 defines the recovered-energy gap G_x(h) and solves for h(x,t) from Eq. (9) by monotonicity (Eqs. 10-11); this is a valid inversion because G_x is nonincreasing and the bracket endpoints match. Section 3.3 differentiates the path to obtain the exact velocity target (Eqs. 12-15), with endpoint stability proved in Appendix A.4. The schedule q(t) and filter width sigma_0 are hyperparameters selected by ablations (Tables 4-5, Fig. 5), not quantities fitted to the headline FID; the headline FID is reported as measured training results on external and controlled baselines. Self-citations to HyperDiT, PixelU, LiWi, and FrequencyBooster appear as backbone implementations, related work, or prior group outputs, but no load-bearing 'uniqueness' theorem or ansatz is imported from them; the uniqueness of h is proved in this paper by monotonicity. The CFG tuning asymmetry in Section 4.4 is an evaluation-fairness risk, not a circular reduction, and Appendix D's stated scope limitations are scope restrictions rather than circular dependencies. No claimed prediction reduces to its own input, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central derivation rests on standard Fourier analysis and flow matching theory; the method-specific components are the heat-kernel endpoint design, the global release clock, and the radial-energy proxy, which are validated only empirically. No new physical entities are introduced.

free parameters (4)
  • sigma_0 = 3.5
    Initial low-pass filter strength in heat-kernel response Eq. (3); chosen by FID sweep on PixelDiT-XL at 80 epochs (Fig. 5), not derived.
  • release_clock q(t) = 6 t^5 - 15 t^4 + 10 t^3 (quintic smootherstep)
    Global energy release clock in Eq. (9); selected via ablation among linear, smoothstep, smootherstep, and sigmoid (Table 5) on 80-epoch FID.
  • CFG scale = 2.55
    Sampling guidance scale tuned for EG-FM in Fig. 7; below the baseline setting, affects reported FID.
  • bisection iterations K = 16
    Numerical solver iterations for heat-time; chosen for accuracy, not optimized for FID.
assumptions (5)
  • standard math Parseval's theorem equates L2 image distance to summed spectral energy under a unitary DFT.
    Used in Eq. (7) and throughout to define G_x and G_tilde_x.
  • domain assumption The heat-kernel response R(h,rho)=exp(-a h rho^2) with a=(pi sigma_0)^2 yields a nested family of low-pass images that releases frequencies monotonically as h decreases.
    Design choice defining the moving endpoint in Eq. (6); assumed to encode a useful coarse-to-fine prior.
  • ad hoc to paper Aligning recovered spectral-energy fraction G_x(h)/G_tilde_x to a global release clock q(t) makes the same path time t represent comparable generative progress across images.
    Central modeling assumption behind Eq. (9); not proven to be optimal, only validated empirically.
  • standard math Conditional flow matching with the moving-endpoint path trains a velocity field whose marginal ODE generates the data distribution.
    Borrowed from flow matching theory (Lipman et al. 2022); the boundary conditions z_0=epsilon and z_1=x are preserved.
  • domain assumption Radial-frequency spectral energy is a sufficient descriptor of coarse-to-fine image structure.
    The schedule uses only radial frequency bins rho, discarding orientation and phase information when defining the release clock.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Energy-Guided Flow Matching." pith.science (2026). https://pith.science/paper/YIXQFVU7

@misc{pith2026260805811,
  author       = {Pith},
  title        = {Pith review of: Energy-Guided Flow Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YIXQFVU7}},
  note         = {Machine review of arXiv:2608.05811}
}
abstract

Pixel-space generative models bypass lossy latent compression, yet necessitate joint learning of global structure and fine-grained details in a high-dimensional space. Standard flow matching interpolates noise toward a fixed clean-image endpoint, leaving the spectral evolution to be learned implicitly. In this paper, we introduce Energy-Guided Flow Matching(EG-FM) that explicitly models a coarse-to-fine generative trajectory by moving endpoint. Specifically, EG-FM replaces the fixed endpoint with a heat-kernel-filtered endpoint that evolves smoothly from low-frequency image to clean image.The fraction of high-frequency signal in moving endpoint is released by an image-specific energy-guided scheduling, leading to the re-targeting of velocity in flow matching.Our framework requires no adaptation of the backbone and training data, bringing negligible cost on the training and inference stages. In our experiment, EG-FM consistently achieves lower FID on the ImageNet class-conditional image generation task at $256 \times 256$ with fewer epochs, reaching an FID of 1.55 at 200 epochs and 1.45 at 600 epochs. We continue training the generation task on the setting of $512 \times 512$ resolution, yielding a FID of 1.58 after only 40 high-resolution adaptation epochs.Furthermore, we transfer EG-FM on text-to-image generation and achieve 0.85 on GenEval score and 83.9 on DPG-Bench. Code is available at https://github.com/ysng123/EG-FM.

Figures

Figures reproduced from arXiv: 2608.05811 by the authors.

Figure 1
Figure 1. Comparison of baselines and Energy-Guided Flow Matching (EG-FM). Marker size increases as check￾points approach the desirable lower-left region, representing fewer training epochs and lower FID. Based on backbones of PixelDiT and DeCo, the EG-FM variants achieve improved FID performance with significant fewer training epochs. assuming that the model should move towards the same tar￾get image throughout generation pr… view at source ↗
Figure 2
Figure 2. Overview of Energy-Guided Flow Matching. Starting from a clean image, we construct a low-frequency endpoint and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of sample-adaptive heat-time. Under [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Qualitative results produced by PixelDiT-XL with EG-FM. From left to right, the panels show ImageNet class [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Initial-endpoint sensitivity after 80 epochs. The left [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 8
Figure 8. Figure 8: Visualization of the release-clock ablation. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: FID across training checkpoints computed with torch-fidelity and the ADM evaluation suite. The two implementations [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Additional text-to-image samples generated by EG-FM-T2I at [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Representative GenEval samples generated by EG-FM-T2I. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: ImageNet class 105 [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 15
Figure 15. Figure 15: ImageNet class 666 [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 17
Figure 17. Figure 17: ImageNet class 950 [PITH_FULL_IMAGE:figures/full_fig_p019_17.png]
Figure 18
Figure 18. Figure 18: ImageNet class 970 [PITH_FULL_IMAGE:figures/full_fig_p019_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 15 canonical work pages

  1. [1]

    International conference on machine learning , pages=

    Pixel recurrent neural networks , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  2. [2]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  3. [3]

    arXiv preprint arXiv:2011.13456 , year=

    Score-based generative modeling through stochastic differential equations , author=. arXiv preprint arXiv:2011.13456 , year=

    arXiv 2011

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  5. [5]

    International Conference on Machine Learning , pages=

    simple diffusion: End-to-end diffusion for high resolution images , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  6. [6]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Simpler Diffusion: 1.5 FID on ImageNet512 with pixel-space diffusion , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  7. [7]

    Forty-first International Conference on Machine Learning , year=

    Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers , author=. Forty-first International Conference on Machine Learning , year=

    work page

  8. [8]

    arXiv preprint arXiv:2511.13720 , year=

    Back to basics: Let denoising generative models denoise , author=. arXiv preprint arXiv:2511.13720 , year=

    arXiv

Show all 63 references
  1. [9]

    arXiv preprint arXiv:2504.07963 , year=

    Pixelflow: Pixel-space generative models with flow , author=. arXiv preprint arXiv:2504.07963 , year=

    arXiv

  2. [10]

    arXiv preprint arXiv:2507.23268 , year=

    Pixnerd: Pixel neural field diffusion , author=. arXiv preprint arXiv:2507.23268 , year=

    arXiv

  3. [11]

    arXiv preprint arXiv:2511.20645 , year=

    Pixeldit: Pixel diffusion transformers for image generation , author=. arXiv preprint arXiv:2511.20645 , year=

    arXiv

  4. [12]

    arXiv preprint arXiv:2410.10629 , year=

    Sana: Efficient high-resolution image synthesis with linear diffusion transformers , author=. arXiv preprint arXiv:2410.10629 , year=

    arXiv

  5. [13]

    arXiv preprint arXiv:2506.23543 , year=

    Pyramidal Patchification Flow for Visual Generation , author=. arXiv preprint arXiv:2506.23543 , year=

  6. [14]

    arXiv preprint arXiv:2511.18822 , year=

    Dip: Taming diffusion models in pixel space , author=. arXiv preprint arXiv:2511.18822 , year=

  7. [15]

    arXiv preprint arXiv:2605.15741 , year=

    Hyperdit: Hyper-connected transformers for high-fidelity pixel-space diffusion , author=. arXiv preprint arXiv:2605.15741 , year=

    arXiv

  8. [16]

    arXiv preprint arXiv:2606.27760 , year=

    PixelU: A U-Shaped Transformer for Efficient End-to-End Pixel Diffusion , author=. arXiv preprint arXiv:2606.27760 , year=

    arXiv

  9. [17]

    The eleventh international conference on learning representations , year=

    Flow matching for generative modeling , author=. The eleventh international conference on learning representations , year=

  10. [18]

    International conference on learning representations (ICLR) , year=

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. International conference on learning representations (ICLR) , year=

  11. [19]

    European Conference on Computer Vision , pages=

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  12. [20]

    arXiv preprint arXiv:2410.06940 , year=

    Representation alignment for generation: Training diffusion transformers is easier than you think , author=. arXiv preprint arXiv:2410.06940 , year=

    arXiv

  13. [21]

    generation: Taming optimization dilemma in latent diffusion models , author=

    Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  14. [22]

    arXiv preprint arXiv:2504.05741 , year=

    Ddt: Decoupled diffusion transformer , author=. arXiv preprint arXiv:2504.05741 , year=

    arXiv

  15. [23]

    arXiv preprint arXiv:2510.11690 , year=

    Diffusion transformers with representation autoencoders , author=. arXiv preprint arXiv:2510.11690 , year=

    arXiv

  16. [24]

    arXiv preprint arXiv:2206.13397 , year=

    Generative modelling with inverse heat dissipation , author=. arXiv preprint arXiv:2206.13397 , year=

    arXiv

  17. [25]

    arXiv preprint arXiv:2207.11192 , year=

    Progressive deblurring of diffusion models for coarse-to-fine image synthesis , author=. arXiv preprint arXiv:2207.11192 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv:2209.05557 , year=

    Blurring diffusion models , author=. arXiv preprint arXiv:2209.05557 , year=

    arXiv

  19. [27]

    arXiv preprint arXiv:2511.19365 , year=

    Deco: Frequency-decoupled pixel diffusion for end-to-end image generation , author=. arXiv preprint arXiv:2511.19365 , year=

    arXiv

  20. [28]

    arXiv preprint arXiv:2604.15521 , year=

    Frequency-Aware Flow Matching for High-Quality Image Generation , author=. arXiv preprint arXiv:2604.15521 , year=

    arXiv

  21. [29]

    arXiv preprint arXiv:2602.19461 , year=

    Laplacian Multi-scale Flow Matching for Generative Modeling , author=. arXiv preprint arXiv:2602.19461 , year=

  22. [30]

    arXiv preprint arXiv:2605.06421 , year=

    FREPix: Frequency-Heterogeneous Flow Matching for Pixel-Space Image Generation , author=. arXiv preprint arXiv:2605.06421 , year=

    arXiv

  23. [31]

    arXiv preprint arXiv:2606.02177 , year=

    Low-Pass Flow Matching , author=. arXiv preprint arXiv:2606.02177 , year=

    arXiv

  24. [32]

    arXiv preprint arXiv:2209.15571 , year=

    Building normalizing flows with stochastic interpolants , author=. arXiv preprint arXiv:2209.15571 , year=

    arXiv

  25. [33]

    2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR) , pages=

    Taming transformers for high-resolution image synthesis , author=. 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR) , pages=. 2021 , organization=

    2021

  26. [34]

    Advances in neural information processing systems , volume=

    Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=

  27. [35]

    arXiv preprint arXiv:2207.12598 , year=

    Classifier-free diffusion guidance , author=. arXiv preprint arXiv:2207.12598 , year=

    arXiv

  28. [36]

    Journal of Machine Learning Research , volume=

    Cascaded diffusion models for high fidelity image generation , author=. Journal of Machine Learning Research , volume=

  29. [37]

    Advances in neural information processing systems , volume=

    Elucidating the design space of diffusion-based generative models , author=. Advances in neural information processing systems , volume=

  30. [38]

    Advances in neural information processing systems , volume=

    Deep generative image models using a laplacian pyramid of adversarial networks , author=. Advances in neural information processing systems , volume=

  31. [39]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  32. [40]

    International conference on machine learning , pages=

    On the spectral bias of neural networks , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  33. [41]

    Advances in neural information processing systems , volume=

    Variational diffusion models , author=. Advances in neural information processing systems , volume=

  34. [42]

    International conference on machine learning , pages=

    Improved denoising diffusion probabilistic models , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  35. [43]

    Cold diffusion: Inverting arbitrary image transforms without noise , author=

  36. [44]

    International conference on learning representations , volume=

    Pixart- alpha : Fast training of diffusion transformer for photorealistic text-to-image synthesis , author=. International conference on learning representations , volume=

  37. [45]

    European Conference on Computer Vision , pages=

    Pixart- : Weak-to-strong training of diffusion transformer for 4k text-to-image generation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  38. [46]

    Forty-first international conference on machine learning , year=

    Scaling rectified flow transformers for high-resolution image synthesis , author=. Forty-first international conference on machine learning , year=

  39. [47]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  40. [48]

    Advances in neural information processing systems , volume=

    Gans trained by a two time-scale update rule converge to a local nash equilibrium , author=. Advances in neural information processing systems , volume=

  41. [49]

    Advances in neural information processing systems , volume=

    Improved techniques for training gans , author=. Advances in neural information processing systems , volume=

  42. [50]

    Advances in neural information processing systems , volume=

    Improved precision and recall metric for assessing generative models , author=. Advances in neural information processing systems , volume=

  43. [51]

    Advances in Neural Information Processing Systems , volume=

    Geneval: An object-focused framework for evaluating text-to-image alignment , author=. Advances in Neural Information Processing Systems , volume=

  44. [52]

    arXiv preprint arXiv:2403.05135 , year=

    Ella: Equip diffusion models with llm for enhanced semantic alignment , author=. arXiv preprint arXiv:2403.05135 , year=

    arXiv

  45. [53]

    Advances in neural information processing systems , volume=

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps , author=. Advances in neural information processing systems , volume=

  46. [54]

    Forty-third International Conference on Machine Learning , year=

    Latent forcing: Reordering the diffusion trajectory for pixel-space image generation , author=. Forty-third International Conference on Machine Learning , year=

  47. [55]

    arXiv preprint arXiv:2605.17759 , year=

    FrequencyBooster: Full-Frequency Modeling for High-Fidelity Pixel Diffusion , author=. arXiv preprint arXiv:2605.17759 , year=

    arXiv

  48. [56]

    arXiv preprint arXiv:2512.10794 , year=

    What matters for Representation Alignment: Global Information or Spatial Structure? , author=. arXiv preprint arXiv:2512.10794 , year=

  49. [57]

    arXiv preprint arXiv:2605.16147 , year=

    Registers matter for pixel-space diffusion transformers , author=. arXiv preprint arXiv:2605.16147 , year=

    arXiv

  50. [58]

    arXiv preprint arXiv:2605.11061 , year=

    Hidream-o1-image: A natively unified image generative foundation model with pixel-level unified transformer , author=. arXiv preprint arXiv:2605.11061 , year=

    arXiv

  51. [59]

    arXiv preprint arXiv:2607.17585 , year=

    Pixel-Space Diffusion Transformers , author=. arXiv preprint arXiv:2607.17585 , year=

    arXiv

  52. [60]

    arXiv preprint arXiv:2510.04504 , year=

    Asynchronous Denoising Diffusion Models for Aligning Text-to-Image Generation , author=. arXiv preprint arXiv:2510.04504 , year=

  53. [61]

    arXiv preprint arXiv:2605.14552 , year=

    LiWi: Layering in the Wild , author=. arXiv preprint arXiv:2605.14552 , year=

    arXiv

  54. [62]

    arXiv preprint arXiv:2408.00118 , year=

    Gemma 2: Improving open language models at a practical size , author=. arXiv preprint arXiv:2408.00118 , year=

    arXiv

  55. [63]

    arXiv preprint arXiv:2505.09568 , year=

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset , author=. arXiv preprint arXiv:2505.09568 , year=

    arXiv

Pith tools

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