Pith. sign in

REVIEW 4 major objections 4 minor 87 references

Letting high-frequency wavelet bands stay pure noise until coarse structure emerges improves all three measured quality axes and halves sampling compute, setting a pixel-space FID record of 1.30 on ImageNet 512 and transferring to video.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 00:26 UTC pith:GPSXGP5I

load-bearing objection The delayed-HF schedule is clean and well ablated, but the sampling pseudocode contradicts the paper's own equations—it omits the 1/(1−t*) chain-rule factor for the HF update—so the results as specified are not reproducible without code. the 4 major comments →

arxiv 2607.28760 v1 pith:GPSXGP5I submitted 2026-07-30 cs.CV cs.AIcs.LGstat.ML

WaiT for the Signal: Simple Frequency-Aware Flow-Matching

Krunoslav Lehman Pavasovic , Th\'eophane Vallaeys , St\'ephane Mallat , Giulio Biroli , Luke Zettlemoyer , Brian Karrer , Jakob Verbeek This is my paper
classification cs.CV cs.AIcs.LGstat.ML
keywords wavelet flow matchingfrequency-aware noise schedulespixel-space image generationtexture fidelityFréchet wavelet distancefive-crop FIDhigh-resolution synthesisvideo generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper is trying to establish that standard flow matching wastes sampling effort: high-frequency texture becomes indistinguishable from pure noise far earlier than coarse structure, yet all frequencies are denoised on the same schedule. WaiT's response is a single, architecture-free change — decompose each image losslessly with a wavelet transform, run the coarse band on the usual schedule, and let the fine bands wait as pure noise until 25 percent of the way through, then refine them jointly. The claim is that this simultaneously improves global coherence, local detail, and texture fidelity, cuts sampling compute by up to half because the early phase runs on a quarter of the tokens, and transfers unchanged to text-to-image and video. On ImageNet 512 it reports a pixel-space FID record of 1.30, and on Kinetics-600 a state-of-the-art FVD of 0.84. A sympathetic reader would care because the gain comes from matching the natural frequency hierarchy of images rather than from a larger or more complex network.

Core claim

The paper's central claim is that the temporal asymmetry of noise corruption — fine bands lose signal far earlier than coarse structure — can be turned into a generative prior. WaiT applies a lossless Haar wavelet transform and runs two coupled schedules: the low-frequency band follows the standard flow on [0,1], while the high-frequency bands follow t_HF = max(0, (t − t*)/(1 − t*)) with t* = 0.25, where fine-band mutual information falls below 0.01 nats. Until t* the fine bands are exactly unit Gaussian noise, so injecting fresh noise at t* matches training with no discontinuity; both bands then refine jointly in pixel space. The paper reports this halves sampling compute (the coarse phase

What carries the argument

The carrying mechanism is a delayed linear noise schedule in the wavelet domain on an otherwise unchanged pixel-space transformer. A single-level 2D Haar discrete wavelet transform — up to scale, 2×2 average pooling plus the horizontal, vertical, and diagonal differences needed to invert it — losslessly splits the image into a low-frequency band of D/4 coefficients and detail bands of 3D/4. The LF band follows global time t on [0,1]; the HF bands follow t_HF = max(0, (t − t*)/(1 − t*)), so before t* = 0.25 they are exactly N(0, I), the same distribution the model sees at training time. Sampling runs a coarse phase on a quarter of the tokens, injects fresh HF noise at the matching variance, a

Load-bearing premise

The method stands on the premise that fine detail carries essentially no usable information early in the denoising trajectory, so holding it as pure noise until a quarter of the way through costs nothing in final quality — and that the two new scoring measures, a five-crop variant of FID and a high-frequency wavelet distance, genuinely track what people see in generated textures, a link the paper validates only on restored-photo distortions, not on generated images.

What would settle it

Measure the paper's own mutual-information quantity — noisy versus clean fine wavelet bands — on a broader, texture-rich corpus at fine time resolution: if fine-band MI climbs well above the 0.01-nat threshold before t = 0.25, the delay is discarding real signal. Likewise, a forced-choice human study at native resolution pitting WaiT outputs against strong latent-space generators on texture-heavy classes would settle whether the 5cFID and hFWD gains are perceptually real, since those metrics are currently validated only on restored-image distortions.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Pixel-space models can match or beat latent-space rivals on texture fidelity at native resolution without a trained autoencoder, because the wavelet transform is lossless and parameter-free.
  • The same two-phase recipe transfers with no algorithmic change to text-to-image at 1024×1024 (up to 3× throughput) and to video, where spatiotemporal wavelets reach a state-of-the-art 0.84 FVD on Kinetics-600 at roughly 30% lower compute.
  • Because only the schedule changes, the method layers onto other designs: applied to a decoupled latent-space transformer from scratch, it matched the baseline's FID with about a third of the training budget and 35% less inference compute.
  • Adding a third frequency level raises compute savings to 33% and further improves the fine-detail and texture metrics, so the gains appear to grow with the number of staggered bands.
  • Standard FID's 299×299 downsampling systematically hides the texture axis; the three-axis protocol (FID, 5cFID, hFWD) gives a more complete picture and changes how pixel- and latent-space methods rank.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 'fine detail waits for the signal' rule is a general design principle that should transfer to any generative process with a known multi-scale basis — audio, 3D shapes, or scientific fields — since the paper's own ablations show the benefit comes from the schedule, not from the specific wavelet chosen.
  • The crossover t* = 0.25 is implicitly tuned to natural-image statistics; datasets with more persistent high-frequency content (text, fur, fabric, medical or satellite imagery) would plausibly want a later crossover, and treating t* as a per-dataset hyperparameter is a cheap immediate extension the paper does not explore.
  • The latent-space transfer suggests that latent generators also waste early compute on high-frequency bands; applying the delay to an already-pretrained generator, rather than training from scratch as the paper does, is a direct testable follow-up.
  • If 5cFID and hFWD are adopted more widely, published rankings of high-resolution generators could shift substantially, because models ordered by FID alone are ordered differently once local detail and texture are measured — as the paper demonstrates for pixel- versus latent-space families.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes WaiT, a frequency-aware variant of pixel-space flow matching. It decomposes images with a Haar DWT, denoises the low-frequency band first on [0,1] while the high-frequency bands remain pure noise until a transition time t*, then injects fresh HF noise and jointly denoises all bands on [t*,1]. This 'delayed linear' schedule is claimed to remove the train–test mismatch of cascaded generation, to improve FID, 5-crop FID, high-frequency FWD, and to reduce sampling compute by up to 50%. The method is implemented on the JiT backbone and evaluated on ImageNet 512, OpenImages, text-to-image at 1024, and video (Taichi-HD, Kinetics-600), reporting a pixel-space ImageNet 512 FID of 1.30 with a 2B model, and FVD 0.84 on Kinetics-600. The paper also introduces a three-axis evaluation protocol (FID, 5cFID, hFWD) and provides detailed training/sampling pseudocode, deterministic dataset construction, and extensive ablations.

Significance. If the claims hold, the delayed-schedule idea is an elegant and potentially influential contribution: it is architecture-agnostic, lossless, and yields both quality and compute gains in pixel-space generation and in transfer to video and text-to-image. The paper is unusually transparent in providing side-by-side pseudocode, deterministic dataset recipes, and cumulative ablations that isolate design choices. The proposed metrics respond to a real gap, because standard FID downsampling discards high-frequency content. However, the significance is conditional on two load-bearing points: the sampling pseudocode must actually integrate the trained HF schedule, and the new metrics must track human perception of generative artifacts, not only restoration distortions. The paper also overstates Pareto dominance in places where Table 2 shows mixed results.

major comments (4)
  1. [Appendix A.2, Table 11] The Phase-1 HF update in the sampling pseudocode omits the chain-rule factor for the time change. With the forward interpolation z_HF = t_HF x_HF + (1 - t_HF) eps_HF and t_HF = (t - t*)/(1 - t*), differentiating with respect to global time gives dz_HF/dt = (1/(1 - t*)) (x_HF - z_HF)/(1 - t_HF). Table 11 updates z_HF += (t_{i+1} - t_i) (x_hat_HF - z_HF)/(1 - t_HF), which omits the factor 1/(1 - t*). As written, the sampler does not follow the training-time interpolation after the first Phase-1 step, so the claimed 'mathematical alignment' at t* does not persist. If the implementation includes the factor, the pseudocode is wrong; if it does not, the reported quality gains may be an artifact of a different stochastic process. Please correct the pseudocode or the derivation and state explicitly which one matches the released code.
  2. [Appendix D.2; Sections 4.2–4.3] 5cFID and hFWD are load-bearing for the 'texture fidelity' and 'Pareto-optimal' claims, but their perceptual validation is performed only on PIPAL, an image-restoration benchmark. PIPAL distortions (super-resolution artifacts, blur, noise, compression) are not necessarily representative of generative-model failure modes such as checkerboard patterns, flattened texture, or structural hallucinations. A correlation on restoration distortions does not establish that hFWD tracks perceived quality of generated images. Please either add a human study on model samples or an established generative-image IQA benchmark, or temper the texture-fidelity claims to 'on the proposed metrics'.
  3. [Table 2, Section 4.2] Several /32 configurations contradict the claimed dominance. WaiT-H/32 has FID 1.99 and hFWD 1.71, whereas JiT-H/32 has FID 1.94 and hFWD 1.33; WaiT-B/32 and WaiT-L/32 similarly have worse FID and hFWD than their JiT counterparts. Even among /16 variants, hFWD is sometimes slightly worse (WaiT-L/16 1.03 vs JiT-L/16 0.69). The statements that WaiT 'dominates' the JiT Pareto front and is 'Pareto-optimal across all three axes' are therefore not supported by the table as the claim is phrased. Please report the exact Pareto frontier per model scale and patch size, or restrict the claim to the configurations that actually lie on it.
  4. [Section 4.2 and Appendix A.2] Headline numbers are selected after sweeping t*, alpha, m, and CFG against the evaluation metrics, and no confidence intervals or multiple-seed runs are reported. For the claimed SOTA FID of 1.30, where differences between some models are as small as 0.03–0.1, sampling and selection noise may change the ranking. Please report variance across seeds or at least a sensitivity analysis over the tuned hyperparameters for the key compute-matched comparisons.
minor comments (4)
  1. [Section 3.1 and Table 1] Notation is inconsistent: Section 3.1 uses z_LF,t and z_HF,t, while Table 1 uses zlow/zhigh. Unify the symbols.
  2. [Appendix B] The mutual-information estimator is computed on only 50 real images and 50 generated trajectories. Since the choice t* = 0.25 is partly justified by this curve, please report the variability of the crossover estimate across subsamples or give more details on the estimator's variance.
  3. [Section 4.4] The sentence 'we are confident that minor, targeted tuning could yield further substantial GFLOP savings' is speculative and not needed for the empirical claims; remove or replace with a concrete experiment.
  4. [Conclusion] The conclusion calls the wavelet transform 'parameter-free', which is fine, but nearby text lists t*, SLF, alpha, and m as method components. Clarify that 'parameter-free' refers only to the DWT, not to the full method, to avoid confusion.

Circularity Check

0 steps flagged

No circularity: t* is swept, SLF is a fixed normalizer, and the metrics are not training targets.

full rationale

The paper's central claim is not constructed from its evaluation metrics or from a fitted parameter. The transition threshold t* is found by sweeping (Fig. 4b: 'Sweeping t* confirms this ... t*=0.25 gives a further −0.24 FID') and only afterwards rationalized with the mutual-information curves; the MI analysis is a post-hoc explanation, not a fitted target. SLF is a fixed data-dependent percentile used to normalize the LF band, and the appendix ablates normalization schemes rather than tuning SLF to the test metric. The training losses (Lcoarse, Lfine) supervise per-sample wavelet coefficients, while the reported hFWD and 5cFID are distribution-level statistics; hFWD uses a level-4 wavelet packet transform, whereas training uses a single-level Haar DWT, so the metric is not identical to the training signal. The theoretical self-citation [14] is used only as motivation for scale-wise stability, and the method's benefit is independently supported by the cumulative ablation in Fig. 4c, so that citation is not load-bearing. I also note a non-circular correctness/reproducibility concern: Table 11's Phase-1 HF update 'zHF← zHF + (ti+1−ti)· vHF' with 'vHF = (x̂HF−zHF)/(1−tHF)' omits the chain-rule factor 1/(1−t*) implied by tHF=(t−t*)/(1−t*), so the pseudocode as written would not follow the trained delayed schedule. This is a train-test mismatch in the written algorithm, not an equivalence-by-construction, and does not constitute circularity. Overall, no prediction reduces to a fitted parameter or to a self-citation chain, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

WaiT introduces no new physical entities. It relies on standard wavelet math, inherited flow-matching assumptions, and the domain assumption that HF bands are uninformative below t*. The main fitted quantities are t*, SLF, alpha, and CFG scale, all selected against evaluation metrics rather than derived from first principles.

free parameters (4)
  • t* (transition threshold) = 0.25
    Chosen by FID sweep on ImageNet 256 (Figure 4b) and used across all datasets, resolutions, and modalities; central to the method.
  • SLF (LF band normalization constant) = 1.94 (ImageNet 256), 1.95 (OpenImages 512)
    95th percentile of absolute LF wavelet coefficients computed over the training set; data-derived and dataset-specific.
  • alpha (timestep shift) = swept over {1,2,3}, best per setting
    Controls step concentration near t=0 and directly affects compute-quality trade-off; tuned per experiment.
  • CFG scale = swept [1.0,5.0] images, [1.0,10.0] text-to-image, best selected
    Classifier-free guidance scale affects every reported FID/FVD number and is selected on evaluation metrics per setting.
axioms (5)
  • standard math The Haar DWT is orthogonal, invertible, and lossless, so the LF band can act as a lossless latent space.
    Invoked in Section 3.1; standard wavelet theory.
  • standard math Flow matching with linear interpolants z_t = t x + (1-t) epsilon and x-prediction is a valid generative framework.
    Inherited from JiT and Lipman et al.; not rederived.
  • domain assumption High-frequency bands carry negligible signal below t* and can be treated as pure noise until coarse structure emerges.
    Central motivation, empirically supported by Figure 3's plug-in MI estimate on 50 images; assumed to transfer across datasets and modalities.
  • domain assumption Scale-wise factorization is more stable than joint modeling, so delaying fine details is beneficial.
    Cited to Guth et al. [14]; used in Section 3 to justify coarse-first generation.
  • domain assumption The self-curated OpenImages subsets and the 40M text-to-image corpus are representative benchmarks for the claimed SOTA comparisons.
    New datasets are constructed in Appendix E; they are not established community benchmarks, so comparisons on them are less standardized.

pith-pipeline@v1.3.0-alltime-deepseek · 26645 in / 11350 out tokens · 129226 ms · 2026-08-03T00:26:41.325324+00:00 · methodology

0 comments
read the original abstract

As image generation models scale to ever higher resolutions, global coherence, local detail, and texture fidelity become critical axes for generation quality. However, standard flow matching treats all spatial frequencies uniformly, ignoring the natural frequency hierarchy where high-frequency bands become indistinguishable from pure noise far earlier than coarse structures. We introduce WaiT, a Wavelet-aware image Transformer that decomposes generation into coarse and fine bands via lossless wavelets. True to its name, the high-frequency bands wait for the signal: staying pure noise until coarse structure has emerged, then joining the flow for joint refinement. Since standard FID discards fine-grained detail through aggressive downsampling, we introduce a more stringent three-axis evaluation protocol to assess quality at native resolution. On ImageNet 512x512, WaiT achieves a pixel-space FID of 1.43 and is Pareto-optimal across all three axes, reducing sampling compute by up to 50%. With our largest 2B model, we set a new state-of-the-art FID of 1.3 for pixel-space models on ImageNet 512 resolution. Our formulation outperforms even the strongest latent-space models on texture fidelity, and scales seamlessly to high-resolution OpenImages and to video generation, achieving a state-of-the-art FVD of 0.84 on Kinetics-600 with no algorithmic modifications.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

87 extracted references · 1 canonical work pages

  1. [1]

    Edify image: High-fidelity image generation with laplacian pyramids

    Matan Atzmon, Axel Sauer, and Tero Karras. Edify image: High-fidelity image generation with laplacian pyramids. NVIDIA Technical Report, 2024

    2024

  2. [2]

    Flux: State-of-the-art text-to-image generation

    Black Forest Labs. Flux: State-of-the-art text-to-image generation. https://blackforestlabs.ai/, 2024

    2024

  3. [3]

    A short note about kinetics-600

    Joao Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arXiv preprint arXiv:1808.01340 , 2018

    Pith/arXiv arXiv 2018

  4. [4]

    PixelFlow: Pixel-space generative models with flow

    Shoufa Chen, Chongjian Ge, Shilong Zhang, Peize Sun, and Ping Luo. PixelFlow: Pixel-space generative models with flow. arXiv preprint , 2504.07963, 2025. https://arxiv.org/abs/2504.07963

    Pith/arXiv arXiv 2025

  5. [5]

    DiP: Taming diffusion models in pixel space

    Zhennan Chen, Junwei Zhu, Xu Chen, Jiangning Zhang, Xiaobin Hu, Hanzhen Zhao, Chengjie Wang, Jian Yang, and Ying Tai. DiP: Taming diffusion models in pixel space. In Conference on Computer Vision and Pattern Recognition, 2026. https://arxiv.org/abs/2511.18822

    arXiv 2026

  6. [6]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems , 2021. https://arxiv.org/abs/2105.05233

    Pith/arXiv arXiv 2021

  7. [7]

    Spectral autoregression

    Sander Dieleman. Spectral autoregression. https://sander.ai/2024/09/02/spectral-autoregression.html, September 2024. Blog post

    2024

  8. [8]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.22706 , 2024

    arXiv 2024

  9. [9]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Danilo Lorber, Robin Rombach, and Björn Ommer. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206 , 2024

    Pith/arXiv arXiv 2024

  10. [10]

    A fourier space perspective on diffusion models, 2025

    Fabian Falck, Teodora Pandeva, Kiarash Zahirnia, Rachel Lawrence, Richard Turner, Edward Meeds, Javier Zazo, and Sushrut Karmalkar. A fourier space perspective on diffusion models, 2025. https://arxiv.org/abs/ 2505.11278

    Pith/arXiv arXiv 2025

  11. [11]

    Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smber, Julia Kreutzer, and Alec et al. Ross. Datacomp: In search of the next generation of multimodal datasets. In Advances in Neural Information Processing Systems, 2024

    2024

  12. [12]

    PIPAL: A large-scale im- age quality assessment dataset for perceptual image restoration

    Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. PIPAL: A large-scale im- age quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision (ECCV), 2020

    2020

  13. [13]

    Photorealistic video generation with diffusion models

    Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision ,

  14. [14]

    Wavelet score-based generative modeling

    Florentin Guth, Simon Coste, Valentin De Bortoli, and Stéphane Mallat. Wavelet score-based generative modeling. In Advances in Neural Information Processing Systems (NeurIPS) , 2022

    2022

  15. [15]

    Unified latents (UL): How to train your latents

    Jonathan Heek, Emiel Hoogeboom, Thomas Mensink, and Tim Salimans. Unified latents (UL): How to train your latents. arXiv preprint arXiv:2602.17270 , 2026. https://arxiv.org/pdf/2602.17270

    arXiv 2026

  16. [16]

    CLIPScore: A reference-free evaluation metric for image captioning

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In EMNLP, 2021

    2021

  17. [17]

    Fleet, Mohammad Norouzi, and Tim Salimans

    Jonathan Ho, Chitwan Saharia, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research , 23(47), 2022. https: //jmlr.org/papers/v23/21-0635.html

    2022

  18. [18]

    Simple diffusion: End-to-end diffusion for high resolution images

    Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. Simple diffusion: End-to-end diffusion for high resolution images. In International Conference on Machine Learning , 2023. https://arxiv.org/abs/2301.11093

    Pith/arXiv arXiv 2023

  19. [19]

    Simpler diffusion (SiD2): 1.5 FID on ImageNet512 with pixel-space diffusion

    Emiel Hoogeboom, Thomas Mensink, Jonathan Heek, Kay Lamerigts, Ruiqi Gao, and Tim Salimans. Simpler diffusion (SiD2): 1.5 FID on ImageNet512 with pixel-space diffusion. In Conference on Computer Vision and Pattern Recognition, 2025. https://arxiv.org/abs/2410.19324

    Pith/arXiv arXiv 2025

  20. [20]

    Scalable adaptive computation for iterative generation

    Allan Jabri, David Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. In International Conference on Machine Learning , 2023. https://arxiv.org/abs/2212.11972. 12

    Pith/arXiv arXiv 2023

  21. [21]

    Pyramidal flow matching for efficient video generative modeling

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. In ICLR, 2025. https://arxiv.org/abs/2410.05954

    arXiv 2025

  22. [22]

    Kingma and Ruiqi Gao

    Diederik P. Kingma and Ruiqi Gao. Understanding diffusion objectives as the ELBO with simple data augmen- tation. arXiv preprint , 2303.00848, 2023. https://arxiv.org/abs/2303.00848

    Pith/arXiv arXiv 2023

  23. [23]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4015–4026, 2023

    2023

  24. [24]

    The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision , 1...

    1956

  25. [25]

    Just image transformers: Pixel-space diffusion without pretrained autoencoders

    Xiang Li, Kai Chen, and Zhenguo Wang. Just image transformers: Pixel-space diffusion without pretrained autoencoders. arXiv preprint , 2025

    2025

  26. [26]

    Flow matching for generative modeling

    Yaron Lipman, Ricky Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint , 2022. https://openreview.net/forum?id=PqvMRDCJT9t

    2022

  27. [27]

    Albergo, Nicholas M

    Nanye Ma, Mark Goldstein, Michael S. Albergo, Nicholas M. Boffi, Eric Vanden-Eijnden, and Saining Xie. SiT: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In European Conference on Computer Vision , 2024. https://arxiv.org/abs/2401.08740

    Pith/arXiv arXiv 2024

  28. [28]

    Deco: Decoupled low-frequency semantics and high-frequency details for image generation

    Yuxin Ma, Wei Zhang, and Chen Liu. Deco: Decoupled low-frequency semantics and high-frequency details for image generation. arXiv preprint , 2025

    2025

  29. [29]

    Dctdiff: Intriguing properties of image generative modeling in the dct space

    Mang Ning, Mingxiao Li, Jianlin Su, Albert Ali Salah, and Itir Onal Ertugrul. Dctdiff: Intriguing properties of image generative modeling in the dct space. arXiv preprint arXiv:2401.02992 , 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In International Conference on Computer Vision , 2023. https://arxiv.org/abs/2212.09748

    Pith/arXiv arXiv 2023

  31. [31]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research , 21(1):5485–5551, 2020

    2020

  32. [32]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. International Journal of Computer Vision , 115(3):211–252, 2015

    2015

  33. [33]

    Jerome M. Shapiro. Embedded image coding using zerotrees of wavelet coefficients. IEEE Transactions on Signal Processing, 41(12):3445–3462, 1993. doi: 10.1109/78.258085

    work page doi:10.1109/78.258085 1993

  34. [34]

    First order motion model for image animation

    Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems , 2019

    2019

  35. [35]

    Skodras, C

    A. Skodras, C. Christopoulos, and T. Ebrahimi. The jpeg 2000 still image compression standard. IEEE Signal Processing Magazine, 18(5):36–58, 2001. doi: 10.1109/79.952804

    doi:10.1109/79.952804 2000

  36. [36]

    Frequency-aware flow matching for high-quality image generation

    Ren Sucheng, Yu Qihang, He Ju, Shen Xiaohui, and Chen Liang-Chieh. Frequency-aware flow matching for high-quality image generation. In CVPR, 2026

    2026

  37. [37]

    Qwen3-vl: Scaling vision-language models with improved architecture and training

    Qwen Team. Qwen3-vl: Scaling vision-language models with improved architecture and training. arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Jetformer: An autoregressive generative model of raw images and text

    Michael Tschannen, André Susano Pinto, and Alexander Kolesnikov. Jetformer: An autoregressive generative model of raw images and text. In International Conference on Learning Representations , volume 2025, pages 85995–86016, 2025

    2025

  39. [39]

    Towards accurate generative models of video: A new metric & challenges

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. In ICLR Workshop on Deep Generative Models for Highly Structured Data , 2019. https://arxiv.org/abs/1812.01717

    Pith/arXiv arXiv 2019

  40. [40]

    Fréchet wavelet distance: A domain-agnostic metric for image generation, 2025

    Lokesh Veeramacheneni, Moritz Wolter, Hildegard Kuehne, and Juergen Gall. Fréchet wavelet distance: A domain-agnostic metric for image generation, 2025. 13

    2025

  41. [41]

    PixNerd: Pixel neural field diffusion

    Shuai Wang, Ziteng Gao, Chenhui Zhu, Weilin Huang, and Limin Wang. PixNerd: Pixel neural field diffusion. In International Conference on Learning Representations , 2025. https://arxiv.org/abs/2507.23268

    Pith/arXiv arXiv 2025

  42. [42]

    DDT: Decoupled diffusion transformer

    Shuai Wang, Zhi Tian, Weilin Huang, and Limin Wang. DDT: Decoupled diffusion transformer. arXiv preprint , 2504.05741, 2025. https://arxiv.org/abs/2504.05741

    Pith/arXiv arXiv 2025

  43. [43]

    Demystifying clip data

    Hu Xu, Saining Xie, Xin Tan, Po-Yao Huang, et al. Demystifying clip data. arXiv preprint arXiv:2309.16671 , 2023

    Pith/arXiv arXiv 2023

  44. [44]

    Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G

    Lijun Yu, José Lezama, Nitesh B. Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion – tokenizer is key to visual generation. In International Conference on Learning Representations , 2024. http...

    Pith/arXiv arXiv 2024

  45. [45]

    Representation alignment for generation: Training diffusion transformers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. In International Conference on Learning Representations , 2025. https://arxiv.org/abs/2410.06940

    Pith/arXiv arXiv 2025

  46. [46]

    Pixeldit: Pixel diffusion transformers for image generation

    Yongsheng Yu, Wei Xiong, Weili Nie, Yichen Sheng, Shiqiu Liu, and Jiebo Luo. Pixeldit: Pixel diffusion transformers for image generation. arXiv preprint arXiv:2511.20645 , 2025

    Pith/arXiv arXiv 2025

  47. [47]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Wang, Alexander Kramer, et al. Sigmoid loss for language image pre-training. arXiv preprint arXiv:2303.15343, 2023

    Pith/arXiv arXiv 2023

  48. [48]

    Zhang, P

    R. Zhang, P. Isola, A. Efros, E. Shechtman, and O. Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Conference on Computer Vision and Pattern Recognition , 2018. https://arxiv.org/abs/ 1801.03924

    Pith/arXiv arXiv 2018

  49. [49]

    Laplacian multi-scale flow matching for generative modeling

    Zelin Zhao, Petr Molodyk, Haotian Xue, and Yongxin Chen. Laplacian multi-scale flow matching for generative modeling. arXiv preprint arXiv:2602.19461 , 2026

    arXiv 2026

  50. [50]

    Diffusion transformers with representation autoen- coders

    Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie. Diffusion transformers with representation autoen- coders. In International Conference on Learning Representations , 2026. https://openreview.net/forum?id= 0u1LigJaab. 14 Contents of Appendix A Implementation details 15 A.1 Training and inference . . . . . . . . . . . . . . . . . . . . . . . . . . ....

    2026

  51. [52]

    Sample tLF∼U (0, 1)

    Sample t∼U (ϵ, 1−ϵ) 1. Sample tLF∼U (0, 1)

  52. [53]

    xLF, xHF = DWT(x)

    ε∼N (0, I) 2. xLF, xHF = DWT(x)

  53. [54]

    ˜xLF = xLF / SLF — Phase 0 (coarse, H/2× W /2) —

    z = t· x + (1−t)· ε 3. ˜xLF = xLF / SLF — Phase 0 (coarse, H/2× W /2) —

  54. [55]

    εc∼N (0, I)

    ˆx = fθ(z, t, y) 4. εc∼N (0, I)

  55. [56]

    zc = tLF· ˜xLF + (1−tLF)· εc

    v = (x− z)/(1−t) 5. zc = tLF· ˜xLF + (1−tLF)· εc

  56. [57]

    ˆxLF = fθ(zc, tLF, y, res = 0) 7.L =∥v− ˆv∥2 7.Lc =∥ˆxLF− ˜xLF∥2/(1−tLF)2 — Phase 1 (fine, H× W ) —

    ˆv = (ˆx− z)/(1−t) 6. ˆxLF = fθ(zc, tLF, y, res = 0) 7.L =∥v− ˆv∥2 7.Lc =∥ˆxLF− ˜xLF∥2/(1−tLF)2 — Phase 1 (fine, H× W ) —

  57. [58]

    Sample t′ LF∼U (t∗, 1)

  58. [59]

    εLF∼N (0, I), εHF∼N (0, I)

  59. [60]

    zLF = t′ LF· ˜xLF + (1−t′ LF)· εLF tHF = (t′ LF−t∗)/(1−t∗)

  60. [61]

    zHF = tHF· xHF + (1−tHF)· εHF

  61. [62]

    zf = IDWT(zLF· SLF, z HF)

  62. [63]

    ˆx = fθ(zf , t′ LF, y, res = 1)

  63. [64]

    Left: standard flow-matching (JiT baseline)

    (ˆxLF, ˆxHF) = DWT(ˆx) 15.Lf =∥ˆxLF/SLF− ˜xLF∥2 (1−t′ LF)2 +∥ˆxHF− xHF∥2 (1−tHF)2 Return:L Return:Lc,Lf Table 10 Training pseudocode comparison. Left: standard flow-matching (JiT baseline). Right: our Wavelet-aware image Transformer extension. Both phases are packed into a single forward pass via FlashAttention’s variable-length interface; fθ is the share...

  64. [65]

    Split at t∗: K1, K2 via multiplier m

  65. [66]

    z∼N (0, I) [H/2× W /2] — Phase 0 ( t=0→ t∗, coarse, normalized LF only) —

    z∼N (0, I) [H× W ] 3. z∼N (0, I) [H/2× W /2] — Phase 0 ( t=0→ t∗, coarse, normalized LF only) —

  66. [67]

    for i = 0 to K−1: 4. for i = 0 to K1−1: ˆx = (1+w)fθ(z, ti, y)− wfθ(z, ti,∅) ˆxLF = (1+w)fθ(z, ti, y, 0)− wfθ(z, ti,∅, 0) v = (ˆx− z)/(1−ti) v = (ˆxLF− z)/(1−ti) z← z + (ti+1− ti)· v z← z + (ti+1− ti)· v — Transition at t∗: un-normalize LF, inject fresh HF noise —

  67. [68]

    zHF∼N (0, I) [HF resolution]

  68. [69]

    z = IDWT(zLF, zHF) [H× W , pixel space] — Phase 1 ( t∗→ 1, fine, joint LF + HF in pixel space) —

  69. [70]

    a hummingbird hovering in front of a fuchsia flower, wings frozen mid-beat

    for i = K1 to K−1: ˆx = (1+w)fθ(z, ti, y, 1)− wfθ(z, ti,∅, 1) (zLF, zHF) = DWT(z), (ˆxLF, ˆxHF) = DWT(ˆx) tHF = (ti−t∗)/(1−t∗) vLF = (ˆxLF− zLF)/(1−ti) vHF = (ˆxHF− zHF)/(1−tHF) zLF← zLF + (ti+1−ti)· vLF zHF← zHF + (ti+1−ti)· vHF z = IDWT(zLF, zHF) Return: z Return: z Table 11 Sampling pseudocode comparison. Left: standard single-phase sampling. Right: ou...

    2048

  70. [71]

    All classes ranked by human-verified positive label count; top 1,000 selected

    Class selection. All classes ranked by human-verified positive label count; top 1,000 selected

  71. [72]

    Each multi-labeled image is assigned to its rarest valid class: arg min c∈Ci |Sc|

    Image assignment (rarest-first). Each multi-labeled image is assigned to its rarest valid class: arg min c∈Ci |Sc|. Tie-breaking: when multiple classes share the same count, Python’s stable sort preserves the order in which labels appear in the source CSV. Images are iterated in dictionary insertion order (order of first appearance in the CSV). Both behav...

  72. [73]

    Per-class cap at 1,280 images ( ≈ ImageNet’s 1,281/class)

  73. [74]

    For classes below 1,280 images, machine-labeled images with confidence ≥ 0.90 are added

    Machine label supplementation. For classes below 1,280 images, machine-labeled images with confidence ≥ 0.90 are added. Dynamic deduplication ensures each image is assigned to at most one class

  74. [75]

    ∼103 classes remain below 1,280 (min ∼111); retained as-is

    Thin-class retention. ∼103 classes remain below 1,280 (min ∼111); retained as-is

  75. [76]

    65 images/class (50 target + 15 buffer), random.seed(42)

    Validation carve-out. 65 images/class (50 target + 15 buffer), random.seed(42). Buffer absorbs ∼6% resolution filter loss in Stage 2. Table 19 Manifest parameters for OpenImages-512. Parameter Value Number of classes 1,000 Max images per class 1,280 Machine label confidence ≥ 0.90 Validation target / buffer per class 50 / 15 Random seed 42 Stage 2: Image ...

  76. [77]

    Discard if min (w, h) < 512 (strict ≥; the code is if min(w,h) < MIN_DIM )

    Resolution filter. Discard if min (w, h) < 512 (strict ≥; the code is if min(w,h) < MIN_DIM ). No upsampling. ∼6% filtered

  77. [78]

    Convert to RGB (before resize; handles grayscale, CMYK, palette)

  78. [79]

    Resize shorter side to 512 px via Lanczos

  79. [80]

    Center crop to 512×512: left=⌊(w′−512)/2⌋, top=⌊(h′−512)/2⌋

  80. [81]

    Validation is capped at exactly 50/class after processing; surplus deleted

    Save JPEG quality=100, subsampling=0 (4:4:4), icc_profile=None. Validation is capped at exactly 50/class after processing; surplus deleted. Final counts: 848,746 train, 49,974 val ( 26 validation images short of 50,000 were skipped due to corrupt/un- readable source files). E.3 OpenImages-1024 Construction The 1024 variant’s key architectural difference: ...

Showing first 80 references.