Pith. sign in

REVIEW 5 major objections 10 minor 38 references

Learned Compression for Compressed Learning

T0 review · 5 major / 10 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read WaLLoC shows that a single linear layer between a wavelet packet transform and its inverse yields a codec that is cheap to encode, compresses harder than generative autoencoders, and lets compressed-domain models beat resolution reduction.

desk verdict WaLLoC's cheap linear encoder is a clever idea with real promise, but the paper's 'drop-in replacement' claim outruns the evidence; the missing ablation on task-critical detail is the key gap. read the letter →

arxiv 2412.09405 v1 pith:PBIME5B6 submitted 2024-12-12 eess.IV cs.CVcs.LGeess.ASeess.SP

classification eess.IVcs.CVcs.LGeess.ASeess.SP
keywords waveletpackettransformlearnedimagecompressioncompressed-domainlearningentropybottleneckdimensionalityreductionlatentdiffusionautoencoderaudiocodecresolutionscaling
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

WaLLoC is a lossy compression architecture that targets compressed-domain learning: operating machine-learning models directly on compact latents instead of on downsampled pixels or audio samples. The paper claims that by placing a single linear learned layer and an entropy bottleneck between a wavelet packet transform and its inverse, a codec can simultaneously achieve very cheap encoding, high compression ratio, and uniform dimensionality reduction up to 20x. On RGB images, the paper reports less than five percent of the encoding cost of other neural codecs and a compression ratio about six times higher than the Stable Diffusion 3 variational autoencoder at similar perceptual quality. On downstream tasks, models trained on WaLLoC latents outperform models trained on resolution-reduced inputs by wide margins, including 50.3% versus 23.1% classification accuracy at 16x reduction. If correct, WaLLoC is a drop-in replacement for resolution reduction in compressed-domain learning.

What carries the argument

The load-bearing object is the invertible wavelet packet transform (WPT) built from Cohen-Daubechies-Feauveau 9/7 filters, the same filterbank used in JPEG 2000. The WPT trades spatial or temporal resolution for frequency resolution without losing information, so the neural part of the codec can operate at low resolution; then a single linear analysis layer projects the WPT coefficients to a low-dimensional latent, and an entropy bottleneck (additive uniform noise during training) makes the latents robust to 8-bit quantization. The synthesis side is a deep convolutional network that untangles the mixed components before the inverse WPT restores the signal. This sandwich is what simultaneously produces cheap encoding, quantizable latents, and uniform dimensionality reduction.

What would settle it

For a fixed WaLLoC codec, compute the per-image residual $X - \hat{X}$ after encode-decode and train the same downstream classifier on those residuals. If classification accuracy on residuals rises well above chance for a task where WaLLoC latents are used, the claim that the linear encoder discards only redundancy is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that the hard part of lossy encoding, discarding signal redundancy, becomes nearly trivial once the signal is in a wavelet packet basis, while the hard part of decoding, reassembling a clean signal, can be left to a deep nonlinear synthesis network. WaLLoC exploits this asymmetry by making the encoder a single linear layer that projects wavelet-packet coefficients to a low-dimensional latent, with a Gaussian companding and additive-noise entropy bottleneck to make the latents resilient to 8-bit quantization, and a convolutional decoder that reconstructs before the inverse wavelet packet transform. The consequence is an autoencoder whose encoder is cheap enough for edge devices, whose latents are uniformly reduced in dimension by a fixed factor, and whose compression ratio beats the generative autoencoders used in latent diffusion models while preserving detail without adversarial or perceptual losses. The paper demonstrates the framework on RGB images and stereo audio, and shows that models trained on the latents outperform resolution-reduced models on classification, colorization, document understanding, and music source separation.

Load-bearing premise

The whole method depends on the empirical premise that after a wavelet packet transform, a single learned linear layer can throw away redundant information without losing the details a downstream task needs; the paper gives no guarantee that the discarded components are never task-critical.

Editorial extensions

If this is right

  • At the same compute budget as 16x resolution reduction, a ViT-Ti on WaLLoC latents reaches 50.3% ImageNet accuracy versus 23.1% for resized pixels.
  • Document understanding follows the same pattern: PaliGemma on WaLLoC latents at 16x reduction scores 81.1 ANLS versus 43.7 for Lanczos-resampled downsampling.
  • For signal enhancement, WaLLoC latents improve colorization LPIPS by 0.40 dB at 16x and source separation PSNR by 2.4 dB at both 5x and 18x rates.
  • WaLLoC reaches 35:1 compression at 16x dimension reduction on RGB images, roughly six times the compression ratio of the Stable Diffusion 3 VAE at similar DISTS quality.
  • Because the entropy bottleneck makes latents robust to 8-bit quantization, off-the-shelf lossless codecs such as WebP add up to 12x extra compression beyond the dimension reduction.

Reading between the lines

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

  • Beyond the paper: if the single-linear-layer encoder hypothesis holds, the same sandwich recipe of an invertible sparsifying transform plus a minimal learned projection should transfer to video, hyperspectral, and microscopy data; the paper lists these as future work.
  • Beyond the paper: the uniform latent geometry suggests WaLLoC latents could be a stable tokenization for pre-training vision-language or audio-language models, since the fixed reduction factor keeps sequence length predictable; the paper does not test this.
  • Beyond the paper: the entropy-bottleneck design is trained for 8-bit quantization with a specific noise scale; an experimental check of how quality degrades when latents are quantized to 4 or 6 bits would map the operating range more fully.
  • Beyond the paper: comparing WaLLoC latents against resolution-reduced inputs on a fixed compute budget, rather than fixed resolution, would directly quantify the accuracy-per-watt gain; the paper compares efficiency and accuracy separately.
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

5 major / 10 minor

Summary. The paper introduces WaLLoC, a lossy compression architecture that sandwiches a shallow linear analysis transform and a nonlinear synthesis transform between an invertible wavelet packet transform, with an entropy bottleneck for quantization resilience. The authors claim that WaLLoC simultaneously achieves computationally efficient encoding, high compression ratios, and uniform dimensionality reduction, making it a suitable drop-in replacement for resolution reduction in compressed-domain learning. They compare WaLLoC against existing codecs on image and audio compression metrics and demonstrate downstream gains on image classification, colorization, document understanding, and music source separation.

Significance. If the results hold, WaLLoC is a meaningful contribution to compressed-domain learning. The architecture is simple, modality-agnostic, and the encoding pipeline is dominated by linear operations, which is a practical advantage for mobile and remote sensing scenarios. The paper provides a systematic comparison with several codec families and shows large accuracy improvements over resolution reduction on three of the four evaluated tasks. The strength of the paper lies in the concrete downstream evaluations and the insight that a wavelet packet transform can expose redundancy so that a single linear layer suffices for dimensionality reduction. However, the significance is tempered by the fact that one of the four downstream experiments (document understanding) does not actually perform compressed-domain learning, and by the absence of uncertainty quantification and ablations that would substantiate the general 'drop-in replacement' claim.

major comments (5)
  1. [§4.2(b), Table 3] The DocVQA experiment does not evaluate compressed-domain learning. The 'compress' condition decodes WaLLoC latents to full 896×896 resolution before passing them to PaliGemma, while the 'resize' condition downsamples to 224×224 and then Lanczos-upsamples to 896×896. In both conditions the model processes 896×896 inputs, so the compute, latency, and memory are identical; only the input image quality differs. The reported ANLS improvement (81.1 vs 43.7) therefore reflects reconstruction quality, not a compressed-domain benefit, and the claim that WaLLoC provides 'the same improvement in latency and memory consumption' for document understanding is unsupported. Please either fine-tune PaliGemma on WaLLoC latents directly, or explicitly relabel this experiment as a reconstruction-quality comparison.
  2. [Tables 1–3] All quantitative results are single-run point estimates with no error bars, confidence intervals, or significance tests. The central claims are empirical, so the reader cannot assess the reliability of the reported differences, particularly in Table 3 where the classification improvement at 16× is 50.3% vs 23.1%. Please report means and standard deviations over at least three independent seeds for the downstream tasks, and bootstrap confidence intervals for the codec metrics in Tables 1 and 2.
  3. [§3.1, §4.2] The analysis transform is trained solely to minimize MSE reconstruction under additive noise, with no coupling to downstream tasks. The paper provides no ablation comparing the learned linear projection to a random projection, to the low-pass subband alone, or to a task-optimized projection at the same dimensionality. The null space of the fixed linear map is selected by reconstruction error on natural images and audio, and tasks whose discriminative information lies in that null space (e.g., fine text edges, small anomalies, subtle texture) may fail even while reconstruction metrics look acceptable. Please add such ablations and state this limitation explicitly, or restrict the 'drop-in replacement' claim to the evaluated tasks.
  4. [Table 1] The comparison between WaLLoC 16× and the Stable Diffusion 3 VAE is made at different dimensionality reductions (16× vs 12×) and different compression ratios. The 'similar quality' claim relies on DISTS dB, while PSNR differs by 6.6 dB (27.5 vs 20.9). Please provide a rate-distortion comparison at matched dimensionality reduction (e.g., a WaLLoC 12× variant) and report PSNR, MS-SSIM, LPIPS, and DISTS at the same operating point, or justify why DISTS is the single decisive metric for this comparison.
  5. [§3.4] The paper omits essential training details: learning rate, optimizer, batch size, number of steps, and hardware used to train the four codecs. The statement 'Additional implementation details are available in our public code repositories' is not sufficient for a stand-alone methods section. Please include these details in an appendix so the reader can reproduce the codecs without consulting external references.
minor comments (10)
  1. [Appendix, Figures 10–11] Figures 10 and 11 are never referenced in the main text; they appear to probe the learned latent channels by decoding a single nonzero latent element. Please add a sentence in the main text or appendix describing what these probes show and what conclusion the reader should draw.
  2. [§3.4] The companding operation is described as Φ(z) but the exact scaling to the range [-127, 127] is not specified. Please give the precise formula.
  3. [§3.4] The sentence 'We optimize the noise scale for 8-bit quantization' is vague. Clarify whether the noise scale is a fixed hyperparameter and how the value was chosen.
  4. [Table 1 and Table 2 captions] The throughput units are given as 'Megapixels/sec, CPU' but the specific CPU model is not named. Please list the exact platforms used for the low-power CPU, high-power CPU, and GPU measurements.
  5. [§3.4] The statement 'We found that WebP’s entropy coding provided the best compression ratio—even for audio signals' is surprising. Please describe how WebP is applied to audio latents and why it outperforms zlib and PNG for this data type.
  6. [§3.2] The claim that the entropy coding step provides 'an additional compression multiplier of up to 12× compared to reducing the dimension only' is not derived. Please show the calculation or provide a reference.
  7. [References] Reference [22] is a blog post ('On the speed of ViTs and CNNs'); consider citing a peer-reviewed source or a technical report with more permanent archival status.
  8. [§3.1] The assertion 'the encoder objective—discarding signal redundancies—becomes trivial after applying the WPT' is made without quantitative support. Provide evidence such as cumulative energy concentration in subbands, or soften the wording to avoid overstating the claim.
  9. [Table 1] The comparison with Cheng2020 is at a single rate point. A rate-distortion curve across multiple bitrates would better illustrate the trade-off between WaLLoC and end-to-end learned codecs.
  10. [Abstract and Introduction] The claim that 'WaLLoC does not require perceptual or adversarial losses to represent high-frequency detail' is only demonstrated for the specific codecs trained on LSDIR and MUSDB18-HQ. State the scope of this claim more carefully.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: WaLLoC's claims are measured against external baselines; the latent dimension is a chosen hyperparameter, and downstream improvements are empirical comparisons, not derived from fitted parameters.

full rationale

I found no circular reasoning in the paper. The WaLLoC architecture is described concretely: a fixed CDF 9/7 wavelet packet transform, a single learned linear analysis layer with a chosen latent dimension C_z, an entropy bottleneck, and a learned nonlinear synthesis transform. The training objective is explicitly mean squared reconstruction error under additive uniform noise, and none of the claimed downstream results are obtained by fitting a parameter to the evaluation target and then relabeling that fit as a prediction. Compression ratios, encoding/decoding throughput, PSNR, DISTS, SSDR, and downstream accuracies are all measured against external baselines (Cheng2020, Stable Diffusion 3 VAE, EnCodec, Stable Audio, Opus, WebP) or against resolution-reduced inputs in controlled comparisons. The dimension reduction factors (4x, 16x, 5x, 20x) are architectural hyperparameters, not predicted outcomes. The central claim that a linear projection of wavelet-packet coefficients preserves task-critical detail is an empirical hypothesis, and the paper's own limitation is that no bound or ablation establishes it generally; this is a correctness/robustness risk, not circularity. The paper does not invoke a uniqueness theorem or a self-citation to justify its design choice; the cited entropy-bottleneck method (Ballé et al.) and decoder blocks (Stable Diffusion / Stable Audio) are independent prior work used as components, not as evidence that WaLLoC's own measurements must hold. Therefore the derivation chain is self-contained with respect to the evaluated claims, and the appropriate circularity score is 0.

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

The central result rests on four hand-chosen design parameters and four assumptions. The most important free parameter is the latent dimension Cz, which sets the dimensionality reduction factor. The most fragile assumption is that a single linear projection after the wavelet packet transform discards only redundant information, which is not formally guaranteed. No invented physical entities are introduced.

free parameters (4)
  • Latent dimension Cz = image: 48 and 12; audio: about 102 and 27
    Chosen by hand to set the dimensionality reduction factor; it directly determines downstream sequence length and compression ratio. Section 3.4.
  • Wavelet packet transform depth J = 3 for RGB images, 8 for stereo audio
    Hand-chosen to balance energy compaction against channel expansion; J determines how much spatial or temporal resolution is traded for frequency channels. Section 3.4.
  • Decoder hidden dimension Chidden = 768 for both image and audio decoders
    Fixed design choice; a larger hidden dimension would improve decoder quality at higher compute cost. Section 3.4.
  • Entropy bottleneck noise scale = U[-0.5, 0.5] scaled to signed 8-bit range [-127, 127]
    Set to make rounding to signed 8-bit integers safe; the noise scale controls the trade-off between rate and distortion. Sections 3.2 and 3.4.
assumptions (4)
  • standard math The CDF 9/7 wavelet packet transform is linear and invertible, so information is only discarded by the learned projection and quantization.
    Invoked in Section 3.1 and Figure 4; the whole design relies on WPT being invertible and linear.
  • domain assumption Natural signal wavelet subbands follow a generalized Gaussian distribution, justifying Gaussian CDF companding.
    Stated in Section 3.4 with citation [25]; if false, companding would be a poor match and quantization behavior would differ.
  • domain assumption Additive uniform noise during training produces latents that are resilient to hard rounding at inference.
    Adopted from Balle et al. [8] in Section 3.2; the paper optimizes noise for 8-bit quantization and relies on this transfer for the compression pipeline.
  • ad hoc to paper A single linear analysis layer after the WPT suffices to discard redundant information without harmful loss.
    Core design hypothesis in Section 3.1; it is justified only by the reported experiments, not derived or formally bounded.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learned Compression for Compressed Learning." pith.science (2026). https://pith.science/paper/PBIME5B6

@misc{pith2026241209405,
  author       = {Pith},
  title        = {Pith review of: Learned Compression for Compressed Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PBIME5B6}},
  note         = {Machine review of arXiv:2412.09405}
}
read the original abstract

Modern sensors produce increasingly rich streams of high-resolution data. Due to resource constraints, machine learning systems discard the vast majority of this information via resolution reduction. Compressed-domain learning allows models to operate on compact latent representations, allowing higher effective resolution for the same budget. However, existing compression systems are not ideal for compressed learning. Linear transform coding and end-to-end learned compression systems reduce bitrate, but do not uniformly reduce dimensionality; thus, they do not meaningfully increase efficiency. Generative autoencoders reduce dimensionality, but their adversarial or perceptual objectives lead to significant information loss. To address these limitations, we introduce WaLLoC (Wavelet Learned Lossy Compression), a neural codec architecture that combines linear transform coding with nonlinear dimensionality-reducing autoencoders. WaLLoC sandwiches a shallow, asymmetric autoencoder and entropy bottleneck between an invertible wavelet packet transform. Across several key metrics, WaLLoC outperforms the autoencoders used in state-of-the-art latent diffusion models. WaLLoC does not require perceptual or adversarial losses to represent high-frequency detail, providing compatibility with modalities beyond RGB images and stereo audio. WaLLoC's encoder consists almost entirely of linear operations, making it exceptionally efficient and suitable for mobile computing, remote sensing, and learning directly from compressed data. We demonstrate WaLLoC's capability for compressed-domain learning across several tasks, including image classification, colorization, document understanding, and music source separation. Our code, experiments, and pre-trained audio and image codecs are available at https://ut-sysml.org/walloc

Figures

Figures reproduced from arXiv: 2412.09405 by the authors.

Figure 1
Figure 1. In discriminative models (left), resolution reduction increases training and in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of our proposed method (WaLLoC) with other autoencoder designs [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. WaLLoC’s encode-decode pipeline. The entropy bottleneck and entropy coding [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Example of forward and inverse WPT with J = 2 levels. Each level applies filters LA and HA independently to each of the signal channels, followed by downsampling by a factor of two (↓ 2). An inverse level consists of upsampling (↑ 2) followed by LS and HS, then summing…
Figure 5
Figure 5. Figure 5: Cheng et al. 2020 [11] [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Stable Diffusion 3 VAE [12] [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: WaLLoC 4× [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: WaLLoC 16× [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Stereo reconstruction of an audio segment from the MUSDB test set. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Result of using the Cz = 12 RGB codec (WaLLoC 16×) to decode a 12 × 3 × 3 latent with all elements equal to zero except except for channel i, which is set to   0 0 0 0 31 0 0 0 0   [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Result of using the Cz = 48 RGB codec (WaLLoC 4×) to decode a 48 × 3 × 3 latent with all elements equal to zero except except for channel i, which is set to   0 0 0 0 31 0 0 0 0   [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

  1. [1]

    WaLLoC uses the computationally cheap and invertible wavelet packet transform [14] to expose signal redundancies prior to autoencoding

    Computationally efficient encoding to reduce overhead in compressed-domain learning and support resource constrained mobile and remote sensors. WaLLoC uses the computationally cheap and invertible wavelet packet transform [14] to expose signal redundancies prior to autoencoding. This allows us to replace the encoding DNN with a single linear layer ( <100k...

  2. [2]

    Lossy codecs typically achieve high compression by combining quantization and entropy cod- ing

    High compression ratio for storage and transmission efficiency. Lossy codecs typically achieve high compression by combining quantization and entropy cod- ing. However, naive quantization of autoencoder latents leads to unpredictable and unbounded distortion. Instead, we apply additive noise during training as an entropy bottleneck [8], leading to quantiz...

  3. [3]

    WaL- LoC’s encoder projects high-dimensional signal patches to low-dimensional latent representations, providing a reduction of up to 20 ×

    Dimensionality reduction to accelerate compressed-domain modeling. WaL- LoC’s encoder projects high-dimensional signal patches to low-dimensional latent representations, providing a reduction of up to 20 ×. This allows WaLLoC to be Distortion (PSNR) Compression Ratio Dimension Reduction Throughput (MPixels/sec) Perceptual Quality (DISTS) Distortion (SDR) ...

  4. [4]

    Paligemma: A versatile 3b vlm for transfer,

    L. Beyer et al., “Paligemma: A versatile 3b vlm for transfer,” arXiv:2407.07726, 2024

  5. [5]

    End-to-end learned compression

    but does not address the computational issues of scaling DNNs to high resolution. End-to-end learned compression. Nonlinear autoencoders that are jointly opti- mized for both rate and distortion [8] achieve higher compression ratios than LTC, but require more computation [16] and offer limited dimensionality reduction—typically 4× [11]. Efficient decoding...

  6. [6]

    Imagenet classification with deep convolutional neural net- works,

    A. Krizhevsky et al., “Imagenet classification with deep convolutional neural net- works,” NeurIPS, 2012

  7. [7]

    Large-scale audio classification,

    S. Hershey et al., “Large-scale audio classification,” in ICASSP, 2017

  8. [8]

    Segment anything for microscopy,

    A. Archit et al., “Segment anything for microscopy,” bioRxiv, 2023

Show all 38 references
  1. [9]

    Deep residual learning in the jpeg transform domain,

    M. Ehrlich and L. Davis, “Deep residual learning in the jpeg transform domain,” in ICCV, 2019

  2. [10]

    Seit: Storage-efficient vision training with tokens using 1% of pixel storage,

    S. Park et al., “Seit: Storage-efficient vision training with tokens using 1% of pixel storage,” in ICCV, 2023

  3. [11]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach et al., “High-resolution image synthesis with latent diffusion models,” in CVPR, 2022

  4. [12]

    End-to-end optimized image compression,

    Johannes et al. Ball´ e, “End-to-end optimized image compression,” in ICLR, 2017

  5. [13]

    Stable audio open,

    Z. Evans et al., “Stable audio open,” arXiv:2407.14358, 2024

  6. [14]

    Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks,

    M. Goldblum et al., “Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks,” NeurIPS, 2024

  7. [15]

    Learned image compression with discretized gaussian mixture likeli- hoods and attention modules,

    Z. Cheng et al., “Learned image compression with discretized gaussian mixture likeli- hoods and attention modules,” in CVPR, 2020

  8. [16]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser et al., “Scaling rectified flow transformers for high-resolution image synthesis,” in ICML, 2024

  9. [17]

    High fidelity neural audio compression,

    A. D´ efossez et al., “High fidelity neural audio compression,” arXiv:2210.13438, 2022

  10. [18]

    Mallat, A Wavelet Tour of Signal Processing: The Sparse Way, 2008

    S. Mallat, A Wavelet Tour of Signal Processing: The Sparse Way, 2008

  11. [19]

    The jpeg ai standard: Providing efficient human and machine visual data consumption,

    J. Ascenso et al., “The jpeg ai standard: Providing efficient human and machine visual data consumption,” IEEE Multimedia, 2023

  12. [20]

    Advancing the rate-distortion-computation frontier for neural image compression,

    David Minnen and Nick Johnston, “Advancing the rate-distortion-computation frontier for neural image compression,” in ICIP, 2023

  13. [21]

    Computationally-efficient neural image compression with shallow decoders,

    Yibo Y. Yang and S. Mandt, “Computationally-efficient neural image compression with shallow decoders,” in ICCV, 2023

  14. [22]

    Maskgit: Masked generative image transformer,

    H. Chang et al., “Maskgit: Masked generative image transformer,” in CVPR, 2022

  15. [23]

    Simple and controllable music generation,

    J. Copet et al., “Simple and controllable music generation,” NeurIPS, 2024

  16. [24]

    Taming transformers for high-resolution synthesis,

    P. Esser et al., “Taming transformers for high-resolution synthesis,” in CVPR, 2021

  17. [25]

    EfficientNet: Rethinking model scaling for convolutional neural networks,

    M. Tan and L. Quoc, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in ICML, 2019

  18. [26]

    On the speed of ViTs and CNNs,

    L. Beyer, “On the speed of ViTs and CNNs,” lb.eyer.be/a/vit-cnn-speed, 2024

  19. [27]

    Taesd: Tiny autoencoder for stable diffusion,

    O. Bohan, “Taesd: Tiny autoencoder for stable diffusion,” 2023

  20. [28]

    Elic: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding,

    D. He et al., “Elic: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding,” in CVPR, 2022

  21. [29]

    Subband coding of color images,

    P. Westerink et al., “Subband coding of color images,” Subband Image Coding, 1991

  22. [30]

    The musdb18 corpus for music separation,

    Z. Rafii, “The musdb18 corpus for music separation,” 2017

  23. [31]

    Lsdir: A large scale dataset for image restoration,

    Y. Li et al., “Lsdir: A large scale dataset for image restoration,” in CVPR, 2023

  24. [32]

    Image quality assessment: from error visibility to structural similar- ity,

    Z. Wang et al., “Image quality assessment: from error visibility to structural similar- ity,” IEEE transactions on image processing, 2004

  25. [33]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang et al., “The unreasonable effectiveness of deep features as a perceptual metric,” in CVPR, 2018

  26. [34]

    Image quality assessment: Unifying structure and texture similarity,

    K. Ding et al., “Image quality assessment: Unifying structure and texture similarity,” IEEE transactions on pattern analysis and machine intelligence, 2020

  27. [35]

    Quantifying spatial audio quality impairment,

    K. Watcharasupat and A. Lerch, “Quantifying spatial audio quality impairment,” in ICASSP, 2024

  28. [36]

    Cdpam: Contrastive learning for perceptual audio similarity,

    P. Manocha et al., “Cdpam: Contrastive learning for perceptual audio similarity,” in ICASSP, 2021

  29. [37]

    Maxvit: Multi-axis vision transformer,

    Z. Tu et al., “Maxvit: Multi-axis vision transformer,” in ECCV, 2022

  30. [38]

    Docvqa: A dataset for vqa on document images,

    M. Mathew et al., “Docvqa: A dataset for vqa on document images,” in IEEE/CVF winter conference on applications of computer vision, 2021. Appendix Figure 5: Cheng et al. 2020 [11] Figure 6: Stable Diffusion 3 V AE [12] Figure 7: WaLLoC 4 × Figure 8: WaLLoC 16 × Figure 9: Stere...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.